a proposed platform for intelligent identification of organs from medical images

Proceedings

of the Third International Advanced Research Workshop on

In Silico Oncology: Advances and Challenges

Zografeio Lyceum, Istanbul, Turkey September 23-24, 2008

Organizing Committee

Georgios Stamatakos, ICCS, National Technical University of Athens Metin Akay, Harrington Department of Bioengineering, Arizona State University

Dimitra Dionysiou, ICCS, National Technical University of Athens and Nikolaos Uzunoglu, ICCS, National Technical University of Athens

The workshop has been technically co‐sponsored by

Sponsors

Editors G. Stamatakos and D. Dionysiou

Galenica S.A. Wyeth Hellas S.A

i

All enquiries should be addressed to Dr G. S. Stamatakos, National Technical University of Athens, Institute of Communication and Computer Systems, MFOL, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR‐157 80, Greece (Phone: + 30 210 772 2288, Fax: + 30 210 772 3557, E‐mail: [email protected] ).

ii

3rd International Advanced Research Workshop on In Silico Oncology: Advances and Challenges

Zografeio Lyceum, Istanbul, Turkey, September 23‐24, 2008

Introduction

The workshop aims at contributing to the shaping of the emerging field of in silico (computational) oncology. In Silico Oncology is a complex and multiscale combination of sciences and technologies intending to simulate malignant tumor growth and tumor and normal tissue response to therapeutic modalities at all levels of biocomplexity. The long term goal is to quantitatively understand cancer and related phenomena and optimize therapeutic interventions by performing in silico (on the computer) experiments based on the individual patient’s clinical, imaging, histopathological, molecular and pharmacogenomic data. In order to achieve such an ambitious goal translation of cancer models into the clinical trials arena is a sine qua non condition and therefore this aspect is particularly emphasized throughout the contributed papers. Contributions address basic science, technology and clinical aspects of in silico oncology.

Organizing Committee

Georgios Stamatakos, ICCS, National Technical University of Athens Metin Akay, Harrington Department of Bioengineering, Arizona State University

Dimitra Dionysiou, ICCS, National Technical University of Athens and Nikolaos Uzunoglu, ICCS, National Technical University of Athens

Website

http://www.3rd‐iarwiso.iccs.ntua.gr/

The workshop has been technically co‐sponsored by

Institute of Electrical and Electronics Engineers, Engineering in Medicine and Biology Society

Sponsors

National Technical University of Athens Institute of Communication and Computer Systems

Galenica S.A. Wyeth Hellas S.A

iii

TABLE OF CONTENTS

Introduction...........................................................................................................................................................ii Table of Contents.................................................................................................................................................iii Scientific Program................................................................................................................................................iv N. Uzunoglu, Scientific and Technological Background of In Silico Oncology.............................................................................1 E. Pirogova, T. Istivan, P. Coloe, M. Akay, I. Cosic, Investigation of Protein Activity Using de‐novo Designed Peptides and LaserLight Radiation: the RRM Computational Predictions..................................................................................................................................... 4 Georgios S. Stamatakos , In Silico Oncology: a Paradigm for Clinically Oriented Living Matter Engineering............................................7 A.W. Hoppe, A Clinical View on In Silico Oncology..................................................................................................................10 D. D. Dionysiou and G. S. Stamatakos, Glioblastoma Multiforme Response to Radiotherapy: Critical Parameters and In Silico Trials......................................................................................................................................................................13 B.A.Lloyd and G.Szekely, Modeling Growth Saturation in Avascular Tumors.............................................................................................17 E.Skounakis, K. Banitsas, and K. Marias, A Proposed Platform for Intelligent Identification of Organs from Medical Images............................................20 L. Skarlas, A. Antonacopoulou and H. Kalofonos, POLR2F, ATP6V0A1, PRNP and their Prognostic Value in Dukesʹ B Colon Cancer.......................................23 E. Ch. Georgiadi, E. A. Kolokotroni, D. D. Dionysiou, N. M. Graf, A. Hoppe, N. K. Uzunoglu and G. S. Stamatakos, Simulating the Response of Nephroblastoma Tumor to Chemotherapy in the Clinical Context...................................................................................................................................................................27 E. A. Kolokotroni, E. Ch. Georgiadi, D. D. Dionysiou and G. S. Stamatakos, A 4‐D Simulation Model of Tumor Free Growth and Response to Chemotherapy in Vivo: The Breast Cancer Case........................................................................................................................................................................31 S. G. Giatili, G. S. Stamatakos, D. D.Dionysiou, E. A. Kolokotroni, E. Ch. Georgiadi, Response to Chemotherapeutic Schemes in the Context of the ACGT Oncosimulator.............................................................35 D.Lavenier and J. Jacques, Paralellizing the ACGT Oncosimulator.................................................................38 T. Athanaileas, A. Menychtas, D. Dionysiou, G. Stamatakos, D. Kaklamani, T. Varvarigou, I. Venieris and N. Uzunoglu, Applying Grid Computing Technologies to In Silico Oncology...........................................................................41 K. Marias, Multi‐level Image Analysis for Extracting Pathophysiological Parameters Related to Cancer Modeling................................................................................................................................................................44 LATE ARRIVED PAPER: E.Konukoglu, O. Clatz, H. Delingette and N. Ayache, Parameter Estimation for Reaction‐Diffusion Tumor Growth Models from Time Series of Images…………………………………………………………………………………………………………....48

iv

3rd International Advanced Research Workshop on In Silico Oncology: Advances and Challenges Venue Zografeio Lyceum, Turnacibasi Sok. 27, 80060 Beyoglu, Galatasaray, Istanbul, Turkey Scientific Program Tuesday 23 September 2008 9:00‐9:15

G. Stamatakos, M. Akay and N. Uzunoglu

Welcome address

9:15‐10:00 N. Uzunoglu Scientific and technological background of in silico oncology

10:00‐10:45 E. Pirogova, T. Istivan, P. Coloe, M. Akay, I. Cosic

Investigation of protein activity using de‐novo designed peptides and laser light radiation: the RRM computational predictions

10:45‐11:30 G. Stamatakos In silico oncology: a paradigm of clinically oriented living matter engineering

11:30‐12:15 A. Hoppe A clinical view on in silico oncology 12:15‐14:00 Lunch 14:00‐14:45 D. Dionysiou and

G. Stamatakos Glioblastoma multiforme response to Radiotherapy: critical parameters and In Silico trials.

14:45‐15:30 B. A. Lloyd and G. Szekely

Modeling growth saturation in avascular tumors

15:30‐16:15 E. Skounakis, K.Banitsas, and K. Marias

A proposed platform for intelligent identification of organs from medical images

16:15‐17:00 L. Skarlas, A. Antonacopoulou and H. Kalofonos

POLR2F, ATP6V0A1, PRNP and their prognostic value in Dukesʹ B colon cancer

Social Program [City landmark sightseeing and dinner]

v

Wednesday 24 September 2008 9:00‐9:45 E. Ch. Georgiadi, E.

A. Kolokotroni , D. Dionysiou, N. M. Graf, A. Hoppe, N. K. Uzunoglu and G. Stamatakos

Simulating the nephroblastoma response to chemotherapy in the clinical context

9:45‐10:30 E. A. Kolokotroni, E. Ch. Georgiadi, D. D. Dionysiou and G. S. Stamatakos

A 4‐D simulation model of tumour free growth and response to chemotherapy in vivo: the breast cancer case

10:30‐11:15 S. G. Giatili, G. S. Stamatakos, D. Dionysiou, E. A. Kolokotroni, E. Ch. Georgiadi

Geometrical and mechanical aspects of tumor growth and response to chemotherapeutic schemes in the context of the ACGT project

11:15‐12:00 D. Lavenier and J. Jacques

Parallelizing the ACGT OncoSimulator

12:00‐14:00 Lunch 14:00‐14:45 T. Athanaileas, A.

Menychtas, D.Dionysiou, G. Stamatakos, D. Kaklamani, T. Varvarigou, I. Venieris and N. Uzunoglu

Applying grid computing technologies to in silico oncology

14:45‐15:30 K. Marias Multi‐level image analysis for extracting pathophysiological parameters related to cancer modeling

15:30‐16:30 Round table discussion

16:30‐17:00 G. Stamatakos, M. Akay and N. Uzunoglu

Conclusions

Social Program [City landmark sightseeing and dinner]

Proceedings of the 3rd International Advanced Research Workshop on In Silico Oncology, Istanbul, Turkey. Sept. 23-24, 2008. 1

Scientific and Technological Background of In Silico Oncology Nikolaos K. Uzunoglu

Abstract—Due to its strongly multidisciplinary character and high biocomplexity in silico oncology depends upon a successful combination and cross-polination of a host of sciences and technologies. This paper briefly outlines the contribution of several fields such as electromagnetic theory and numerical techniques to in silico oncology in the Laboratory of Microwaves and Fiber Optics, National Technical University of Athens. A short outline of the radiation therapy modeling paradigm is also provided.

I. CROSS-POLLINATING SCIENTIFIC DOMAINS HEab

scientific discipline of electrical engineering arose out 120 years ago just after the emergence of the first

applications of Maxwell’s electromagnetic theory. The foundations of electromagnetic theory are very similar to the Newtonian theory of dynamics and essentially both physical theories have a structure very similar to Euclidean geometry. The beauty of the electromagnetic theory has been enhanced by its wide applicability in explaining the dynamical behavior of all phenomena related to electric and magnetic fields varying along the time axis. Furthermore, when the Maxwell electromagnetic theory was recruited in order to explain the interaction of matter with electromagnetic energy its failure to explain the involved phenomena led to the emergence of quantum theory while electromagnetism in moving media gave the way to the special theory of relativity and afterwards to the theory of general relativity. The success of Maxwell theory is based on its reductionist concepts of scientific thought. What we call science today was born in western Anatolia and in particular in the town of Miletus from the 6th century B.C. onwards. It was based on the cultivation of the reductionist method. The observation of natural phenomena in combination with a pioneering spirit resulted into the concepts through which human mind can model nature by exploiting the power of thought. In this way by explaining simple phenomena one can build up a scientific method able to explain complex phenomena that may be viewed as

combinations of the simpler ones.

Manuscript received September 1, 2008. N.K.Uzunoglu is with the National Technical University of Athens,

School of Electrical and Computer Engineering,, 9 Iroon Polytechniou, GR 157 80, Greece (phone: +30 210 772 3556; fax: +30 210 772 3557; e-mail: nnap@)otenet.gr )

The research activities in the framework of electrical engineering science (and latter on with the addition of computer science and informatics) has been much influenced by the previously mentioned scientific way of thinking. Especially during the 1960’s many fundamental theories of physics such as applied quantum mechanics and quantum electrodynamics paved the way to many new inventions and technological applications. Advances in computer technology during the same period provided the opportunity to numerically model complex structures involving the interaction of matter with radiation. Problems considered impossible to analyze until that time became readily solvable and accurate predictions of the systems behavior were obtained. In the mid-1990’s it was possible to numerically compute electromagnetic structures with complexity of high degree. The main thrust to these efforts came from the calculation of the radar signals that were reflected from flying objects. This spirit of exploiting the capability of performing large scale computations penetrated the research community working on electromagnetics modeling. However, this advance did not come without any adverse effects since in many cases the physical intuition started to gradually disappear. Meanwhile the field of medical physics that was initiated as early as 1950s started to expand into new domains of medical applications. The technological applications in medicine, being restricted in 1940’s to only X-ray screening and a few electrical measurements, was continuously expanding to new fields such as linear accelerator applications, computers in medicine and to an extensive degree computers in biological process modeling. However, even by the end of 1990’s it was difficult to propose a detailed multiscale model of malignant cell cluster behavior relying on computational methods for funding. In many cases proposals were rejected with the dogmatic statement “cancer cannot be modeled by mathematics and physics”.

Research on biomedical engineering at the Microwave and Fiber Optics (MFOL)– National Technical University of Athens started in 1985 in two specific topics i.e hyperthermia and barin modeling. The influence of the above mentioned reductionist attitude combined with an engineering approach was implemented in order to design and construct new

T


systems. This led to the combination of theoretical modeling work with immediate clinical applications in the two aformentioned fields. Hyperthermia is an adjuvant cancer treatment modality based on the cytotoxic properties of biologically high temperatures (43o C) achieved in cancer cells and tissues. Hyperthermia is usually achieved by using microwaves to heat the tumor. Soon it was realized that in order to achieve the optimal treatment it was necessary to model the electromagnetic and thermal response of the anatomic area of interest in conjunction with the response of the region of interest to radiation therapy. This led the MFOL research team to develop detailed computational models based on the past experience of electromagnetic computations. A similar approach was adopted for the case of brain studies since the problem to be dealt in brain modeling was again of enormous complexity. Electromagnetics methods proved to be very useful and several brain response phenomena were succesfully treated.

Based on the above concepts and after having received stimuli by medical doctors and in particular by Prof. N. Zamboglou, Klinikum Offenbach, Germany as well as by the pioneering work of Prof.W. Duchting an effort started in late 1990’s with the aim to model in detail the dynamic response of cancerous cell clusters (initially in vitro and subsequently in vivo). Despite the initial negative precondition of numerous medical doctors the recognition of the fact that many practices in clinics fail because of the lack of accurate knowledge of system’s biology helped in the acceptance of the proposed detailed methods in cancer modeling. After a decade a change of mood occurred and many former unbelievers of the potential of cancer modeling began to cooperate with engineering and physics groups so that several interdisciplinary research teams are presently intensively working on the topic. The long term goal is to develop individualized treatment methods taking into account all the biological details in each patient case. There is substantiated hope that this approach will provide concrete results in achieving optimization of the therapeutic methods in the clinical environment.

II. THE RESPONSE TO RADIATION THERAPY MODELING PARADIGM

In the radiation therapy setting current treatment planning algorithms are based on the concept of physical optimization of the dose distribution and rely on rather crude biological models of tumour and normal tissue response. Such algorithms practically ignore the highly complicated dynamic behavior of malignant and normal cells and tissues. The introduction of advanced biosimulation methods based on

cell proliferation mechanisms and also on information drawn from the cellular and molecular properties of each individual malignancy and each individual patient are expected to substantially improve the radiation therapy efficiency.

This would be accomplished by using alternative fractionations, spatial dose distributions and even combination with other therapeutic modalities such as chemotherapy, hyperthermia etc. Therefore, efficient modelling, simulation and visualization of the biological phenomena taking place before, during and after irradiation is of paramount importance.

Discrete time algorithmic descriptions (simulations) of the various phenomena offer the possibility of taking into account a large number of involved mechanisms and interactions. The same philosophy has already been extensively applied to purely technological problems and the emerged numerical methods (e.g. the Finite Difference Time Domain (FDTD) technique) have proved to be very efficient and reliable.

A further prominent characteristic of the biological phenomena under consideration is stochasticity. The fate of a single irradiated cell cannot be accurately predicted for example. Only survival probabilities can be assigned to the cell based on the accumulated experimental and clinical observations made on large cell populations. Furthermore, the exact spatiotemporal distribution of the various cell cycle phases within the tumor volume is generally unknown, although some plausible macroscopic hypotheses can be made.

Therefore, stochastic techniques such as the generic Monte Carlo method seem to be particularly appropriate for the prediction of tumor growth and response to radiation therapy. The practical usefulness of such methods is both to improve understanding of the cancer behavior and to optimize the spatiotemporal treatment plan by performing in silico ( = on the computer) experiments before the actual delivery of radiation to the patient.

The clinician would be able to perform computer simulations of the likely tumor and adjacent normal tissue response to different irradiation scenarios based on the patient’s individual imaging, histologic and genetic data. The simulation predictions would support him or her in selecting the most appropriate fighting strategy. To this end a substantial number of experimental and analytical models have been developed. On the contrary a rather small number of actual three-(or four-) dimensional computer simulation models have appeared in the literature. Exploitation of the potential of current visualization techniques is even more limited. The basic philosophy is to develop detailed four


dimensional simulation models of the biological systems under consideration whereas at the same time to make use of advanced technology (e.g. visualization systems, client-server architectures, parallelization etc. )

III. CONCLUSIONS Indicative contributions of various scientific domains to the development of in silico oncology have been outlined. The importance of cross-pollination of scientific and technological ideas and approaches has been revealed through the paradigm of the MFOL’s involvement in in silico oncology. Finally, a high level categorization of simulation techniques applicable to the paradigmal case of radiation therapy has been presented and discussed.

Proceedings of the 3rd International Advanced Research Workshop on In Silico Oncology, Istanbul, Turkey. Sept. 23-24, 2008.

4

Investigation of Protein Activity Using de-novo Designed Peptides and Laser Light Radiation: the RRM Computational Predictions

E. Pirogova, T. Istivan, P. Coloe, M. Akay, senior member IEEE, I. Cosic, senior member IEEE

Abstract—In this study we have examined the specific function of cell transformation through the computational analysis of activity of oncogene and proto-oncogene proteins using the Resonant Recognition Model (RRM). We have identified two distinct features for oncogene products (transforming proteins) and proto-oncogenes products (non-transforming proteins), which are used to design a set of bioactive peptides that can mimic the transforming and non-transforming protein activity. We also discuss here the possibility of protein activation using laser radiation of the defined wavelengths [3,5]. The experimental design of testing the activity of the de-novo designed peptides on transformed (HeLa; MCF-7; and CEF/MV infected with herpes virus) and non-transformed cell cultures (HaCaT; MCF-10A; and CEC), is presented and discussed.

I. INTRODUCTION BILITY to predict the functions and three-dimensional shapes of biological molecules would certainly be

useful in designing therapeutic drugs. The structure of the drug molecule that can specifically interact with a particular bio-molecule could be modeled using computational tools. These tools allow drug molecules to be constructed using knowledge of its structure and the nature of its active site. In order to design biologically active peptides it is of primary importance to determine which amino acids are responsible for the biological activity of the native protein. The RRM was designed for

analysis of protein (DNA) interactions and their interaction with EMR [2,3]. In the RRM a protein primary structure is presented as a numerical series by assigning to each amino acid a physical parameter value relevant to the protein’s biological activity. Once the characteristic frequency for a particular protein function/interaction is identified, it is possible then to utilize the RRM approach to predict the amino acids in the protein sequence, which predominantly contribute to this frequency and thus, to the observed function, as well as to designed de novo peptides having the desired periodicities [2-4].

Manuscript received September 1, 2008. E. Pirogova is with the School of Electrical and Computer Engineering,

RMIT University, GPO Box 2476V Melbourne VIC Australia ( phone: 613-9925-3015; fax: 613-9925-2007; e-mail: [email protected]).

T. Istivan is with the Department of Biotechnology and Environmental Biology, School of Applied Sciences, RMIT University, Bundoora West, Australia (phone: 613-9925-7107; e-mail: [email protected]).

P. Coloe , Pro Vice-Chancellor, SET Portfolio Office, RMIT University, Melbourne, VIC Australia (phone: 613-9925-9518, e-mail: [email protected])

M. Akay is with the Harrington Department of Bioengineering, Fulton School of Engineering, Arizona State University Tempe Arizona 85287-9709 US (e-mail: [email protected]).

I. Cosic, Dean Research & Innovation, SET Porfolio Office, RMIT University, Melbourne VIC Australia (phone: 613-9925-9903, e-mail: [email protected]).

Oncogene proteins are a specific group of growth factors which promote uncontrolled cell growth and proliferation. These proteins derived from normal cellular growth factors (proto-oncogenes) via a limited number of modifications. Here, using the RRM methodology, we have analyzed oncogene and proto-oncogene proteins and designed de-novo peptides which would be able to modulate activity of transformed tissue. We have also determined the activation frequency of external laser radiation for oncogene and proto-oncogene proteins.

II. METHODOLOGY

The existence of frequency selective effects of light, that can affect protein function, imply that protein activation involves energies of the same order and nature as the electromagnetic radiation of light [1]. In the RRM [2,3] it is proposed that protein activities (interactions) are based on resonant electromagnetic energy transfer in a range of infra-red (IR) and visible light. The application of the RRM approach involves two stages of calculation. The first is the transformation of the amino acid sequence into a numerical sequence. Each amino acid is represented by its Electron-Ion Interaction Potential (EIIP) value which describes the average energy states of all valence electrons in a given amino acid [2,3]. Then the numerical series obtained are analyzed by digital signal analysis methods, Fourier and Wavelet transform, in order to extract information pertinent to the biological function. A multiple cross-spectral function is defined and calculated to obtain the common frequency components from the spectra of a group of proteins. Peak frequencies in such a multiple cross-spectral function denote common frequency (feature) components for all sequences analyzed. This characteristic frequency was shown to be related to protein biological

A


5

function [1-5]. It was also shown that proteins and their targets share a characteristic frequency. Thus, protein interactions can be viewed as a resonant energy transfer between the interacting molecules. This energy can be transferred through oscillations of a physical field, possibly electromagnetic in nature [2,3]. It has been found that the strong linear correlation exists between the predicted and experimentally determined frequencies corresponding to the absorption of electromagnetic radiation of light absorbing proteins [2,5]. A linear correlation between the absorption spectra of proteins and their RRM spectra with a regression coefficient of K=201 has been established [2,3]. Using the RRM postulates, a computationally identified characteristic frequency for a protein functional group can be used to calculate the wavelength of applied irradiation, λ, which assumingly would activate this protein sequence and modify its bioactivity [2,3,5]: λ=201/fRRM. The frequency range predicted for protein interactions is from 1013 Hz to 1015 Hz. This estimated range includes IR, visible and UV light. In this study we utilize this relationship to calculate the frequencies (wavelengths) of laser light radiation aiming to investigate the effects of light of defined wavelengths on cell growth and proliferation.

III. RESULTS AND DISCUSSION

We have analyzed 46-oncogene and 15 proto-oncogene sequences.

The common frequency for oncogene proteins, characterizing the ability to promote an uncontrolled cell growth and proliferation, was identified at f1=0.0322±0.023, S/N=408.77 (Fig.1a), and common frequency for proto-oncogene proteins was identified at f2=0.0566±0.067, S/N=222.39 (Fig.1b). On the basis of characteristic frequencies (f) and corresponding phases (φ) determined for oncogene and proto-oncogene proteins, we designed peptides having those spectral characteristics only [3,4]. The strategy for the design of such defined peptides is presented in [2]. As a model protein for this computational study we have chosen the Transforming protein (K-ras, Mouse): MTEYKLVVVGAGGVGKSALTIQLIQNHFVDEYDPTIEDSYRKQVVIGETCLLDILDTAGQEEYSAMRDQYMRTGEGFLCVFAINNTKSFEDIHHYREQIKRVKDSEDVPMVLVGNKCDLPSRTVDTKQAQELARSYGIPFIETSAKTRQRVEDAFYTLVREIRQYRLKKISKEEKTPGCVKIKKCVIMGVDDAFYTLVREIRKHKEKMSKDGKKKKKKSRTRCTVM

(a)

f1

f2

(b)

f2

f1

Figure 1. Multiple cross-spectral function of a) Oncogene proteins and b) ptoto-oncogene proteins. The abscissa represents the RRM frequencies, and the ordinate is the normalized intensity. Here, we present a set of the designed peptides (f1=0.0322, φ1=-0.425, and f2=0.0566, φ2=0.416) that can mimic the non-transforming protein activity of the chosen proto-oncogene: (A): QTRDDDDRTQWKPVNLNEPHAYWWWYYAKKKK (B): WMRDDDDRCWAPELIEHA (C): RRDDDDDDDRTMQWAKPEEIIINEPHAYWQTR (D): TRDDDDDRTMQQWQQQSCTTCQWYKPEILGPK (E): RDDDDRQWKPGLNPKYQR (F): YWQCRRDDDDDDDRFCQWYKHPENLINEPHKY

Using the ration λ=201/fRRM , we have determined the activation frequency of external laser light radiation for oncogene (λ1=6,242 nm) and proto-oncogene (λ2=3,551 nm) proteins that assumingly will modulate their bioactivity and thus, affect the cell activity. We will irradiate transformed (HeLa; MCF-7; and CEF/MV infected with herpes virus) and non-transformed cell cultures (HaCaT; MCF-10A; and CEC) with the computationally predicted oncogene (λ1=6,242 nm) and non-oncogene (λ2=3,551 nm) wavelengths of laser light. The changes in cell viability/morphology after this treatment will be assessed by phase contrast microscopy and confocal microscopy. Fluorescent microscopy will be used with fluorescent tagged peptides. The intracellular stability and migration, and binding of the introduced peptides will be determined and the effect of the peptides on metabolism and growth of the cells will be quantified. The efficiency of the


6

peptides alone, or bound to gold nano-particles to enhance cellular uptake will be determined.

IV. CONCLUSION

The approach presented here is intended to investigate cell transformation from the perspective of the energy/field point of view on protein activation. The approach is based on the RRM and can be used in design of bioactive peptides for oncogene and proto-oncogene proteins as well as use of radiation at the specific electromagnetic frequencies to affect protein function/cell transformation. This approach present a novel technology of modulating protein/cell biological activity by exposure to electromagnetic fields, resulting in a significant cost saving and an improvement in current biotechnology, cancer research and quality of new biomaterials.

REFERENCES

[1] T. Karu, 1999, “Primary and Secondary Mechanisms of Actions of Visible to Near-IR Radiation on Cells”, J. Photochem. Photobiol, Biol 49:1-17.

[2] I. Cosic, 1994, “Macromolecular bioactivity: Is it Resonant Interaction between Molecules? - Theory and Applications”, IEEE Trans. on BME. 41:1101-1114.

[3] I. Cosic, The Resonant Recognition Model of Macromolecular Bioactivity, Birkhauser Verlag, 1997.

[4] E. Pirogova, Q. Fang, M. Akay, I. Cosic, 2002, “Investigation of the structure and function relationships of Oncogene proteins”, Proceeding of the IEEE, Vol. 90, No. 12:1859-1867.

[5] Vojisavljevic, V., Pirogova, E., Cosic, I. (2007) “The effect of Electromagnetic Radiation (550nm-850nm) on l-Lactate Dehydrogenase Kinetics”, Int. Journal Radiation Biology, 83(4), 221-230.


7

In Silico Oncology: a Paradigm for Clinically Oriented Living Matter Engineering

Georgios S. Stamatakos

Abstract— The emerging scientific, technological and medical domain of in silico oncology, primarily aiming at supporting the optimization of cancer treatment in the patient individualized context, is seen as a paradigm for living matter information engineering. An outline of this approach is provided along with a brief overview of related work done by the In Silico Onolcogy Group, ICCS, National Technical University of Athens. The concept of the onco-simulator, a centrally positioned system of in silico oncology, is briefly introduced. Current international activities related to the domain are cited.

I. INTRODUCTION CCUMULATION of multilevel experimental and clinical information pertaining to cancer is increasing at a

remarkably high rate. In order to thoroughly exploit such information at all levels of biocomplexity - from the atomic up to the ecosystem level - new information processing strategies have to be developed, clinically validated and eventually translated into clinical routine. Due to the fact that cancer apart from a disease is a hypercomplex natural phenomenon, such strategies should refer to both space and time and efficiently capture the dynamics of the corresponding biological systems. The latter dictates the recruitment of several engineering methods such as reaction-diffucion modeling, four dimensional numerical techniques based on discretization meshes, discrete state modeling etc. Within the context of both modeling and simulating biosystems at multiple levels of biocomplexity a new form of engineering seems to emerge: living matter information engineering. A fundamendal feature of such an engineering discipline should be the ability to describe living systems at

many biocomplexity levels concurrently while pretty different methods and approaches are applicable to each level separately. Furthermore, there must be continuous passing of “distilled” information between biocomplexity levels and subsequent updating of the state of each level. Despite the high usability of techniques originally developed for non living matter, completely new concepts, techniques and algorithms have to be developed in order to tackle the particularities of living matter (e.g. gene-protein interactions, cell cycling, tissue behavior their spatiotemporal interdependences etc.). Within this framework in silico oncology holds a paradigmal place. In silico oncology is a complex and multiscale combination of sciences and technologies intending to simulate malignant tumor growth and tumor and normal tissue response to therapeutic modalities at all levels of biocomplexity. Its practical goal is to quantitatively understand cancer and related phenomena and optimize therapeutic interventions by performing in silico (on the computer) experiments based on the individual patient’s clinical, imaging, histopathological, molecular and pharmacogenomic data. It should be specially stressed that in order to achieve such an ambitious goal translation of cancer models into the clinical trials arena is a sine qua non condition.

Manuscript received September 1, 2008. G. S. Stamatakos (corresponding author) is with the National Technical

University of Athens, Institute of Communications and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (phone: + 30 210 772 2288, fax: + 30 210 772 3557, e-mail: [email protected] ).

This work was supported by the European Commission under the project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996), http://www.eu-acgt.org/.

In this context In Silico Oncology Group (ISOG), a research module of the Microwave and Fiber Optics Laboratory of ICCS, National Technical University of Athens [1] has developed a number of tumor and treatment affected normal tissue simulation models over the past 11 years [2]-[7]. The models concern in vitro and in vivo tumor growth, the response of both in vitro and in vivo types of tumors to radiotherapeutic and chenotherapeutic schemes as well as the response of normal tissues to radiotherapeutic schedules. Currently ISOG is leading the development of the Oncosimulator i.e. an advanced information system able to simulate the response of tumors and affected normal tissues to therapeutic schemes based on the individual multiscale data of a given cancer patient. This action is taking place within the framework of the European Commission (EC) funded ACGT (Advancing Clinicogenomics Trials on Cancer) project [8]. The Oncosimulator aims at optimizing

A


8

cancer treatment on a patient-individualized basis by performing in silico (on the computer) experiments of candidate therapeutic schemes. In parallel ISOG is leading the development of the integrated onco-simulator of the recently started EC funded project ContraCancrum: Clinically Oriented Transaltional Cancer Multilevel Modeling. It is also participating in the EC Virtual Physiological Human Newtork of Excellene [9] and in the US based Center for the Development of a Virtual Tumor[10] funded by NCI-NIH. ISOG is a coorganizer of the First Transatlantic Workshop on Multiscale Cancer Modeling [11].

II. THE ONCOSIMULATOR Following its thorough clinical validation, currently taking

place within the ACGT framework, the Oncosimulator will function as follows:

A. First Step: Obtain Patient’s Specific Data The following sets of data is collected for each patient: • Clinical data (age, eventual previous treatments etc.) • Imaging data (images of MRI, ultrasound, PET, CT etc.) • Histopathological data (histopathology slide images

whenever biopsy is allowed and feasible) • Molecular data (specific molecular marker values and/or

DNA array data based on biopsy and/or blood samples) B. Second Step: Preprocess Patient’s Data The data collected is preprocessed in order to take an

adequate form allowing its introduction into the Oncosimulator system. For example the imaging data are segmented, registered, interpolated, 3-D reconstructed. Similarly the molecular data are combined via molecular interaction networks in order to perturb the average pharmacodynamic or radiobiological cell survival parameters and so on.

C. Third Step: Describe CandidateTherapeutic Schemes The clinician describes a number of candidate therapeutic

schemes to be simulated in silico i.e. on the computer D. Fourth Step: Run the Simulation The tumor growth and therapy response computer code is

executed on distributed GRID computational resources so that several candidate treatment schemes incorporating many possible unknown tumor parameter values combinations are

simulated concurrently. Predictions concerning the toxicological acceptability of each candidate treatment scheme are also produced.

E. Fifth Step: Visualize the Predictions The expected reaction of the tumor as well as indications

of the toxicological side effects for all scenarios simulated are visualized using several techniques ranging from graph plotting to virtual reality rendering.

F. Sixth Step: Evaluate the Predictions and Decide on the

Optimal Scheme to be Applied The Oncosimulator’s predictions are carefully evaluated by

the clinician by taking into account their logic, education and even experience. If no serious conflicts are detected, the predictions can be used to support the clinician in taking their final (expectedly optimal) decision on the actual treatment to be administered to the patient.

G. Seventh Step: Apply the Optimal Therapeutic Scheme

and Further Optimize the Oncosimulator The expectedly optimal therapeutic scheme (schedule) is

applied on the patient. In parallel the prediction versus reality comparison data is collected and used as a continuous optimization feedback to the Oncosimulator.

Other envisaged application areas of the Oncosimulator

include: • Basic science (dynamic integration of multilevel biodata

and biomechanisms, in silico experimentation) • Design of new clinicogenomic trials • Medical education • Education of interested patients and/or parents

III. CONCLUSION In silico oncology has been viewed as a paradigm of living

matter information engineering. The concept of the onco-simulator, a central simulation system of in silico oncology, has been briefly outlined. Examples of large scale international projects related to the domain have been provided. It is becoming clear that the hypercomplexity of cancer and treatment affected normal tissues in conjunction with the particularly high interdisciplinarity of the cancer simulation endeavor dictate a global deployment of the development of in silico oncology. This implies that large research and development teams consisting of engineering, basic science and clinical specialists are to be formed if an


9

analytical, manageable and clinically exploitable understanding of cancer is to be achieved. Development and clinical translation of onco-simulators is a concrete example of such an exploitation. Last but not least it should be stressed that a thorough clinical validation of any onco-simulator through carefully designed clinical trials is a sine qua non necessity.

REFERENCES [1] http://www.in-silico-oncology.iccs.ntua.gr/ [2] G.S.Stamatakos, D.D.Dionysiou, E.I.Zacharaki, N.A.Mouravliansky,

K.Nikita, N.Uzunoglu, “In silico radiation oncology: combining novel simulation algorithms with current visualization techniques”, Proceedings of the IEEE, vol. 90, No 11, pp.1764-1777, Nov. 2002.

[3] D. D. Dionysiou, G. S. Stamatakos, N.K. Uzunoglu, K. S. Nikita, A. Marioli, “A four-dimensional simulation model of tumour response to radiotherapy in vivo: parametric validation considering radiosensitivity, genetic profile and fractionation,” Journal of Theoretical Biology 230 (2004) 1–20

[4] G. S. Stamatakos, V. P. Antipas, N. K. Uzunoglu “A Spatiotemporal, Patient Individualized Simulation Model of Solid Tumor Response to Chemotherapy in Vivo: The Paradigm of Glioblastoma Multiforme Treated by Temozolomide,” IEEE Trans. Biomedical Engineering, vol. 53, No.8, pp. 1467-1477, 2006.

[5] G. Stamatakos, D. Dionysiou, and N. Uzunoglu, “In Silico Radiation Oncology: A Platform for Understanding Cancer Behavior and Optimizing Radiation Therapy Treatment,”in M.Akay Ed. “Genomics and Proteomics Engineering in Medicine and Biology,” Wiley-IEEE Press, 2007

[6] G.S.Stamatakos, D.D.Dionysiou, N.M.Graf, N.A.Sofra, C.Desmedt, A.Hoppe, N.Uzunoglu, M.Tsiknakis , The Oncosimulator: a multilevel, clinically oriented simulation system of tumor growth and organism response to therapeutic schemes. Towards the clinical evaluation of in silico oncology , Proceedings of the 29th Annual International Conference of the IEEE EMBS, Cite Internationale, August 23-26, SuB07.1: 6628-6631 , Lyon, France , 2007

[7] N.Graf,C.Desmedt,F.Buffa,D.Kafetzopoulos,N.Forgo,R.Kollek,A.Hoppe,G.Stamatakos,M.Tsiknakis Post-genomic clinical trials - The perspective of ACGT, Ecancermedicalscience Vol. 2, 2008

[8] http://eu-acgt.org/acgt-for-you/researchers/in-silico-oncology.html [9] http://www.vph-noe.eu/ [10] https://www.cvit.org/ [11] http://ec.europa.eu/information_society/events/ict_bio/2008/ta-cancer-

wkshp/index_en.htm


10

A Clinical View on In Silico Oncology

Alexander W. Hoppe

Abstract— Over the last years the role of medical practitioners has become a mixed blessing between the daily clinical work and the knowledge out of an explosively expanding world of medical facts and research. Today, the increasing amount of data and findings from multi-level based research (from genetics to epidemiology) is at risk by missing the clinical adaptability and reliability. At the same time the proceedings in bioinformatics and computational science offer opportunities to be the “connecting piece” in the search to use the findings and realize an individual treatment for patients. One of the major players in the new developments and approaches to fight cancer is to understand and effectively model the dynamics of cancer and treatment affected normal tissues at all biocomplexity levels by using any efficient combination of mathematical and computer modeling approaches. At the end these models have to be provided in a simple and applicable form for daily clinical work.

I. INTRODUCTION HE document Hippocratic Oath sworn by physicians receiving the licence to practice medicine says:“ I will

respect the hard-won scientific gains of those physicians in whose steps I walk, and gladly share such knowledge as is mine with those who are to follow.“ It becomes more and more difficult today to fulfil this demand, as the findings are multi-scaled and received from a number of scientific fields aside of medicine while efficient ways to combine and convert from research to daily practice is difficult to realize. Still the direct benefit for the patient from these research results is missing most of the time today.

Good clinical practice, Evidence-based medicine, randomized, controlled trials and documentation, treatment algorithms, administrative work and in the first line the emphatic and ethical work in line with legal directives are the principle items in daily oncologists work.

Always focused on patients’ needs and health, based on the basic principle to offer them the best possible treatment and to respect the often life-threatening situation they are facing.

During the last decades conventional therapies like radiotherapy, chemotherapy, newly developed drugs and optimized treatment regimes enabled significant increase of the overall survival in cancer patients. Today our

understanding of cancer biology, tumour growth, invasion, and metastases; the evasion of apoptosis; self-sufficiency in growth signalling; insensitivity to antigrowth signals; sustained angiogenesis; and limitless replicative potential allows the identification of “drugable targets” [1]. Some of these target pathways are well identified and drugs that aim to these pathways are approved and used in clinics.

Manuscript received September 1, 2008. A.W. Hoppe is with University Hospital of Saarland, Paediatric

Haematology and Oncology, D – 66421 Homburg Germany (phone: 004968411628045, email:[email protected])

An example on how biological findings can be translated to daily practice is the pathway of the epidermal growth factor receptor (EGFR). Activation of the EGFR pathways leads to in-creased cellular proliferation, motility, adhesion, invasion, angiogenesis, and inhibition of apoptosis [2]. Agents have been developed that inhibit this pathway by binding to the extracellular domain of the receptor, as in the case of monoclonal antibodies or by binding to the intracellular tyrosine kinasis. After all there are several mutations in the pathways. Patients whose tumours have an activating mutation in the tyrosine kinase domain and good respond to EGFR tyrosine kinase inhibitors have a therefore prolonged survival. But the percental fraction carrying this activating mutation is approximately 10% [3]. Moreover secondary mutations were identified in patients who developed resistance towards these target drugs [4].

Currently cancer treatment decision and planning are mainly based on a large extent on disease behaviour, and not on the behaviour of all individual tumour characteristics, of the statistically “mean value” patient and clinical trials in which genetic and molecular individuality of tumours are not respected. To improve cancer treatment it is essential to take the characteristics of a tumour in account. Therefore critical details of the particular patient’s tumour biology in combination with new imaging methodologies and techniques like 4 DMT [5], patients’ individual pharmacokinetic and metabolism up to epidemiology need to be related and simulated in an usable and efficient way for the physician to support him by finding the best possible therapy decision.

Genetic research, microarray analysis and immune response analysis belong to the major fragments to characterise tumours and tumour development. The immune system responds to exogenous structures in several ways. Towards tumours it responds either by reacting against tumour-specific antigens (surface structures that are typical to cancer cells) or against tumour-associated antigens (molecules that are expressed differently by cancer cells and normal cells) [6].

T


11

The use of Microarray analysis, the availability of sequenced genomes in combination with the development of new technologies and powerful computational tools that are capable of exploiting these genomic data [7] accelerate and offer new insights in the tumour characteristics and biology and drug development. All these findings and achievements are important to treat a patient individually by taking all characteristics of his disease into account; research findings that are not possible to be used efficiently in daily work today.

To make the multi-level findings usable for daily medical work, computational systems and models that combine the results (e.g. from cancer biology, radiobiology and pathology, cancer bioinformatics, anatomy, clinical and radiotherapeutic oncology, medical physics, pharmacology, image processing, visualization technologies, discrete and continuous mathematics, probability theory, statistics, epidemiology) automatically are needed. These models can build the basis to predict the tumour response respecting all characteristics of the individual tumour and simulate the tumour behaviour by altering parameters like drugs, drug combinations, drug dose, time points of administration, response to radiotherapy or newly developed “target” therapies. Therefore it is essential to develop research platforms allowing seamless integration and comparability from different sources for reliable modelling.

The EC funded project “Advancing clinico-genomic trials on cancer ACGT” (www.eu-acgt.org) is currently developing such a grid-based platform, based on the idea of seamless integration of basic research data in combination with clinical data from clinical trials and the integration and simulation of data for the prediction of tumour behaviour. [8].Within ACGT multiple data of real patients with breast cancer and nephroblastoma are collected, anonymized patient individualised tumour growth as well as tumour and normal tissue response simulation models are developed, with a frequent validation of the proceedings to realise four dimensional, patient specific computer simulation models of the biological behaviour and their changes at treatment modalities [9].

The first evaluations and simulation results gained in the ACGT project from the “Oncosimulator” are highly promising.

Important for the realisation and collection of the critical data was the implementation of specially developed scenarios in clinico-genomic trials. To assure completeness of the data the trial protocols and patients informed consents were extended and clinical and research findings are now stored together centrally in a single data management system, providing access to the anonymized data for further research.

It is obvious that the development of such simulations does not need strict clinical validation because of

correctness and legal requirements, but to provide simulations trusted and accepted by physicians.

If the confidence towards correctness and safety of simulations is missed by clinicians, the simulation models will fail clinical utilisation.

Many factors determine the usability for daily clinical work. The simulations will be based on an amount of parameters and complex mathematic calculations and computational procedures that make it impossible for the physician to evaluate its correctness, yet only judge its plausibility. The clinician is expecting to have a “black box” where the data of the patient will be entered and by changing a defined set of parameters the response of the tumour will be visualised and hence support the physician in finding the optimal therapy for the individual patients’ situation. At the end In Silico Oncology will provide an additional interpretation of another sentence found in the Hippocratic Oath saying: „I will apply, for the benefit of the sick, all measures that are required, avoiding those twin traps of over treatment and therapeutic nihilism.”

II. CONCLUSION The approaches and first simulations in In Silico

Oncology appear to be an effective future weapon in the fight against cancer. Taking into account sets of multilevel parameters of tumours, patients’ characteristics, computational simulation of tumours, behaviour and treatment response, it is on the way to be a powerful application for physicians to support them in finding the right therapy decision for the “individual” patient and is expected to dramatically accelerate the achievement of cancer cure on a patient individualized basis through treatment optimization.

REFERENCES [1] Schiller, Joan H., M.D. “Non-invasive Monitoring of Tumors.” N

Engl J Med, vol. 359, pp. 418-420, 2008. [2] P.A. Jänne, J.A. Engelman, B.E. Johnson. “Epidermal growth factor

receptor mutations in non-small-cell lung cancer: implications for treatment and tumor biology, ”J Clin Oncol, vol. 23, pp. 3227-3234, 2005.

[3] E. Avizienyte, R.A. Ward, A.P. Garner. ”Comparison of the EGFR resistance mutation profiles generated by EGFR targeted tyrosine kinase inhibitors and the impact of drug combinations.,” Biochem J. Jun 30, 2008.

[4] T.J. Lynch, D.W. Bell, R. Sordella, et al. “Activating mutations in the epidermal growth factor receptor underlying responsiveness of non–small-cell lung cancer to gefitinib.” N Engl J Med, vol. 350, pp. 2129-39, 2004.

[5] G. Li, D. Citrin, K. Camphausen, B. Mueller, C. Burman, B. Mychalczak, R.W. Miller, Y. Song, “Advances in 4D medical imaging and 4D radiation therapy.” Technol Cancer Res Treat., vol. 7(1), pp. 67-81, 2008.

[6] D.F. Graziano, O.J. Finn. “Tumor antigens and tumor antigen discovery.” Cancer Treat Res, vol. 123, pp. 89-111, 2005.

[7] C.V. Murty, V.M. Chennathukuzhi, D.S. Johnston, P.E. Stevis, G.S. Kopf. “Gene Expression Profiling and its Practice in Drug Development,” Genomics Curr,, vol. 8, pp. 262-270, 2007.


12

[8] N. Graf, C. Desmedt, F. Buffa, M. Kafezopoulos, N. Forgó, R. Kolleg, A. Hoppe, G. Stamatakos, M. Tsiknakis, Post-genomic clinical trials – the perspective of ACGT,” Vol. 2, Article Number 66, 2008.

[9] G.S. Stamatakos, D.D Dionysiou, N.M. Graf, N.A. Sofra, C.Desmedt, A.Hoppe, N.K. Uzunoglu, M. Tsiknakis, “The "Oncosimulator": a multilevel, clinically oriented simulation system of tumor growth and organism response to therapeutic schemes. Towards the clinical evaluation of in silico oncology.” EMBS, 22-26 Aug., pp. 6628 – 6631, 2007.


13

Glioblastoma Multiforme Response to Radiotherapy: Critical Parameters and In Silico Trials

Dimitra D. Dionysiou and Georgios S. Stamatakos

Abstract—. The aim of this paper is to demonstrate characteristic research applications of the In Silico Oncology Group (ISOG) imageable glioblastoma multiforme response to radiotherapy model. The model can be used in order to investigate the most critical parameters determining radiotherapy treatment outcome in terms of tumor cell kill for glioblastoma multiforme tumors. It can be also used to simulate clinical trials, with the long-term goal of both better designing clinical studies and understanding their outcome based on basic biological science.

I. INTRODUCTION ADIATION therapy holds a central position in the management of cancer. Cancer biology plays a critical

role in determining radiotherapy treatment outcome. In this context, efficient modeling of tumor response to radiotherapy can contribute to the elucidation of the involved biological mechanisms and to the emergence of patient-individualized treatments, on the presupposition of adequate clinical adaptation and testing.

The aim of this paper is to demonstrate how multilevel tumour growth and response to therapeutic treatment models can be used in order to perform characteristic research studies, such as exploring the influence of critical parameters on the response to radiotherapy and reproducing the results of clinical trials. conclusions

The comparison of the simulation results with clinical experience and experimental knowledge was selected as a means to reveal and substantiate the potential and flexibility of the particular simulation model in order to study biological phenomena related to cancer and in the long-term serve as a patient individualized treatment optimization tool, following a strict clinical validation

procedure.

Manuscript received September 1, 2008. D. D. Dionysiou is with the In Silico Oncology Group, Institute of

Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: +30 210 772 2288; e-mail: [email protected]).

G. S. Stamatakos is with the In Silico Oncology Group, Institute of Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: + 30 210 772 2288, e-mail: [email protected] ).

This work was supported by the European Commission under the project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996), http://www.euacgt. org/.

II. CRITICAL PARAMETERS DETERMINING GLIOBLASTOMA MULTIFORME RESPONSE TO RADIOTHERAPY

The In Silico Oncology Group (ISOG) tumor response to radiotherapy model [1]-[3] has been used to perform an investigation of the relative impact of the most critical parameters determining radiotherapy treatment outcome in terms of tumor cell kill for glioblastoma multiforme (GBM) tumors by performing adequate simulations. For each of the selected critical parameters a series of simulation runs covering the whole range of values that have appeared in the literature for GBM have been performed, while adjusting the remaining parameters at the most typical GBM values.

A thorough study of glioblastoma multiforme literature preceded the simulations, so as to define the real clinical range of the values of various model parameters [4]-[28]. The parameters selected for the study of their relative impact in GBM’s response, based on accumulated clinical and experimental knowledge, were the α and β radiosensitivity parameters of the Linear-Quadratic (LQ) model, the cell cycle duration (TC) and the cell loss factor (CLF) of the tumor [4]-[5].

Figures 1-2 depict characteristic simulation results in the form of the number of living tumor cells as a function of time. The time point t=0 corresponds to the start of the radiotherapy treatment. The standard fractionation scheme of 2Gy/day, 5 days/week, 60Gy in total, no irradiation during weekends, was used in all simulation cases.

According to the results obtained for the parameter ranges considered, the parameters with the major impact on the tumor’s response to radiotherapy were the alpha parameter of the LQ model and the cell cycle duration of tumor cells, while the effect of the other parameters was less pronounced. More specifically, starting with an otherwise identical tumor in all cases, by varying the values of the alpha parameter the resulting number of living tumor cells one week after the end of treatment (t=6 weeks) spans an interval of approximately 10 orders of magnitude (logs) (Figure 1). The corresponding interval in the cases of TC (Figure 2), beta and CLF parameters is approximately of the order of 7, 2 and 1.5 logs correspondingly.

R


14

III. AN IN SILICO TRIAL Multilevel tumour growth and response to therapeutic

treatment models can be used in order to simulate clinical trials, with the long-term intention of both better designing clinical studies and understanding their outcome based on basic biological science. For this purpose, the ISOG simulation model of GBM response to radiotherapy has been used and a clinical study concerning GBM response to radiotherapy has been simulated. In order to facilitate the simulation of such virtual trials, a toolkit enabling the user-friendly execution of the simulations on grid infrastructures has been designed and developed. The results of the conducted virtual trial are in agreement with the outcome of the real clinical study.

The Radiation Therapy Oncology Group (RTOG) 83-02 clinical study [6] has been selected to serve as a realistic paradigm of clinical trials in which modeling could play a useful role; a series of simulations corresponding to its various arms have been performed. This was a randomized Phase I/II study of escalating doses for Hyperfractionated radiotherapy (HF, 1.2Gy twice daily to doses of 64.8, 72, 76.8, or 81.6Gy) and Accelerated Hyperfractionated radiotherapy (AHF, 1.6Gy twice daily to doses of 48 or 54.4Gy) with carmustine (BCNU) for adults with supratentorial GBM or anaplastic astrocytoma. The study has revealed that GBM patients who received the higher HF doses had survival superior to the patients in the AHF arms or lower HF doses.

Fig. (1). Number of living tumor cells as a function of time from the start of radiotherapy treatment (at t=0), for hypothetical GBM tumors with different values of the alpha parameter of the LQ model. α is in Gy-1.

In order to create the in silico counterpart of the real study cohort, the study comprised a total of 462 radiotherapy simulations based on the 6 radiotherapeutic schemes of the RTOG 83-02 clinical study. 11 values of the cell cycle duration and 7 pair values of the radiosensitivity α and β parameters of the LQ model have been considered.

Fig. (2). Number of living tumor cells as a function of time from the start of radiotherapy treatment (at t=0), for hypothetical GBM tumors with different values of the cell cycle duration.

Fig. (3) The number of living tumour cells as a function of time from the start of the radiotherapy treatment for a hypothetical case of a GBM tumour with TC=40h and (α,β) = (0.31Gy-1, 0.04Gy-2). HF: HyperFractionation, AHF: Accelerated Hyperfractionation. For each scheme the total dose is given.

Fig. (4) The number of alive tumour cells as a function of time from the start of the radiotherapy treatment for hypothetical cases of GBM tumours with the same cell cycle duration, TC=40h, but differing in their cellular radiosensitivity and treated according to the HF 72Gy schedule. α is in Gy-

1, β is in Gy-2.

For the cell cycle duration a parameter range has been identified (20h-120h), whereas for the LQ model pair (α,β)


15

a number of experimentally determined values have been considered reflecting specific molecular profiles. The parameter values considered seem to cover the whole range of values of TC and (α, β) that have been reported in literature for GBM and therefore are assumed to reflect GBM interpatient variability. All combinations of TC and (α,β) pair values have been considered. All other parameters of the simulation model (e.g. cell density, cell loss factor, oxygen enhancement ratio etc.) have been selected so as to reflect a typical value for GBM tumors as dictated in the relevant literature.

The simulation results are in agreement with the results of the RTOG 83-02 clinical study, as they reveal that for GBM radiotherapy the use of high-dose hyperfractionation schemes is advantageous in terms of tumour cell kill compared to low-dose hyperfractionation schemes or accelerated hyperfractionation schemes. This holds true for the whole range of parameter values that have been tested and is therefore interpreted as being valid for GBM tumours in general, under the assumption mentioned before that the values of the parameters used cover the whole spectrum of GBM tumours. Indicative results drawn from the whole series of simulations are presented in Fig. 3and 4. In Fig. 3 the clear advantage of HF schedules and in particular high-dose HF schedules is evident for a hypothetical case of a moderately radiosensitive GBM tumour with TC=40h and (α,β) = (0.31Gy-1, 0.04Gy-2). In Fig. 4 the comparative tumor cell kill effect of the HF 72Gy radiotherapy scheme for hypothetical clinical cases differing in their cellular radiosensitivity is depicted. The increased cell survival in the more radioresistant tumours is evident.

REFERENCES [1] G.S. Stamatakos, D.D. Dionysiou, E.I. Zacharaki, N.A.

Mouravliansky, K.S. Nikita, N.K. Uzunoglu. “In Silico Radiation Oncology: Combining Novel Simulation Algorithms With Current Visualization Techniques.” Proc. of the IEEE, vol. 90, pp. 1764-1777, Nov. 2002.

[2] D.D. Dionysiou, G.S. Stamatakos, N.K. Uzunoglu, K.S. Nikita, A. Marioli. “A four-dimensional simulation model of tumour response to radiotherapy in vivo: parametric validation considering radiosensitivity, genetic profile and fractionation.” J theor. Biol., vol. 230, pp. 1-20, Sept.2004

[3] D.D. Dionysiou, G.S. Stamatakos, N.K. Uzunoglu, K.S. Nikita. “A computer simulation of in vivo tumour growth and response to radiotherapy: New algorithms and parametric results.” Comp. Biol. Med., vol. 36, pp. 448-464, May 2006.

[4] G. Steel. Basic Clinical Radiobiology. London: Arnold, 2002. [5] C. Perez, L. Brady L. Principles and Practice of Radiation Oncology.

Philadelphia: Lippincott-Raven, 1998. [6] M. Werner-Wasik, C.B. Scott, D.F. Nelson, L.E. Gaspar, K.J.

Murray, J.A. Fischbach, J.S. Nelson, A.S. Weinstein, W.J. Curran. “Final report of a Phase I/II trial of hyperfractionated and accelerated hyperfractionated radiation therapy with carmustine for adults with supratentorial malignant gliomas: Radiation therapy oncology group study 83-02.” Cancer, vol. 77, pp. 1535-1543, 1996.

[7] J. Denekamp. “Cell kinetics and radiation biology.” Int. J Radiat. Biol.,vol. 49, pp. 357-380, 1986.

[8] D.A. Haas-Kogan, G. Yount, M. Haas, D. Levi, S.S. Kogan, L. Hu, C. Vidair, S.F. Deen, W.C. Dewey, M.A. Israel. “p53-dependent G1 arrest and p53 independent apoptosis influence the radiobiologic response of glioblastoma.” Int J Radiat. Oncol. Biol. Phys., vol. 36, pp. 95-103, Aug. 1996.

[9] K. Jellinger. “Glioblastoma multiforme: morphology and biology.” Acta Neurochir., vol. 42, pp. 5-32, Mar. 1978.

[10] B. Hegedues, A. Czirok, I. Fazekas, T. Babel, E. Madarasz, T. Viscek. “Locomotion and proliferation of glioblastoma cells in vitro: statistical evaluation of videomicroscopic observations.” J Neurosurg., vol. 92, pp. 428-434, Mar. 2000.

[11] T. Hoshino. “Cell kinetics of glial tumors.” Rev Neurol (Paris), vol. 148, pp. 396-401, 1992.

[12] L.E. Dillehay. “A model of cell killing by low dose rate radiation including repair of sublethal damage, G2 block, and cell division.” Rad. Res., vol. 124, pp. 201-207, 1990.

[13] T. Hoshino, C.B. Wilson. “Cell kinetic analyses of human malignant brain tumors (gliomas).” Cancer, vol 44, pp. 956-962, 1979.

[14] D.C. Crafts, T. Hoshino, C.B. Wilson. “Current status of population kinetics in gliomas.” Bulletin Du Cancer, vol. 64, pp. 115-124, 1977.

[15] T. Hoshino, C.B. Wilson, M.L. Rosenblum, M. Barker. “Chemotherapeutic implications of growth fraction and cell cycle time in glioblastomas.” J Neurosurg., vol. 43, pp. 127-135, 1975.

[16] J. Markert, V.T. DeVita, S.A. Rosenberg, S. Hellman. Glioblastoma Multiforme. USA: Jones and Bartlett Publishers, 2005.

[17] H. Watanabe, M. Miura, T. Sasaki. “Differential effects of the Insulin-Like Growth Factor I Receptor on radiosensitivity and spontaneous necrosis formation of human glioblastoma cells grown in multicellular spheroids.” Exp. Cell Res., vol. 250, pp. 99-111, July 1999.

[18] W. Budach, D. Gioioso, A. Taghian, M. Stuschke, H. D. Suit. “Repopulation capacity during fractionated irradiation of squamous cell carcinomas and glioblastomas in vitro.” Int J Radiat. Oncol. Biol. Phys., vol. 39, pp.743-750, Oct. 1997.

[19] H. Schmidberger, M. Rave-Fraenk, J. Lehmann, E. Weiss, L. Gerl, N. Dettmer, S. Glomme, C. F. Hess. “Lack of interferon beta-induced radiosensitization in four out of five human glioblastoma cell line.” Int. J Radiat. Oncol. Biol. Phys., vol. 55, pp. 1348-1357, Apr. 2003.

[20] J.R. Williams, Y. Zhang, J. Russell, C. Koch, J.B. Little. “Human tumor cells segregate into radiosensitivity groups that associate with ATM and TP53 status.” Acta Oncol, vol. 46, pp. 628-638, 2007.

[21] P. Huang, A. Allam, L.A. Perez, A. Taghian, J. Freeman, H.D. Suit. “The effect of combining recombinant human tumor necrosis factor-alpha with local radiation on tumor control probability of a human glioblastoma multiforme xenograft in nude mice.” Int J Radiat Oncol Biol Phys., vol. 32, pp. 93-98, Apr. 1995.

[22] L. Hlatky, M. Olesiak, P. Hahnfeldt. “Measurement of potential doubling time for human tumor xenografts using the cytokinesis-block method.” Cancer Res., vol. 56, pp. 1660-1663, Apr. 1996.

[23] L.A. Perez, D. Dombkowski, J. Efird, F. Preffer, H.D. Suit. “Cell proliferation kinetics in human tumor xenografts measured with iododeoxyuridine labeling and flow cytometry: a study of heterogeneity and comparison between different methods of calculation and other proliferation measurements.” Cancer Res., vol. 55, pp. 392-398, Jan. 1995.

[24] M. Nakajima, S. Nakasu, S. Morikawa, T. Inubushi. “Estimation of volume doubling time and cell loss in an experimental rat glioma model in vivo.” Acta Neurochir. (Wien), vol. 140, pp. 607-613, Jul. 1999.

[25] D.T. Yew, H.H. Wang, D.R. Zheng. “Apoptosis in astrocytomas with different grades of malignancy.” Acta Neurochir (Wien), vol. 140, pp. 341-347, May 1998.

[26] S. Tribius, A. Pidel, D. Casper. “ATM protein expression correlates with radioresistance in primary glioblastoma cells in culture.” Int J Radiat Oncol Biol. Phys., vol. 50, pp. 511-523, Jun. 2001.

[27] F. Giangaspero, C. Doglioni, M.T. Rivano, S. Pileri, J. Gerdes, H. Stein. “Growth fraction in human brain tumors defined by the monoclonal antibody Ki-67.” Acta Neuropathol., vol. 74, pp. 179-182, Jun. 1987.


16

[28] N. Laperriere, L. Zuraw, G. Cairncross. “The cancer Care Ontario Practice Guidilines Initiative Neuro-oncology Disease Site Group “Radiotherapy for newly diagnosed malignant glioma in adults: a systematic review”.” Rad Oncol., vol. 64, pp. 259-273, 2002.


17

Modeling Growth Saturation in Avascular Tumors Bryn A. Lloyd, Member, IEEE, Gábor Székely, Associate Member, IEEE

Abstract—We comment on a commonly used oxygen diffusion boundary condition in tumor models. The Dirichlet boundary condition at the tumor boundary is unrealistic since it implies that the surrounding tissue is not affected by the presence of the neoplasm. We build an alternative model, which does not assume constant oxygen at the interface and, therefore, predicts lower oxygen tension in the outer layer of the tumor and hence decreased proliferation. This seemingly minor change to the model has consequences on the cell population dynamics and offers a simple alternative to current explanations of the often described phenomenon of growth saturation.

I. INTRODUCTION N the early phases of tumor development most tumors growing in vivo develop as avascular cell clusters, relying

on diffusion from the neighboring healthy vascularized tissue for the supply of oxygen and nutrients and the removal of wastes. Often, these tumors do not grow beyond a certain size and can stay dormant for many years.

The paper of Ward and King [1] analyzes growth saturation and presents a model, which can account for this phenomenon. Their model assumes that nutrients diffuse and are consumed by the cells, which proliferate or die depending on the availability of nutrients (oxygen, glucose, etc.). The model is defined only inside the tumor and assumes constant nutrient concentration at the tumor boundary. The analysis of Ward and King demonstrates that based on these assumptions, growth saturation is only possible if the necrotic material is transported outside of the tumor. This model is representative of many others, in the sense that it assumes a direct influence of the nutrient availability on the proliferative strength of the tumor cells. It is also common to assume constant oxygen levels at the tumor boundary [2], [3], [4]. There are also several models, which do not use constant tumor boundary concentrations [5], [6], [7]. However, the

consequences of this choice have not been discussed.

B. Lloyd is with the Computer Vision Laboratory, ETH Zürich, 8092

Zürich, Switzerland (corresponding author: +41 44 63 27690; fax: +41 44 63 21199; e-mail: [email protected]).

Prof. G. Székely is with the Computer Vision Laboratory, ETH Zürich, 8092 Zürich, Switzerland (e-mail: [email protected]).

This work has been performed within the frame of the Swiss National Center of Competence in Research on Computer Aided and Image Guided Medical Interventions (NCCR Co-Me) supported by the Swiss National Science Foundation.

In this short paper we analyze the significance of the mentioned boundary condition for qualitative and quantitative studies alike. Based on a simple model, which does not treat the concentration at the boundary as constant, we identified an alternative reason for growth saturation.

II. METHODS For our analysis we use a simple model of tumor growth,

which includes oxygen transport and production/ consumption as well as proliferation-based growth and elastic tissue response. The oxygen concentration in the tissue obeys a reaction-diffusion equation of the form

2 ( ) ( ),idc cD R c R cdt

+ −= ∇ + − (1)

where c is the concentration, Di is the diffusion coefficient, R+(c) is the source term, which depends on the vasculature and blood flow, and R-(c) is the consumption of oxygen by the cells. For the consumption we take a logistic

term -sat

1/ 2

( )R (c)=R( ) ( )

p

p

cc c+

I

p

0

, where c1/2 is the

concentration at which the reaction term reaches half maximum and p controls the shape of R-(c). This reflects that the consumption is bounded by the amount of oxygen available but also reaches a maximum if oxygen is unlimited. The delivery of oxygen depends on the partial pressure difference between the blood and the tissue. It increases in hypoxic regions, while in regions with high concentrations, only little oxygen is delivered. Therefore, outside the tumor we assume a constant source with a linear term, which penalizes deviations from the optimal concentration (i.e.

). Since we focus on avascular tumors, the source term inside the tumor is zero. More advanced models are discussed elsewhere [4].

+0R (c)=R - R_p (c - c )

As in our previous paper [4], we treat cell proliferation macroscopically using an elastic growth model. We assume the growth rate to be zero for c<td and increases linearly to a maximum at c=cBC. For simplicity we set the displacement boundary condition to zero.

The parameters have been taken from the literature where


18

available. The diffusion constant in tissue is approximately Di=2.4 × 10-5 cm2s-1 [8]. Based on estimates of oxygen partial pressure in the human breast, we set the oxygen to cBC=2.5 × 10-3 cm3 O2 / cm3 [9] at the far boundary. Since this is the preferred concentration, we set c0=cBC. The saturation consumption must be Rsat=R0 (cBC

p + c1/2p )/ cBC

p in order to balance the source when there is no tumor. The reaction term depends largely on the tissue and varying conditions such as altitude, physical activity etc. Under resting conditions the zero order reaction term has been estimated to be R0=1.7 × 10-4 cm3 O2 / (cm3 s) [8]. Other parameters, such as c1/2 and Rp can only be guessed. For lack of better information we set p=1, c1/2=td=0.6cBC, Rp =0.02s-1.

We solve the model equations in polar coordinates for the spherically symmetric case using Galerkin finite element methods. The domain is described by the tumor interface position r=rI and the far boundary position r=rMAX.

We assume the growth time scale to be much larger than oxygen transport and elastic response time scales. Therefore, these processes can be treated in a quasi-static manner, i.e. in order to compute the current growth rate we use the static solution from Eq.1 for the current domain description.

III. RESULTS The Dirichlet boundary condition at the tumor surface can

be enforced in our framework by an infinite diffusion constant Di outside the tumor, or an infinite linear source coefficient Rp. While such settings are of course hardly realistic, they allow to study the effects of applying Dirichlet boundary conditions to the tumor growth problem. The results in this section are shown for the default parameters mentioned above - and for the case when either Di or Rp is increased by a factor of 100.

In Fig. 1 the evolution of the tumor volume is plotted against time. The shape of the curves (first exponential, then linear and finally saturation) is reminiscent of the result in Ward and King [1]. Note that although we have not assumed any volume loss due to necrosis, the growth curve does reach a saturation value.

In Fig. 2 the oxygen concentration profile of a tumor with a diameter of 1mm is shown for different parameter settings. The figure illustrates the effect of increasing Di or Rp in the healthy tissue and demonstrates that the actual oxygen tension can be significantly reduced at the tumor boundary. In Fig. 3 the reaction terms are depicted. The effect of increasing Rp is to raise the source by Rp (c0-c), which in the case where Rp is a factor 100 higher leads to a significantly larger source close to the tumor.

0 500 10000

0.05

0.1

0.15

0.2

0.25

Rad

ius

[cm

]

Time [days]0 100 200

0

0.01

0.02

0.03

0.04

0.05

Time [days]

DefaultHigh D

i

High Rp

Figure 1 The development of a spherically symmetric avascular tumor. The image on the right is a close-up.

0 0.05 0.1 0.15 0.20

0.5

1

1.5

2

2.5

x 10−3 Concentration profile

Radius [cm]

Con

cent

ratio

n

DefaultHigh D

i

High Rp

Figure 2 The O2 concentration profile is shown for different cases. The circles are plotted at the tumor boundary.

0

1

2

3

4

x 10−4 Source

a)

0 0.05 0.1 0.15 0.20

1

2

3

4

x 10−4

Radius [cm]

Consumption

b)DefaultHigh D

i

High Rp

Figure 3 The source and sink terms are shown for different parameter settings. The source term can reach a peak of R0 + c0Rp.

IV. DISCUSSION Using a simple model we have shown that volume loss and

transport of cell debris across the tumor boundary is not necessary to explain growth saturation. While we do not want to suggest that there is no volume loss due to apoptosis or necrosis, these results might suggest that its influence has


19

been eventually overestimated in the past. Especially removal of necrotic (vs. apoptotic) cells might be slower than implied, since the resulting cell debris must be ingested by neutrophils and macrophages.

REFERENCES [1] J. P. Ward and J. R. King, “Mathematical modelling of avascular

tumour growth. II: Modelling growth saturation.” IMA J Math Appl Med Biol, vol. 16, no. 2, pp. 171–211, Jun 1999.

[2] Y. Jiang, J. Pjesivac-Grbovic, C. Cantrell, and J. P. Freyer, “A multiscale model for avascular tumor growth.” Biophys J, vol. 89, no. 6, pp. 3884–3894, Dec 2005.

[3] G. Schaller and M. Meyer-Hermann, “Continuum versus discrete model: a comparison for multicellular tumour spheroids.” Philos Transact A Math Phys Eng Sci, vol. 364, no. 1843, pp. 1443–1464, Jun 2006.

[4] B. A. Lloyd, D. Szczerba, M. Rudin, and G. Sz´ekely, “A computational framework for modelling solid tumour growth.” Philos Transact A Math Phys Eng Sci, Jul 2008.

[5] A. R. A. Anderson, “A hybrid mathematical model of solid tumour invasion: the importance of cell adhesion.” Math Med Biol, vol. 22, no. 2, pp. 163–186, Jun 2005.

[6] X. Zheng, S. M. Wise, and V. Cristini, “Nonlinear simulation of tumor necrosis, neo-vascularization and tissue invasion via an adaptive finite-element/level-set method.” Bulletin of Mathematical Biology, vol. 67, no. 2, pp. 211–259, Mar 2005.

[7] R. Betteridge, M. R. Owen, H. M. Byrne, T. Alarc´on, and P. K. Maini, “The impact of cell crowding and active cell movement on vascular tumour growth,” Networks and Heterogeneous Media, vol. 1, no. 4, 2006.

[8] J. W. Ji, N. M. Tsoukias, D. Goldman, and A. S. Popel, “A computational model of oxygen transport in skeletal muscle for sprouting and splitting modes of angiogenesis.” J Theor Biol, vol. 241, no. 1, pp. 94–108, Jul 2006.

[9] P. Vaupel, K. Schlenger, C. Knoop, and M. Hckel, “Oxygenation of human tumors: evaluation of tissue oxygen distribution in breast cancers by computerized O2 tension measurements.” Cancer Res, vol. 51, no. 12, pp. 3316–3322, Jun 1991.


20

A Proposed Platform for Intelligent Identification of Organs from Medical Images

Emmanouil Skounakis, Konstantinos Banitsas, and Konstantinos Marias, Member, IEEE

Abstract— The Virtual Physiological Human (VPH) is initiated by the European Commission as a major scientific multi-disciplinary challenge, aiming to develop robust physiology or pathology models, in silico. However, to develop such patient-specific models it is crucial to implement multi-layered models describing different properties (e.g. electrical, mechanical and biochemical), and appropriate image analysis and data assimilation tools to identify their specific parameters from patient images. This paper describes an intelligent system that can identify organs from medical images by applying region-oriented segmentation technique for automatic recognition of objects. The platform has been tested with several CT-Scan images and it is designed with the aim of becoming a platform for applying models of pathophysiology.

I. INTRODUCTION OWDAYS, hospitals are moving toward digitization of the ever increasing number of medical images. It is a

fast growing area with a large volume of information and a variety of modalities such as the X-ray transmission imaging, magnetic resonance imaging (MRI), computerized tomography (CT), ultrasound images, positron emission tomography (PET), and the X-ray mammography (MG). The need to represent such large medical images with the smallest possible number of bits demonstrates the need for data compression which in its turn plays a very important role on minimizing storage requirements and reducing transmission times across low bandwidth communication channels.

In order to develop an intelligent system for organ recognition from medical images, the methods used have to

utilize all-important information and simultaneously minimize any information not necessary for diagnosis. The proposed platform is designated to primarily identify organs but at the same time provide an infrastructure for telemedical purposes. This is implemented in the presented platform by first segmenting the organs present in a given image. Then, the useful information (organs) is compressed by a lossless method while the residual image by a lossy method. Efficiently compressing data is the key to making telemedicine feasible, since the limited bandwidth provided by computer media limits the amount of medical data that can be transmitted. The technique used in this paper, known as Region of Interest (ROI) coding, increases compression in images with large regions of clinically insignificant information [1].

Manuscript received September 1, 2008. E. Skounakis is a PhD researcher at the Electronic and Computer

Engineering Department, Brunel University, West London, England. He is also a researcher at the Institute of Computer Science at FORTH, Vassilika Vouton, GR-70013 Heraklion, Crete, Greece ([email protected])

Dr. K. Banitsas is a researcher at the Electronic and Computer Engineering Department, Brunel University, West London, England ([email protected])

Dr. K. Marias is a researcher at the Institute of Computer Science at FORTH, Vassilika Vouton, GR-70013 Heraklion, Crete, Greece ([email protected]).

K. Marias and E. Skounakis acknowledge support from EC Contra Cancrum Project ICT-2007-223979.

The large proportion of background in CT scans, yield significant compression gains since it can be compressed with a lossy method, while the clinically important features of the image (organs) can be losslessly compressed to preserve the important image information. This concept is exploited in our work in order to develop a platform that isolates different organs from tomographic data as is described in the next section.

II. METHODOLOGY

Fig. 1. The proposed Platform

Fig. 1 is a schematic representation of the proposed platform. The sequence of the process starts with the image

N


21

being loaded (DICOM file) and then the header information is removed in order to process the raw image data, while a bmp version of the image is displayed to the user. The image data is then analyzed. Initially, by applying the region-oriented segmentation method we get some basic features of the regions, including:

• The area of the region which defines how many pixels

are included in the specific region. • The perimeter of the region which defines the perimeter

of the defined object. • The center of gravity which defines the center of the

defined region. • The compactness of the region. The first two properties are of less importance as they

depend on the object’s position, rotation and scaling. The center of gravity and the compactness are features invariant to the position, rotation and scaling of the object, so they are very useful for identifying organs from tomographic datasets.

The Region Oriented Segmentation method is only a part of the whole ROI Compression system. A front-end application has also been developed in order to manage the implemented processes such as lossless compression, lossy compression, segmentation, and also show the results on the screen. The application contains 7 image frames for the presentation of the images created at every step of the method to visualize the effect on the source image. Fig 2 illustrates the functionality of the platform in identifying the liver and aorta in a CT image.

III. INITIAL RESULTS The proposed ROI Compression platform has shown remarkable robustness in accurate identification and segmentation of four organs (Stomach, Liver, Aorta, and Spinal Cord) in a small dataset of 10 abdominal images. The proposed platform is developed to assess the effectiveness of the Region-Oriented segmentation method [2], [3] when used in recognizing organs within medical images. The most important features were found to be the center of gravity (position of the object) and the compactness, since they are independent of the organ’s size. The results presented represent our initial validation work; however it is essential to validate it in a large number of data and at the same time define measures of agreement with a clinical expert (golden truth).

IV. CONCLUSION We presented a platform for identifying different organs

from CT data. The method has been applied for the automatic recognition of four organs (Stomach, Liver, Aorta, and Spinal

Cord) and the initial results are promising. It is our goal to extend the validation of the system by also using heterogeneous sets of images (from different scanners), taking into consideration possible alterations in image contrast. Also, the suggested platform could serve as the basis for analyzing tomographic data and applying models in selected organs and pathologies (e.g. tumors). This could facilitate the simulation process by providing the segmented 3D data of a tumor to a modeling algorithm along with an integrated environment for simulating e.g. the therapy response of a tumor and visualizing the results.

(a)

(b)

Fig. 2. Automatic identification of the liver (a), and of the aorta (b)

Fig.2 illustrates the function of the proposed platform. The top left window displays the source image while the four windows below, present the intermediate results of the “Region Oriented Segmentation Method”. The mark image (bottom right) labels the identified organs/regions in the


22

image. The top middle image displays the contour of the selected area (liver in (a), aorta in (b)), and the top right image the final ROI Compressed Image (the selected area is losslessly compressed, and the rest of the image is heavily compressed).

REFERENCES [1] A. Bruckmann, and A. Uhl. “Selective Medical Image Compression

Techniques for Telemedical and Archiving applications.” Comp. Biol. Med., vol. 30: pp. 153-169, 2000.

[2] S. Bokturk, C.. Tomasi, B. Girod, C. Beaulieu. “Medical Image Compression based on region of interest with application to colon CT Images.” Proc. 23rd Annual International Conference of the IEEE on medical and Biomedical Engineering, Istanbul, Turkey, 2001.

[3] Ad Oculos –Digital Image Processing– Student Version 2.0. Henning Bassmann, Philipp Besslich I.S.B.N 1-850-32132-9

[4] M.Unser and A. Aldroubi. “A review of wavelets in biomedical applications.” Proceedings of the IEEE, vol. 84(4), pp. 626-638, 1996.


23

POLR2F, ATP6V0A1, PRNP and their Prognostic Value in Dukes' B Colon Cancer

Lambros Skarlas, Anna Antonacopoulou and Haralambos Kalofonos

Abstract— Expression microarray data was analyzed using MAPS to reveal elevated expression of POLR2F in relapsed colorectal carcinoma patients. mRNA levels of POLR2F were also evaluated with quantitative RT-PCR in 70 colorectal carcinomas and 17 normal tissue specimens and were correlated with clinicopathologic parameters. POLR2F was upregulated in colorectal carcinomas.

The predictive value of POLR2F expression regarding disease recurrence was confirmed only by in silico analysis. Additionally, POLR2F and PRNP exhibited elevated mRNA levels in carcinomas compared to normal tissue both in silico and in vitro. This finding may suggest a possible role for these molecules in colorectal tumorigenesis.

I. INTRODUCTION1

OLORECTAL cancer is the second leading cause of cancer death in the western world. Colorectal

carcinogenesis is a multistep process during which alterations in the expression of particular genes constitute important steps. The control of gene expression can occur in various steps, but the majority of regulatory events occurs at the transcriptional level.

The computational analysis of available microarray data for Dukes’ stage B colorectal carcinomas suggested that among other genes POLR2F overexpression was strongly associated with disease relapse. Therefore, the purpose of the current study was to investigate the role of specific genes in colorectal cancer, and the concordance of these findings both biologically and computationally. Particular combinations of expression of these three genes may constitute prognostic biomarkers in colorectal cancer. An enzyme with an important role in transcription is DNA dependent RNA polymerase II (POLRII). POLRII is responsible for synthesizing mRNA from protein-encoding genes and it is composed of 10-14 subunits. The sixth largest of these, is the subunit F (POLR2F, RpB6). POLR2F is shared by the other two DNA-directed RNA polymerases (I and III). Its basic function is to catalyze the transcription of DNA into RNA using the four

Manuscript received September 1, 2008. L.Skarlas is with the Computer Engineering and Information Department, University of Patras and with the Clinical Oncology Laboratory, University Hospital of Patras. A. Antonacopoulou is with the Clinical Oncology Laboratory, University Hospital of Patras. H. Kalofonos is with the Clinical Oncology Laboratory, University Hospital of Patras.

ribonucleoside triphosphates as substrates [3]. POLRII facilitates efficient transcription initiation and

elongation of genes encoding antiapoptotic proteins. It has been reported that the inhibition of these cyclins reduces the accumulation of transcripts with short half lives including those encoding anti-apoptotic proteins and cell cycle regulators [7]. Indeed, diminution of POLRII levels leads to decreased expression of the anti-apoptotic proteins MCl-1 and X-linked inhibitor of apoptosis (XIAP) in non-small cell lung cancer and osteosarcoma cells[1]. Therefore, the upregulation of POLRII expression and/or phosphorylation may result in an increase of anti-apoptotic factors.

II. MATERIALS AND METHODS

Gene expression profiles of 16 patients with Dukes B colorectal cancer with or without recurrence within a 5 year period follow up were evaluated using Human 19K Oligo Array slides (Center for Applied Genomics, University of Medicine of New Jersey). The dataset is publicly available from the NCBI GEO database(AccessionNumber:GDS1263, http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE2630).

The intensity value associated to each probe is the result of subtracting a Gaussian function of the noise from the foreground values. After this background subtraction, base 2 logarithms of all data were calculated and genes with more than two missing values were excluded from the analysis. The remaining missing values were replaced by using the KNN imputation method. In order to analyze the microarray data for the 16 samples, MAPS tool [19] was used. More specifically, in order to identify gene markers that can best discriminate between relapsed and non-relapsed patients, a supervised class prediction method was followed. From the 19,200 total genes we divided all 16 patients into a training set of 10 samples and a test set of 6 samples (Table I). The training set was used to select gene markers and to build a prognostic signature. The test set contained the remaining samples and was used for independent validation.

MAPS contains several procedures for finding significant differences in gene expression between the relapsed and non-relapsed patients. Two methods were used to find the differentially expressed genes. The first test is the simple permutation t-test for comparison of two

C


24

means and the second is a novel method that is implemented in MAPS computation tool [9].

This method is entitled MVT (Μean Value Trust) and the computational formula is the following:

if (meansΑln(i)-meansΒln(i)) > ln((threshold)) AND ((min(genesAln(i:)) > max(genesBln(i:)))) then gene(i) is overexpressed in sample A, otherwise in B where A is the group of the non-relapsed patients, and B is the group of the relapsed patients, meansAln(i) is the mean value of the natural logarithm of gene(i) expression for samples belong to A group, min(genesAln(i:)) in the minimum value of gene(i) in A group samples. The ln((threshold)) is a constant value that the user can define according the desired accuracy. For example if we want to have high accuracy of the differentially expressed genes we choose a low value, typically smaller than 2.

TABLE I

THE TRAINING SET CONTAINS 6 SAMPLES OF NON-RELAPSED PATIENTS AND 4 RELAPSED PATIENTS (HIGHLIGHTED). THE SAMPLES WITH THE PREFIX 1 DENOTE THE NON-RELAPSED PATIENTS WHILE THE PREFIX 2 DENOTES THE RELAPSED PATIENTS. THE REMAINING 4 DISEASE FREE SAMPLES AND 2 RELAPSED PATIENTS’ SAMPLES FORM THE TESTING SET.

Samples SexAge at diagnosis Phase Size Location Recurrence

1GSM50474 M 66 T3N0M0 B5cm Colon sigma NO1GSM50496 F 45 T3N0M0 B5cm Left colon NO1GSM50504 M 78 T3N0M0 B5cm Right colon NO1GSM50505 F 69 T3N0M0 B5cm Right colon NO1GSM50506 M 67 T2N0M0 B5cm Colon sigma NO1GSM50507 M 67 T2N0M0 B5cm Colon sigma NO1GSM50508 F 66 T3N0M0 B5cm Colon sigma NO1GSM50509 M 81 T3N0M0 B5cm Colon sigma NO1GSM50511 M 67 T2N0M0 B5cm Right colon NO1GSM50512 F 48 T2N0M0 B5cm Colon sigma NO2GSM50473 M 42 T3N0M0 B5cm Colon sigma YES2GSM50475 M 72 T3N0M0 B5cm Colon sigma YES2GSM50510 M 72 T3N0M0 B5cm Colon sigma YES2GSM50513 M 72 T2N0M0 B5cm Colon sigma YES2GSM50514 M 60 T2N0M0 B5cm Right colon YES2GSM50515 F 58 T3N0M0 B5cm Left colon YES

In order to predict the unknown class of the test set we choose a simple method that is straightforward for the investigators to assess and to visualize.

Using only the differentially expressed genes we plot the unknown sample together with every other known sample (Fig. 1). It is then easy to see the similarity of the “behaviour” of the unknown sample with a labelled sample. For example in Fig. 1, it is easy to see that sample 8 and sample 11 have a high degree of similarity.

More formal though, using the mean squared error (MSE) we can compute the difference of the unknown sample with every other sample and categorize it according the smallest MSE.

III. RESULTS

Expression microarray data was analyzed using two methods to reveal differentially expressed genes in the

training set (tables 2 and 3). The union of the genes identified by both methods yielded a group of 139 differentially expressed genes between the relapsed and non-relapsed groups. This 139 genes signature was then evaluated in the 6 independent patients (test set) to test its predictive performance in Dukes’ B colon cancer patients. Two of 2 relapse patients and 6 out of 6 disease-free patients were predicted correctly.

Fig. 1. MAPS instance of the class prediction. For the “unknown” sample 8, with the use of the Mean of Squared Errors algorithm we calculate the minimum Euclidean distance between this sample and every other sample. The smallest distance of all other samples and the unknown sample classifies the unknown sample to that sample. The unknown sample 11 (RED color) mostly similar with sample 8 (BLUE color) depicted in 2nd column 3rd row.

TABLE II

THREE OF TOTALLY 133 UPREGULATED GENES IN NON-RELAPSED PATIENTS. THE RESULTS OF THE MVT METHOD AND THE T-TEST METHOD ARE ALMOST IDENTICAL EXCEPT FOR THE LAMA2, THE CDKN2C, THE ALG1, THE ZNF148 AND THE SIRT7 GENE, WHICH ARE NOT OVEREXPRESSED ACCORDING TO THE T-TEST. COLUMN CERTAINTY DENOTES THE RESULT OF THE NULL HYPOTHESIS AND IT IS 1, WHEN THERE IS A VIOLATION IN THE NULL HYPOTHESIS AND 0 OTHERWISE.

Probe Gene MVT t-test

s10566 CDKN2C 0.75 0.04 ……… ……… ……… ………

……… ……… ……… ………

s95 TMEM16H 0.66 0.03

s2062 PELO 0.55 0.01

In order to find the most accurate classifier (predictor) we make with the use of the smallest possible subset of the 139 genes. By using the same training and test set but now only one gene the POLR2F, which is found overexpressed


25

in the recurrent, we were able to accurately label every unknown sample.

TABLE III SIX OVEREXPRESSED GENES IN THE RECURRENCE SAMPLES. THE RESULTS OF THE MVT METHOD AND THE T-TEST METHOD ARE ALMOST IDENTICAL EXCEPT FOR THE C21ORF86 GENE, WHICH IS NOT OVEREXPRESSED ACCORDING TO THE T-TEST.

Probe Gene MVT t-test

s11555 NULL 0.97 0.04

s13976 POLR2F 0.86 0.01

s11571 ATP6V0A1 0.83 0.02

s11530 RORA 0.82 0.01 s13073 C21orf86 0.8 0.06

s11507 PRNP 0.76 0.04

Figure 2. Hierarchical clustering for the 139 genes list. All the relapse free samples are in the same clusters. The same is valid for the non-relapsed samples.

In Fig. 3, the hierarchical clustering based on the POLR2F gene is shown for the 16 samples. It can be observed that the “unknown” samples: 1GSM50508, 1GSM50509, 1GSM50511, 1GSM50512, 2GSM50473, 2GSM50475, 2GSM50510, 2GSM50513, 2GSM50514, and 2GSM50515 belong to the same clusters with their real (known) label.

Three of the 139 genes, namely POLR2F, PRNP and ATP6V0A1, that were found to be overexpressed by both methods of microarray data analysis in carcinomas from patients who relapsed, were chosen for further analysis to assess any possible concordance between the computationally obtained results and the biological data.

Seventy patients with stage B and C colorectal carcinomas with or without disease relapse were chosen and the expression levels of POLR2F, PRNP and ATP6V0A1 were evaluated in the 70 colorectal carcinomas and in 17 normal tissue specimens using quantitative RT-PCR. The 3 genes were expressed in all normal tissues with

median values of 0.0545 (0.0138-0.3867) for POLR2F, 0.1266 (0.0781-0.3025) for PRNP and 0.031753 (0.0799-0.2244) for ATP6V0A1. Expression of POLR2F was detected in 68 (97.14%), ATP6V0A1 in 67 (95.71%) and PRNP in all of the carcinomas. The corresponding median levels of POLR2F, PRNP and ATP6V0A1 expression in the carcinomas were 0.1015 (0.0184-9.6910), 0.1874 (0.1061-0.7996) and 0.0443 (0.0092-0.3591), respectively.

Fig. 3. The Hierarchical clustering for the POLR2F gene (probe s13976).

The expression levels of POLR2F and PRNP were statistically significantly higher in carcinomas than in normal tissue (p=0.034 and p<0.001, respectively). In carcinomas, POLR2F levels strongly correlated with PRNP and ATP6V0A1 levels (r= 0.625, p<0.001; r=0.835, p<0.001, respectively). Moreover, PRNP levels strongly correlated with ATP6V0A1 levels (r=0.689, p<0.001).

Additionally, using the median value of expression as a cutoff we divided the carcinomas in two categories, over-expressing (above median) and under-expressing (below median), and examined different combinations of over/underexpression of the three genes. POLR2F, PRNP and ATP6V0A1 levels were simultaneously overexpressed in 27 out of 67 (40.29%) specimens and underexpressed in 21 out of 67 specimens (31.34%) expressing all three genes. Moreover, elevated levels of PRNP with concomitant low levels of POLR2F and ATP6V0A1 were noted in 5 out of 67 cases (7.46%). Kaplan-Meier analysis using the 3-year survival rate as an endpoint, revealed that high expression levels of POLR2F or ATP6V0A1 conferred a survival benefit compared to lower levels (p=0.002, p=0.002, respectively) (Fig. 4).

IV. CONCLUSIONS

We analyzed publicly available expression microarray data in order to pinpoint genes with major roles in tumor development and progression, as suggested by their


26

differential expression between recurrent and non-recurrent colorectal tumors. POLR2F, PRNP and ATP6V0A1 were found to differ significantly. Their expression levels were thus evaluated in primary colorectal carcinomas as well as in adjacent normal mucosa.

POLR2F and PRNP exhibited elevated levels in carcinomas compared to normal tissue samples. Increased expression of POLR2F may reflect a higher transcriptional activity in tumor cells. Although data regarding POLR2F expression in cancer is scarce, POLR2F has been reported to be a metastasis-specific gene in colon cancer cells [5]. With regard to PRNP, elevated levels in carcinomas agree with the recently reported upregulated PRNP levels in gastric carcinomas and gliomas [2] as well as with the antiapoptotic function of this molecule. Increased levels of ATP6V0A1 were expected in carcinomas as they would reflect the effort the cancer cell applies to achieve intracellular pH homeostasis in the hypoxic conditions inside the tumor. This is in agreement with the findings of Sennoune et al, [6] who reported that V-ATPases not only participate in the pH homeostasis but their increased levels are also correlated with higher metastatic potential of breast cancer cells.

Expression levels of POLR2F, PRNP and ATP6V0A1 did not correlate with age, gender, grade or stage of the disease. Expression levels of the three genes were not associated with any other clinicopathologic parameters with the exception of the primary tumor site. In particular, expression levels differed significantly between right colon and rectum. This finding is in agreement with the different epidemiology and distinct gene expression profiles displayed by tumors located in these two anatomical sites [4].

None of the molecules correlated with disease recurrence or disease free survival. This contradicts the microarray data used for the selection of the studied genes. This may be attributed to the different tissue specimens used for the microarrays and the RT-qPCR validation.

However, the prognostic significance of the expression levels of each gene individually or in combination with each other as also the clinicopathologic parameters was assessed for the 3-year survival. Patient’s age and primary tumor site possessed prognostic value. Between the three molecules studied, only PRNP expression appeared to be an independent prognostic factor, when POLR2F and ATP6V0A1 expression levels were also taken into account. In particular, high PRNP levels with concomitant low POLR2F and ATP6V0A1 levels offered an advantage in survival. This may suggest that elevated PRNP levels alone are not sufficient to inhibit apoptosis and invasion. To conclude, PRNP, and POLR2F appear to play a role in colorectal carcinomas although further studies are necessary to confirm our results and elucidate the role of these molecules in colorectal cancer.

REFERENCES [1] D. Cai, V.M.Jr. Latham, V.M., X. Zhang, G.I. Shapiro. “Combined

depletion of cell cycle and transcriptional cyclin-dependent kinase activities induces apoptosis in cancer cells.” Cancer Res., vol. 66, pp. 9270-9280, 2006.

[2] S. Comincini, A. Facoetti, I. Del Vecchio, K. Peoc'h, J.L. Laplanche, L. Magrassi, M. Ceroni, L. Ferretti, R. Nano. “Differential expression of the prion-like protein doppel gene (PRND) in astrocytomas: a new molecular marker potentially involved in tumor progression.” Anticancer Res, vol. 24, pp. 1507-1517, 2004.

[3] P. Cramer, D.A. Bushnell, J. Fu, A.L. Gnatt,, B. Maier-Davis, N.E. Thompson, R.R. Burgess, A.M. Edwards, P.R.David, R.D. Kornberg. “Architecture of RNA polymerase II and implications for the transcription mechanism.” Science, vol. 288, pp. 640-649, 2000.

[4] K. Komuro, M. Tada, E. Tamoto, A. Kawakami, A. Matsunaga, K. Teramoto, G. Shindoh, M. Takada, K. Murakawa, M. Kanai, N. Kobayashi, Y. Fujiwara, N. Nishimura, J. Hamada, A. Ishizu, H. Ikeda, S. Kondo, H. Katoh, T. Moriuchi, T. Yoshiki . “Right- and left-sided colorectal cancers display distinct expression profiles and the anatomical stratification allows a high accuracy prediction of lymph node metastasis.” J Surg Res, vol. 124, 216-24, 2005.

[5] V. Orian-Rousseau, S. Mink, J. Mengwasser, H. HogenEsch, F. Guo, W.G. Thies, M. Hofmann, P. Herrlich, H. Ponta. “Genes upregulated in a metastasizing human colon carcinoma cell line.” Int J Cancer, vol. 113, pp. 699-705, 2005.

[6] S.R. Sennoune, K. Bakunts, G.M. Martinez, J.L. Chua-Tuan, Y. Kebir, M.N. Attaya, R. Martinez-Zaguilan. “Vacuolar H+-ATPase in human breast cancer cells with distinct metastatic potential: distribution and functional activity.” Am J Physiol Cell Physiol, vol. 286, pp. C1443-52, 2004.

[7] G.I. Shapiro. “Cyclin-dependent kinase pathways as targets for cancer treatment.” J Clin Oncol, vol. 24, pp. 1770-83, 2006

[8] M. Taron, R. Rosell, E. Felip, P. Mendez, J. Souglakos, M.S. Ronco, C. Queralt, J. Majo, J.M. Sanchez, J.J. Sanchez, J. Maestre. “BRCA1 mRNA expression levels as an indicator of chemoresistance in lung cancer.” Hum Mol Genet, vol. 13, pp. 2443-9, 2004.

[9] L. Skarlas, D. Tsavachidou, D. Sanoudou, L.Spiridon, H.Kalofonos, B.L. Weber, “MAPS: MicroArray Processing Software for management, data mining and visualization of gene expression data”, Computational Methods in Molecular Biology: Theory and Applications, International Conference of Computational Methods in Sciences and Engineering, 27 October - 1 November 2006, Chania, Crete, Greece.)

[10] A.G. Antonacopoulou, P.D. Grivas, L. Skarlas, M. Kalofonos, C.D. Scopa, H.P. Kalofonos. “POLR2F, ATP6V0A1 and PRNP expression in colorectal cancer: new molecules with prognostic significance?” Anticancer Res. , vol. 28(2B), pp. 1221-7, 2008.


27

Simulating the Response of Nephroblastoma Tumor to Chemotherapy in the Clinical Context

Eleni Ch. Georgiadi, Eleni A. Kolokotroni, Dimitra D. Dionysiou, Norbert M. Graf, A. Hoppe, N.

K. Uzunoglu and Georgios S. Stamatakos.

Abstract—Many efforts have been made during the last

decades to investigate the complex biological mechanisms of cancer with the vision of treatment optimization. Various models have been proposed but their clinical adaptation requires explicit clinical validation. One major component of the EC funded project “Advancing Clinico-Genomic Trials on Cancer” (ACGT) (FP6-2005-IST-026996) is the development of a patient specific four dimensional multilevel cancer model simulating the nephroblastoma tumor response to the administration of chemotherapeutic agents (vincristine, dactinomycin) according to the SIOP 2001/GPOH clinical trial. Individualized histopathological, molecular and clinical data of the SIOP 2001/GPOH trial are appropriately collected and utilized. This paper briefly describes the basic principles of the model developed by the In Silico Oncology Group and focuses on the

Manuscript received July 4, 2008. E. Ch. Georgiadi is with the National Technical University of Athens,

Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected] )and the Foundation for Research and Technology – Hellas, Institute of Computer Science, N. Plastira 100, Vassilika Vouton, GR-700 13 Heraklion, Crete, Greece

E. A. Kolokotroni is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected]).

D. D. Dionysiou is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected]).

N.Graf is with the University Hospital of the Saarland, Paediatric Haematology and Oncology, D-66421 Homburg,, Germany (e-mail: [email protected] ).

A.W. Hoppe is with University Hospital of Saarland, Paediatric Haematology and Oncology, D – 66421 Homburg Germany (phone: 004968411628045, email:[email protected])

N.K.Uzunoglu is with the National Technical University of Athens, School of Electrical and Computer Engineering,, 9 Iroon Polytechniou, GR 157 80, Greece (phone: +30 210 772 3556; fax: +30 210 772 3557; e-mail: nnap@)otenet.gr )

G. S. Stamatakos (corresponding author) is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (phone: + 30 210 772 2288, fax: + 30 210 772 3557, e-mail: [email protected] ).

This work was supported by the European Commission under the project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996).

new categorization of tumor cells to stem/clonogenic, progenitor and differentiated cells. The need for specialized tumor initialization methods is pointed out and the use of a nomogram connecting cell category transition rates with cell category percentages is proposed. Indicative results are presented. The agreement of these results with clinical reality so far is promising, while further clinical validation is in progress.

I. INTRODUCTION

THE simulation model presented in this paper is based

on the top-down modelling approach and has been developed by the In Silico Oncology Group [3-6] (www.in-silico-oncology.iccs.ntua.gr) within the framework of the EC funded project “Advancing Clinico-Genomic Trials on Cancer” (ACGT) (FP6-2005-IST-026996) [1-2]. A case of neoadjuvant nephroblastoma tumor chemotherapeutically treated according to the SIOP 2001/GPOH clinical trial is presented.

II. OUTLINE OF THE SIMULATION MODEL A. Tumor initialization and growth

The anatomical region of interest is quantitized in a mesh with the spatial unit of a geometrical cell (GC) [3-6]. In this paper we consider a homogeneous ellipsoidal tumor as an approximation to a real clinical nephroblastoma tumor. Each geometrical cell can normally accommodate a number of 109 cells/cm3 [7] biological cells (NBC). Biological cells are categorized according to mitotic potential and follow the cytokinetic diagram shown in Fig.1 for the case of free tumor growth. All biological cells within a geometrical cell are considered synchronized.

The following biological phenomena are simulated (Fig.1): • stem cells cycle through the cell cycle phases, may

undergo spontaneous apoptosis, perform symmetric or asymmetric divisions when they reach M phase, may pass to the dormant (G0) phase due to inadequate supply of oxygen and nutrients, re-enter the cell cycle in case of reestablishment of oxygen and nutrient supply or die through necrosis (due to prolonged oxygen and nutrients’ shortage).

http://www.in-silico-oncology.iccs.ntua.gr/

http://www.in-silico-oncology.iccs.ntua.gr/


28

• Limp cells cycle through the cell cycle phases, perform three mitotic divisions and then terminally differentiate. They may pass to the dormant (G0) phase due to inadequate supply of oxygen and nutrients, re-enter the cell cycle if there is local reestablishment of oxygen and nutrients’ supply or otherwise die through necrosis. They also undergo spontaneous apoptosis.

• Diff cells undergo spontaneous apoptosis and are led to necrosis with a specific rate per hour.

• Dead, necrotic and apoptotic cells are removed from the tumor volume with a constant rate, which is higher for apoptotic cells and lower for necrotic.

Celldisappearance

Apoptosis

Spontaneous apoptosis

Necrosis

Celldisappearance

G0G1 S G2 M

Local reoxygation

G0G0

M G2 S G1

AsymmetricDivision

DIFFSTEM LIMP

Symmetric division

Spontaneous apoptosis

Local reoxygation

After n mitosis

Fig. 1. Cytokinetic model for free tumor growth. STEM: stem cells. LIMP: Limited proliferative potential cells. DIFF: terminally differentiated cells. In the present version of the code n has been taken equal to 3. In future versions higher values of n will be considered.

B) Tumor Response to chemotherapy

After the administration of a chemotherapeutic agent, lethally hit cells enter a separate cell cycle before dying through apoptosis. The simulation algorithms developed so far refer to the cases of unilateral stage I-III nephroblastoma tumors treated with preoperative chemotherapy with combined administration of actinomycin-D and vincristine, according to the SIOP2001/GPOH clinical trial in the framework of the ACGT project (Fig. 2).

Fig. 2: SIOP2001/GPOH clinical protocol for Wilm’s tumours.

Vincristine and actinomycin-D are concurrently administered drugs. Hence, as a first approximation the corresponding cell kill fractions computed according to the pharmacodynamics of each drug are added in order to acquire

the total cell kill fraction (cell kill fraction = 1-cell survival fraction) [8], [9].

III. THE NOMOGRAM OF CELL STATE CATEGORY TRANSITION PROBABILITIES AND DISTRIBUTION OF CELL STATE

CATEGORIES IN A FREE GROWING TUMOUR The initial distribution of cells to sub-categories and the

specification of the transition probabilities have a great impact on the tumor development. If the initial cell category relative populations of the tumor are not in accordance with the transition rates, the tumor’s free growth time course curves present abnormalities in the beginning of the simulations as evident in Fig. 3.

0,00

50,00

100,00

150,00

200,00

250,00

300,00

0 100 200 300 400 500

time(h)po

pula

tion(

10^8

cel

ls)

totaldifferentiatedlimpstemdead

Fig. 3. Simulation results of free growth with inappropriately chosen initial percentages of the various populations in relation to given transition probabilities and cell cycle durations

In order to obtain a biologically valid initialization, a single geometrical cell is initialized with a small number of stem cells and the simulation proceeds until it reaches a population of 106 cells. Subsequently, the relative populations of the various cell categories are recalculated and used for the correct initialization of all geometrical cells of the tumor. Fig. 4 presents the simulation results of the tumor’s free growth with appropriately chosen initial percentages of the various populations in relation to given transition probabilities and cell cycle durations.

Various simulation executions have been performed with different initial populations of stem, limp, diff and dead cells and various combinations of the fraction of cells that enter G0 phase after mitosis (sleep_fraction) and the fraction of the stem cells that divide symmetrically (sym_fraction). The relative populations of the stem, limp, diff and dead cells after achievement of equilibrium are obtained and the results form a nomogram used as a baseline to initialize different tumors.

IV. THE NEPHROBLASTOMA CASE - THERAPY RESULTS A triaxial ellipsoidal nephroblastoma tumor with axes 10

mm, 20 cm and 30 cm is initialised. The response of this tumor to the chemotherapeutic scheme defined in Fig. 2 is simulated. During the first 4 days free growth is simulated.


29

Fig. 5, 6 and 7 present typical tumor volume and cell population curves as functions of time. An increase in all cell category populations is obvious during this interval. In Fig.7 repopulation of certain cell categories can be easily observed.

0,00

10,00

20,00

30,00

40,00

50,00

60,00

70,00

80,00

0 100 200 300 400 500

time(h)

popu

latio

n(10

^8 c

ells

)

totaldifferentiatedlimpstemdead

Fig. 4. Simulation results of free growth with appropriately chosen initial percentages of the various populations in relation to given transition probabilities and cell cycle durations

Tumour volume

0500

10001500200025003000350040004500

0 100 200 300 400 500 600 700 800

time(h)

volu

me(

mm

^3)

Fig. 5. The time course effect of the chemotherapeutic session (Fig.1) on the volume of a nephroblastoma tumor characterized by the parameter values shown in Table 1. The time point 0 corresponds to a triaxial ellipsoidal tumour with axes equal to 10 mm, 20mm and 30 mm.

Differentiated population

0,00

5,00

10,00

15,00

20,00

25,00

30,00

35,00

0 100 200 300 400 500 600 700 800

time(h)

popu

latio

n(10

^8 c

ells

)

Fig. 6. The time course effect of the chemotherapeutic session (Fig.1) on proliferating (stem and progenitor) and dead cells’ population of a nephroblastoma tumor characterized by the parameter values shown in Table 1. The time point 0 corresponds to a triaxial ellipsoidal tumour with axes equal to 10 mm, 20mm and 30 mm.

V. DISCUSSION A multiscale simulation model of clinical nephroblastoma

tumor growth and response to chemotherapy has been

outlined. The need for correlation of the initial cell category relative populations to the cell category transition rates has been pointed out and the concept of the corresponding nomogram has been introduced. A method to initialize the various cell category relative populations at the beginning of the simulation has been described. The simulation results for free tumor growth and response to chemotherapy are presented (Fig. 4-7) and are representative of clinical reality. Introduction of tumour layers [3-5] and definition of different cell category transition rates for different layers will produce the model’s potential to study also non homogeneous tumors. Further use of SIOP 2001 / GPOH clinical trial data is also expected to enhance the model’s optimisation.

0

1

2

3

4

5

6

7

0 100 200 300 400 500 600 700 800

time(h)po

pula

tion(

10^8

cel

ls)

limpstemdead

Fig. 7. The time course effect of the chemotherapeutic session (Fig.1) on differentiated cells’ population of a nephroblastoma tumor characterized by the parameter values shown in Table 1. The time point 0 corresponds to a triaxial ellipsoidal tumour with axes equal to 10 mm, 20mm and 30 mm.

REFERENCES [1] ACGT: Advancing Clinicogenomic Trials on Cancer (FP6-2005-IST-

026996), http://www.eu-acgt.org/ [2] N. Graf and A. Hoppe. “What are the expectations of a Clinician from

In Silico Oncology ?.” Proc. 2nd International Advanced Research Workshop on In Silico Oncology, Kolympari, Chania, Greece, 25-26 Sept. 2008, Ed. K. Marias and G. Stamatakos, pp. 36-38 (http://www.ics.forth.gr/bmi/2nd-iarwiso/).

[3] G.S. Stamatakos, D.D. Dionysiou, E.I. Zacharaki, N.A. Mouravliansky, K.S.Nikita, N.K. & Uzunoglu. “In silico radiation oncology: combining novel simulation algorithms with current visualization techniques, Proceedings of IEEE: Special Issue on Bioinformatics: Advances and Chalenges , vol . 90(11), pp. 1764-1777 , 2002.

[4] D. D. Dionysiou, G. S. Stamatakos, N.K. Uzunoglu, K. S. Nikita, A. Marioli. “A four-dimensional simulation model of tumour response to radiotherapy in vivo: parametric validation considering radiosensitivity, genetic profile and fractionation.” J. of Theor. Biol. vol. 230, pp. 1–20, 2004.

[5] G. S. Stamatakos, V. P. Antipas and N. K. Uzunoglu. “A spatiotemporal, patient individualized simulation model of solid tumor response to chemotherapy in vivo: the paradigm of glioblastoma multiforme treated by temozolomide.” IEEE Transactions on Biomedical Engineering, vol. 53, pp. 1467- 1477, 2006.

[6] G.S.Stamatakos, D.D.Dionysiou, N.M.Graf, N.A.Sofra, C.Desmedt, A.Hoppe, N.Uzunoglu, M.Tsiknakis. “The Oncosimulator: a multilevel, clinically oriented simulation system of tumor growth and organism response to therapeutic schemes. Towards the clinical evaluation of in silico oncology.” Proceedings of the 29th Annual International


30

Conference of the IEEE EMBS, August 23-26, SuB07.1: 6628-6631 , Lyon, France , 2007.

[7] G. Steel, Ed., Basic Clinical Radiobiology, 3rd ed. London, U.K.: Arnold, 2002, pp. 165–168, pp. 13, 52–63.

[8] K. Sawada, K. Noda, H. Nakajima, N. Shimbara, Y.Furuichi, M. Sugimoto. “Differential cytotoxicity of anticancer agents in pre- and postimmortal lymphoblastoid cell lines.” Biol Pharm Bull vol. 28, pp. 1202- 1207, 2005.

[9] G.J. Veal, M. Cole, J .Errington, A. Parry, J. Hale, A.D.J. Pearson, K. Howe, J.C. Chiholm, C. Beane, B. Brennan, F. Waters, A. Glaser, S. Hemsworth, H. McDowell, Y. Wright, K.ritchard- Jones, R. Pinkerton, G. Jenner, J. Nikolson, A.M. Elsworth, A.V. Boddy, and UKCCSG Pharmacology Working Groups. “Pharmacokinetics of Dactinomycin in a pediatric patient population: a United Kingdom Children’s Cancer Study group study,” Clin Cancer Res vol. 11(16), pp. 5893-5899, 2005.


31

Abstract— In this paper a four dimensional, top-down, Monte-Carlo, multilevel model of solid tumors’ behavior, integrating the novel approach of stem cell origin of cancer is presented. It involves patient individualized simulation of free tumor growth and response to therapeutic treatment taking into account imaging, histopathological molecular and clinical data. The work is part of the EC funded project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996). The special case of breast cancer treated with epirubicin is considered. However the model can be easily adapted to other types of cancer. Computational, logic and clinical experience based validation of the model is presented along with representative parametric studies. The model’s behavior substantiates its potential to serve as a basis for further development of the treatment planning support system following a strict clinical optimization and validation procedure.

I. INTRODUCTION ARIOUS computational models of tumor response to chemotherapy have been developed in order to assist

clinicians towards treatment optimization. Indicative

Manuscript received September 8, 2008. E. A. Kolokotroni is with the National Technical University of Athens,

Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected]).

E. Ch. Georgiadi is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected] )and the Foundation for Research and Technology – Hellas, Institute of Computer Science, Biomedical Informatics Laboratory, N. Plastira 100, Vassilika Vouton, GR-700 13 Heraklion, Crete, Greece.

D. D. Dionysiou is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (e-mail: [email protected]).

G. S. Stamatakos (corresponding author) is with the National Technical University of Athens, Institute of Communication and Computer Systems, In Silico Oncology Group, Iroon Polytechniou 9, Zografos, GR-157 80, Greece (phone: + 30 210 772 2288, fax: + 30 210 772 3557, e-mail: [email protected] ).

This work was supported by the European Commission under the project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996).

examples are the following: Chuang [1] presented a theoretical study of pharmacokinetic and cell kinetic models for cancer chemotherapeutic systems. Ozawa et al. [2] presented a pharmacodynamic model for the cell cycle phase-specific antitumor agents as well as for the cell cycle phase-nonspecific agents. Gardner [3] developed a computer model, the KInetically Tailored Treatment or KITT model, to predict drug combinations, doses, and schedules likely to be effective in reducing tumor size and prolonging patient survival.

In this paper an in vivo, four-dimensional, patient-specific, Monte Carlo, simulation model of breast cancer solid tumor growth and response to chemotherapy is presented.

II. A SHORT OUTLINE OF THE MODEL A description of the fundamental principles of the model

can be found in previous publications of the in silico oncology group [4]-[7]. The model has been extended to include novel aspects, such as the stem cell theory of cancer origin, drug pharmacokinetics, the “cell categories nomogram concept”, and has been adapted to simulate the special case of breast cancer and its neoadjuvent treatment with the single-agent epirubicin. These novel aspects are discussed below.

Cytokinetic model: One of the novelties embodied in the model refers to the types of cell categories considered. The consideration of stem cells (cells assumed to possess unlimited proliferative potential), limp cells (progenitor cells with limited proliferative potential of three divisions), and diff cells (terminally differentiated cells) is a considerable improvement in respect to previous publications of the In Silico Oncology group, which did not take into account limp and diff cells. Fig.1 describes the path followed by tumor cells from birth to differentiation and death through apoptosis or mitosis. Table I summarizes the code input parameters and the values assigned to them for the results presented in the present work unless otherwise stated.

Cell Categories Population Nomogram: A totally arbitrary initialization of the various cell categories populations (stem, limp cells etc) can lead to an abnormal free growth behavior, i.e. an unexpected decrease of its volume, followed by a volume increase (Fig. 2). In order to avoid such a behavior the tumor must be correctly initialized. A number of exploratory executions have been performed in order to

A 4-D Simulation Model of Tumor Free Growth and Response to Chemotherapy in Vivo: The Breast Cancer Case

Eleni. A. Kolokotroni, Eleni. Ch. Georgiadi, Dimitra. D. Dionysiou and Georgios. S. Stamatakos

V


32

determine the initial percentages of the various populations as they are regulated by model transition probabilities and cell cycle durations. An example is given in Table II. The grey area corresponds to a tumor that shrinks by itself and therefore cannot exist.

Epirubicin Pharmacokinetics: Pharmacokinetics describes a drug’s concentration in the body with time, as regulated by the mechanisms of absorption, distribution, metabolism and excretion. According to literature [8], epirubicin pharmacokinetics is best described by an open three-compartment model (Fig. 3). The parameter of interest, Area Under Curve (AUC), can be calculated based on the inter-compartment rate constants of the model for any given drug dose and volume of distribution. The inter-compartment rate constants were calculated from the fitting of experimental data [8] by the tri-phasic model using the SAAM II software tool [9].

Fig 1. General cytokinetic model for untreated tumor growth. A proliferating tumor cell (stem or limp) passes through the phases Gap 1 (G1), DNA synthesis (S), Gap 2 (G2) and mitosis (M). After mitosis the “daughter” cells, depending on nutrient supply and oxygenation conditions may enter the resting (dormant) G0 phase or (re-)enter G1 phase. A dormant or proliferating cell may die due to ageing and spontaneous apoptosis. A dormant cell can live for a period of TG0. If in the meantime local environmental conditions become adequate it re-enters G1 phase, otherwise it dies via necrosis. Diff cells may die due to apoptosis or necrosis. Stem cells can perform symmetric divisions, giving rise to two stem cells (preserving the properties of the parent cell), or asymmetric divisions, giving rise to one stem and one limp cell.

TABLE I SUMMARY OF CODE INPUT PARAMETERS AND THEIR VALUES THAT

CORRESPOND TO THE RESULTS INCLUDED IN THE PRESENT PAPER UNLESS OTHERWISE STATED

Symbol Description Value Tc

Cell cycle duration 60h TG1 Duration of Gap 1 phase 0.41(Tc - TM) TS Duration of DNA synthesis phase 0.41(Tc - TM) TG2 Duration of Gap 2 phase 0.18(Tc - TM) TM Duration of mitosis phase 1h TG0 Duration of dormant phase 96h TN Time period needed for necrosis’

products to disappear from the tumour 20h

TA Time period needed for apoptosis’ products to be removed from the tumour

6h

NLIMP Number of mitoses performed by 3

progenitor cells before they become differentiated

RA Apoptotic rate of cancer cells 0.001 RNDiff Necrotic rate of differentiated cells 0.001 RADiff Apoptotic rate of differentiated cells 0.001 PG0toG1 Fraction of dormant cells that re-enter the

cell cycle 0.01

PMtoG0 Fraction of cells that enter G0 phase following mitosis

0.1

Psym Fraction of stem cells that perform symmetric division

0.18

*Based on literature, breast cancer cell cycle duration can vary from 23h to 90h. A middle value of 60h has been considered.

1,00E+06

1,00E+07

1,00E+08

1,00E+09

1,00E+10

1,00E+11

0 500 1000 1500 2000 2500 3000 3500 4000 4500

Time (h)

Num

ber o

f cel

ls

Proliferating cells - Execution ATotal cells - Execution AProliferating cells - Execution BTotal cells - Execution B

Fig. 2. The effect of arbitrary cell populations’ initialization on tumor growth. Execution A, which corresponds to the “correct” population initialization, results in a smooth growth from the beginning of the execution. Execution B corresponds to a random population initialization.

Fig. 3 Three-compartment first order model. Ci, Vi are the concentration

and volume of each compartment, kel is the elimination rate constant from the central compartment, and k12, k21, k13, k31 are the rate constants describing drug transfer between the compartments

Epirubicin Pharmacodynamics: Epirubicin is an

anthracycline chemotherapeutic agent that inhibits DNA synthesis. In the model it is assumed that the drug is absorbed by a fraction of proliferating cells residing at all cycling phases. Affected cells die through apoptosis when reaching S phase. At a first approximation the survival fraction (SF) is computed from AUC based on experimental FDA (Food and Drug Administration) data concerning the in vitro cytotoxicity of epirubicin on HeLa cells [10], by means of linear interpolation. More specifically a fraction of the calculated AUC (e.g. 2AUC/3) is considered to account for

Cell Local reoxygenatio Local reoxygenatio

Cell disappearanc

Apoptosi

Spontaneous apoptosi

Necrosis

disappearanc

G0 G1 S G2 M G G0 M G2 S G1

A symmetri

division DIFF

STE LIMP

Symmetric division

Spontaneous apoptosi

After mitosis


33

inadequate drug penetration into the entire tumor [11]. TABLE II

NOMOGRAM: CELL CATEGORIES POPULATION PERCENTAGES

Symmetric division percent

Stem cells

Limp cells

Diff cells

Dead cells

10% 20% 0.037 0.133 0.810 0.021 30% 0.075 0.199 0.706 0.020 40% 0.130 0.248 0.603 0.019 50% 0.203 0.278 0.501 0.018 60% 0.298 0.284 0.400 0.018 70% 0.419 0.263 0.299 0.019 80% 0.570 0.212 0.199 0.020 90% 0.754 0.126 0.099 0.021

III. RESULTS Indicative parametric studies are presented. A

homogeneous spherical tumor with diameter equal to 14mm and cell density equal to 106cells/mm3 is considered. Two case studies, tumor free growth and response to chemotherapy, are simulated for three values of the parameter symmetric division fraction: 0.19, 0.20, 0.21. The drug is assumed to be administrated orally (bolus administration). The fragmentation scheme simulated is 100 mg/m2 i.v. once every 3 weeks for 4 consecutive cycles.

Fig. 4 shows the fluctuation of the total cell population with time. In accordance with clinical experience, the untreated tumor presents a monotonic exponential expansion, whereas the treated tumor shrinks as a result of each therapeutic session and is repopulated by proliferating tumor cells in-between sessions. The tumor behavior depends on the value of the symmetric division fraction. A tumor characterized by a higher fraction of stem cells performing symmetric divisions exhibits a lower doubling time, i.e. a more rapidly evolving nature. Such a tumor, as anticipated, responds poorly to the chemotherapeutic scheme applied. Indeed the tumor with the highest value of Psym has the lowest population reduction.

These findings suggest that by assigning different values to the various input code parameters, different cases regarding the effectiveness of a therapeutic scheme can be demonstrated.

IV. CONCLUSIONS A top-down, spatio-temporal, multilevel, Monte Carlo

simulation model of breast cancer solid tumor free growth and response to epirubicin-based chemotherapeutic treatment in vivo has been presented. Novel approaches have been integrated such as the distinction between stem cells, exhibiting unlimited proliferative potential, and progenitor cells, undergoing a limited number of mitoses, and the

correlation between cell categories relative populations and code input parameters.

Total cells

0.00E+00

5.00E+08

1.00E+09

1.50E+09

2.00E+09

2.50E+09

3.00E+09

3.50E+09

4.00E+09

0 10 20 30 40 50 60 70 80 90

Time (days)

Num

ber o

f cel

ls

Free growth, Psym=0,19Therapy, Psym=0,19Free growth, Psym=0,20Therapy, Psym=0,20Free growth, Psym=0,21Therapy, Psym=0,21

Fig. 4. Total number of tumor cells as a function of time for the following values of the symmetric division fraction (Psym): 0.19, 0.20 and 0.21. Free growth and therapeutic response are simulated. The dose of each fraction is 100mg/m2. The administration times are accompanied by a rapid decrease of the cell categories populations (at days 0, 21, 42 and 63).

The model successfully simulates characteristics of tumor

behavior such as tumor repopulation, expansion and shrinkage. Extensive parametric studies in order to better understand the model’s sensitivity to input code parameters variation are in progress. Quantitative validation and optimization based on pertinent clinical and laboratory data, for example e.g. size reduction after therapy, gene expression etc. is also under way.

The ultimate goal is the development of a reliable simulation tool that would enable clinicians to draw optimized therapeutic schemes, adapted to the patient’s clinical data. The model could also serve as a valuable tool for researchers, professionals or patients to gain insight into the biological mechanisms involved in tumor growth and response to chemotherapy in vivo.

ACKNOWLEDGMENT

The authors acknowledge the support provided by C. Desmedt, Institut Jules Bordet, Belgium and N. Graf, University of Saarland.

REFERENCES [1] S. Chuang, “Mathematic models for cancer chemotherapy:

pharmacokinetic and cell kinetic considerations,” Cancer Chemother. Rep.,vol. 59, no. 4, pp. 827–42, 1975.

[2] S. Ozawa, Y. Sugiyama, J. Mitsuhashi, and M. Inaba, “Kinetic analysis of cell killing effect induced by cytosine arabinoside and cisplatin in relation to cell cycle phase specificity in human colon cancer and chinese hamster cells,” Cancer Res., vol. 49, pp. 3823–3828, 1989.

[3] S. N. Gardner, “Modeling multi-drug chemotherapy: Tailoring treatment to individuals,” J. Theor. Biol., vol. 214, pp. 181–207, 2002.


34

[4] G.S. Stamatakos, D.D. Dionysiou, E.I. Zacharaki, N.A. Mouravliansky, K.S.Nikita, N.K. & Uzunoglu , In silico radiation oncology: combining novel simulation algorithms with current visualization techniques , IEEE Proceedings: Special Issue on Bioinformatics: Advances and Chalenges , 90(11) , 1764-1777 , 2002.

[5] D. D. Dionysiou, G. S. Stamatakos, N.K. Uzunoglu, K. S. Nikita, A. Marioli, “A four-dimensional simulation model of tumour response to radiotherapy in vivo: parametric validation considering radiosensitivity, genetic profile and fractionation,”Journal of Theoretical Biology vol. 230, pp. 1–20, 2004

[6] G. S. Stamatakos, V. P. Antipas and N. K. Uzunoglu, “A Spatiotemporal, Patient Individualized Simulation Model of Solid Tumor Response to Chemotherapy in Vivo: The Paradigm of Glioblastoma Multiforme Treated by Temozolomide,” IEEE Transactions On Biomedical Engineering, vol. 53, pp. 1467- 1477, August 2006

[7] G.S.Stamatakos, D.D.Dionysiou, N.M.Graf, N.A.Sofra, C.Desmedt, A.Hoppe, N.Uzunoglu, M.Tsiknakis , “The Oncosimulator: a multilevel, clinically oriented simulation system of tumor growth and organism response to therapeutic schemes. Towards the clinical evaluation of in silico oncology,” Proceedings of the 29th Annual International Conference of the IEEE EMBS Cite Internationale, August 23-26, SuB07.1: 6628-6631 , Lyon, France , 2007

[8] R.Danesi, F.Innocenti, S.Fogli, A.Gennari, E.Baldini, A.Di Paolo, B.Salvadori, G.Bocci, P.F.Conte and M.Del Tacca, “Pharmacokinetics and pharmacodynamics of combination chemotherapy with paclitaxel and epirubicin in breast cancer patients”, Journal of Clinical Pharmacology, vol. 53, pp 508-518, 2002.

[9] http://depts.washington.edu/saam2/ [10] http://www.fda.gov/cder/foi/nda/99/50-778_Ellence_pharmr.pdf [11] J.Lankelma, H.Dekker, R.F. Luque, S.Luykx, K.Hoekman, P.van der

Valk, P.J.van Diest, H.M.Pinedo. Doxorubicin gradients in human breast cancer,” Clin Cancer Res, vol. 5, pp. 1703-1707, 1999.


35

Geometrical and Mechanical Aspects of Tumor Growth and Response to Chemotherapeutic Schemes in the Context of the

ACGT Oncosimulator Stavroula. G. Giatili, Georgios. S. Stamatakos, Dimitra D. Dionysiou, Eleni A. Kolokotroni, Eleni Ch.

Georgiadi

Abstract—Major advances in the study of several tumor types using various medical imaging modalities have been achieved over the past decades. Coupling multilevel tumor dynamics modeling with medical imaging techniques has proved a particularly powerful approach for the future eventual clinical translation of cancer models. In this context a major component of the European Commission funded project “Advancing Clinico-Genomic Trials on Cancer” (ACGT) (FP6-2005-IST-026996) is the development of patient specific four dimensional multilevel cancer models simulating nephroblastoma and breast cancer tumor response to chemotherapeutic schemes. The models are based on available imaging, histopathological, molecular and clinical data and take into consideration the effect of cancer stem, progenitor, differentiated and dead cells. In this paper a brief outline of a number of novel algorithms simulating the geometrical and mechanical aspects of tumor

growth and response to chemotherapeutic schemes is presented. The algorithms relate to the spatial translation, expansion and shrinkage of the metabolic layers of a tumor.

Manuscript received July 30, 2008. S.G.Giatili is with the In Silico Oncology Group, MFOL, Institute of

Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: +30 210 772 2288; fax +30 210 772 2288; e-mail: [email protected]).

G. S. Stamatakos (corresponding author) is with the In Silico Oncology Group, MFOL, Institute of Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: + 30 210 772 2288, fax +30 210 772 2288; e-mail: [email protected] ).

D. D. Dionysiou is with the In Silico Oncology Group, MFOL, Institute of Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: + 30 210 772 2288; fax +30 210 772 2288; e-mail: [email protected]).

E. A. Kolokotroni is with the In Silico Oncology Group, MFOL, Institute of Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: + 30 210 772 2288; fax +30 210 772 2288; e-mail: [email protected]).

E. Ch. Georgiadi is with the In Silico Oncology Group, MFOL, Institute of Communication and Computer Systems, School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Str., 15780 Zografos, Athens, Greece. (phone: + 30 210 772 2288; fax +30 210 772 2288; e-mail: [email protected] ) and the Foundation for Research and Technology – Hellas, Institute of Computer Science, Biomedical Informatics Laboratory, N. Plastira 100, Vassilika Vouton, GR-700 13 Heraklion, Crete, Greece

This work was supported by the European Commission under the project “ACGT: Advancing Clinicogenomic Trials on Cancer” (FP6-2005-IST-026996), http://www.eu-acgt.org/

I. INTRODUCTION

VER the past years considerable efforts have been made in order to simulate tumor growth and response to

chemotherapeutic agents. These efforts have led to the development of several simulation algorithms. In this context In Silico Oncology Group, Institute of Communication and Computer Systems, National Technical University of Athens is leading the action of the Oncosimulator development within the framework of the European Commission (EC) funded ACGT (Advancing Clinicogenomics Trials on Cancer) project [http://www.eu-acgt.org/]. The Oncosimulator aims at optimizing cancer treatment on a patient-individualized basis by performing in silico (on the computer) experiments regarding several candidate therapeutic schemes. The paper focuses on the modeling of the geometrical and mechanical aspects of tumor dynamics as these are integrated into the Oncosimulator. Specific topics include algorithmic procedures for the expansion and shrinkage of the metabolic layers of the tumor and the effective handling of artifact holes that usually emerge during internal restructuring of the virtual tumor.

II. GEOMETRICAL AND MECHANICAL ASPECTS A. Handling of the Eventual Artifact Holes of the Virtual Tumor

The unit volume of a discretizing mesh covering the region of the tumor has been termed a Geometrical Cell (GC). The size of a geometrical cell is linked to the size of the voxel of the available imaging data [1]-[4]. As a first approximation the expansion of a virtual tumor during free growth and its shrinkage due to treatment response should be algorithmically achieved in such a way that the tumor retains its initial shape (conformal shrinkage). Conformal shrinkage seems to happen quite frequently in treated tumors. This change is computationally achieved by appropriate movements of the geometrical cells of the discretizing mesh

O


36

of the problem. Computational movement of geometrical cells usually leads to the emergence of some geometrical cells that are empty. These cells are visible in the 3D tumor reconstruction as grey dots inside the tumor mass. The problem of such artifact holes has been solved by the implementation of an algorithm based on the evaluation of the number of neighbor geometrical cells of the empty geometrical cell. In that case extra geometrical cell movement takes place in order to fill in the gap. Fig. 1 depicts the results of relevant simulations. Equatorial virtual slices (sections) of the tumor are presented. The ability of the model to adequately fill in the artifact gaps and therefore produce a compact tumor is clear.

B. Movement of Metabolic Layers The development of the simulation model is based on

pertinent imaging, histopathological, molecular and clinical data of the patient. On the imaging data (such as CT, contast enhanced MRI slices etc.), especially when a contrast agent is used (e.g. gadolinium in T1 MRI), two gross regions within the tumor are usually distinguished based on their levels of brightness One region appears lightly coloured whereas another one dark colored. The tumor delineation by the clinician is mainly based on such a brightness factor. By carefully exploiting this type of information the clinician may be able to distinguish between a well and a poorly oxygenated region of the tumor.

Fig. 1. Slices of virtual tumors, illustrating the geometrical and mechanical aspects of tumor growth and response to chemotherapeutic schemes. The left column of images depicts an equatorial section of a layered spherical tumor of initial diameter 50mm whereas the right column of images depicts an equatorial section of a layered tumor with arbitrary shape. A. Initial tumors. B. Tumors at the 41th day of therapy. The artifact holes inside the tumor mass are presented as grey dots. C. The same tumors at the same day of therapy, after the implementation of the algorithm eliminating artifact holes. The tumor mass appears compact. D. The virtual tumors after 71 days without therapy and without the implementation of the artifact hole elimination algorithm. Expansion of the tumor has led to the creation of the artifact holes. E. The virtual tumors after 71 days without therapy but with the implementation of the artifact hole elimination algorithm. The internal area appears compact. In the images A-E the necrotic area which is painted with white colour remains stable with time. That rather unreaslistic situation has been achieved by the implementation of the algorithm for the movement of the metabolic layers as shown in images F and G. Image F refers to tumor shrinkage whereas image G refers to tumor expansion. The size of the necrotic layer changes with time whereas its dimesions depend on the whole tumor volume and shape.

A.

B.

C.

D.

E.

F.

G.


37

The necrotic region may contain only a very small number of functioning neovasculature microvessels. This usually corresponds to a dark (nearly black or black) area in a (contrast enhanced) MRI slice. The proliferative area usually corresponds to a light (nearly white or white) area. The latter contains an extensive functioning microvessel network and is characterized by the production of vascular endothelial growth factors [5]. According to [6] tumor growth rate is assumed to depend on the vascular density. The rate of cell production, cell survival, cell differentiation, stem cell division, multipotent progenitor cell division etc. are influenced by the microenvironment primarily referring to oxygen and nutrient availability.

As a result of that there is need for the creation of two regions in the model termed layers, the proliferative layer and the necrotic layer. The basic difference between them is growth rate of tumor cells including proliferation, loss and death. Management of the shrinkage and expansion of tumor regions being in contact with practically non deformable anatomical structures (such as bones) has proved to be quite demanding.

Expansion of the necrotic and the proliferative layer is of great importance because the cell category transition rates and the cell death due to apoptosis or necrosis depend on the layer that each cell is belonging to.

Considerable improvement has been made in that aspect and the problem has been successfully managed. The algorithm proposed and applied is based on the evaluation of the distances of the cells from the center of the mass of the tumor and the fluctuation of the percentage of the total number of tumor cells. In Fig 1. two different types of macroscopic tumor structures are presented. The pattern of growth and shrinkage of the tumor regions (necrotic and proliferative) is depicted not only for a simple spheroidal structure but also for a tumor with a more complex geometry and internal macroscopic metabolic structure. The macroscopic features of the result of the simulation suggest the effectiveness of the algorithm we applied.

III. DISCUSSION During the initial stages of the Oncosimulator development

three metabolic layers of the tumor had been assumed [1-[4]. However as only two metabolic regions are usually easily distinguisable by the clinician we have resorted to the consideration of two regions. This is also in agreement with [6] for a special case of breast cancer.

Following the implementation of the above mentioned algorithms the entire simulation model showed a remarkable refinement concerning tumor compactness. An extensive

parametric study of the algorithm bahavior is under way.

ACKNOWLEDGMENT The authors acknowledge the support provided by. N. Graf,

University of Saarland and N. Uzunoglu, School of Electrical and Computer Engineering, National Technical University of Athens.

REFERENCES [1] G.S. Stamatakos, D.D. Dionysiou, E.I. Zacharaki, N.A. Mouravliansky,

K.S.Nikita, N.K. & Uzunoglu , In silico radiation oncology: combining novel simulation algorithms with current visualization techniques , Proceedings of IEEE: Special Issue on Bioinformatics: Advances and Chalenges , 90(11) , 1764-1777 , 2002.

[2] D. D. Dionysiou, G. S. Stamatakos, N.K. Uzunoglu, K. S. Nikita, A. Marioli, “A four-dimensional simulation model of tumournresponse to radiotherapy in vivo: parametric validation considering radiosensitivity, genetic profile and fractionation,” Journal of Theoretical Biology vol. 230, pp. 1–20, 2004

[3] G. S. Stamatakos, V. P. Antipas and N. K. Uzunoglu, “A spatiotemporal, patient individualized simulation model of solid tumor response to chemotherapy in vivo: the paradigm of glioblastoma multiforme treated by temozolomide,” IEEE Transactions on Biomedical Engineering, vol. 53, pp. 1467- 1477, 2006

[4] G.S.Stamatakos, D.D.Dionysiou, N.M.Graf, N.A.Sofra, C.Desmedt, A.Hoppe, N.Uzunoglu, M.Tsiknakis , “The Oncosimulator: a multilevel, clinically oriented simulation system of tumor growth and organism response to therapeutic schemes. Towards the clinical evaluation of in silico oncology,” Proceedings of the 29th Annual International Conference of the IEEE EMBS, August 23-26, SuB07.1: 6628-6631 , Lyon, France, 2007

[5] Ivan L Cameron, Nicholas Short, LuZhe Sun and W Elaine Hardman, “Endothelial cell pseudopods and angiogenesis of breast cancer tumors,” Cancer Cell International 2005, 5:17 doi:10.1186/1475-2867-5-17, Published: 26 May 2005

[6] Stoll, B.R., Migliorini, C., Kadami, A., Munn, L.L., Jain, R.K., 2003. A mathematrical model of the contribution of the endothelial progenitor cells to angiogenesis in tumour: implications for antiangiogenic therapy. Blood 102(7), 2555-2561.


38

Parallelizing the ACGT OncoSimulator

Dominique Lavenier and Julien Jacques

Abstract--Nowadays, a clinician can take advantage of in-silico technologies to get faster estimations on clinical treatments without performing time-consuming and expensive in vivo experimentations. Among these techniques, tumor growth simulators can estimate tumor volume and the quantity of cells (stem, proliferating etc.) as a function of time based on a set of parameters. Within the framework of the ACGT European project [2], we are working on the development of the ACGT Oncosimulator which is based on simulation models developed by ICCS. In this paper, we describe current computational techniques for increasing the ACGT Oncosimulator efficiency.

I. MOTIVATIONS

O have a powerful simulator, the underlying program must be reliable, precise and fast. Firstly, the reliability comes from the consistence of the mathematical model

compared with in vivo experiments. Second, the precision depends of the sampling area, that is the spatial 3D unit dedicated to the tumor discretization. Finally, speed is both related to the quality of the code and to the performance of the hardware resources. This paper deals with the last point.

Initially, the Oncosimulator has been developed for a stand-alone PC equipped with a single core processor. This approach obviously limits the program to sequential executions, even with today machines which are now integrating double or quad core processors. Furthermore, one of the goals of the ACGT project is to deploy a European computational grid able to support fast execution of the Oncosimulator. It was thus desirable to adapt the Oncosimulator code to benefit from the use of the ACGT grid power together with the latest technology improvements.

Basically, the ACGT end-users may exploit the Oncosimulator as follows:

• intensively: to find the best treatment, the clinician has to try many combinations of parameters provided by the Oncosimulator (size of the tumor, duration of the treatment, quantity of drugs, interval between two injections, etc.). In that case, many runs with different parameters have to be executed, and the execution times are directly correlated with the number of runs. It may be not unusual to have a few hundred of parameter combinations to test.

• with high precision: the better the definition of the tumor, the better the estimation of the tumor behavior. In our case, the definition of the tumor is directly linked to the 3D discretization. From a computational point of view, shrinking the discretization by 10 will increase the computational complexity by 1000 since the modelization operates in the 3D space.

• interactively: another way to focus to the right treatment is to run successive simulations guided by a human expert. Depending of the results of one simulation, the clinician will slightly tune a few parameters according to the tumor evolution. In that specific case, the response time of the simulator is critical. Ideally, an execution should not take more than ten to twenty seconds in order to have an efficient interacting tool.

In the rest of the paper, we explore how the various technologies available today can support these different kinds of requirements [1].

II. TECHNOLOGIES A. Grid computing

Grid computing consists in a set of machines (called nodes) geographically spread and connected through Internet (Fig. 1). A grid generally provides a high computing power and a huge storage capacity. It is accessed through server managers as, for instance, the Globus-SGE [3] environment. These servers drive the node allocations according to the grid workload.

Grids are well adapted to simultaneous executions of many independent programs. The execution time of one program is the same as if it is executed into a single machine. However, since several programs are run in parallel, the global time T needed to achieve the whole computation corresponds to:

g

Deploying a code on a grid is easy since no

modification is required. Only the execution management is crucial in order to distribute program instances on the available resources.

T

Tg = Tp ×Np

Nnodes

⎡

⎢ ⎢

⎤

⎥ ⎥ ,

where Tp : execution time of single programN p : number of single programs Nnodes : number of nodes

⎧

⎨ ⎪

⎩ ⎪

D. Lavenier is with Symbiose Project Team, CNRS, Irisa, Campus de Beaulieu, 35000 Rennes, France (e-mail: [email protected])

J. Jacques is with Symbiose Project Team, INRIA, Irisa, Campus de Beaulieu, 35000 Rennes, France (e-mail: [email protected])

This work is supported by the European Community Framework Programme for Research, Technological Development and Demonstration (FP6)

mailto:[email protected]


39

The grid solution is well adapted to the first use of the Oncosimulator: a lot of executions can be run in parallel with different parameters, each run being assigned to a different machine. Results are automatically collected back to the user.

B. Cluster

A cluster is a set of identical machines connected through a high-speed network (Fig. 2). From an implementation point of view, there are basically two possibilities to execute a code on this support. The first one follows the grid idea: many instances of the same program are dispatched on the different nodes of the cluster. Difficulties and implementation are thus identical.

Fig 1. Example of a Grid infrastructure with repartition of the computational power and the storage capacity

The second one required to modify the code for parallelizing the various parts of the program on several nodes. To get a fast program, computations must be shared between the processors in order to have a number of communications and synchronizations as small as possible.

Fig 2. Single Program Multiple Data (SPMD) Architecture

That makes the parallelization task quite difficult: Even by assuming well adapted distribution of data in the processor memories, it always remains, in many cases, a part of the program which cannot be parallelized and have to be done sequentially. We also need to keep in mind that communications add costs between processors, and these costs do not exist in the monoprocessor version. As a consequence, the execution time can be stated as follows:

commasksparallel_t

// CTN

TT seq

p

++=

In the case where Tg < (N p × T// ) then the parallelization is inadequate.

As for the grid technology, clusters are well adapted for running a large set of independent simulations with different parameters. If a smart parallelization can be done at the cluster level, this technology is thus able to satisfy all end-user requirements.

C. Multicore processors

Fig. 3 Two-core processor organization

A multicore processor is a component having, on the

same die, several cores -- or processing units -- connected to the same memory (Fig. 3). In this architecture, the use of threads (or light processes) enables to explicitly express in the source code the parallel execution of various parts of the program.

Using this technology, we need to extract parallel blocks that do not have data, time and spatial dependencies. In a sequential program, the process is composed of only one thread. Therefore, its execution on a multicore processor will be done only on one core. In a multithreaded program, the process is divided into many threads. As a result, the execution can be shared among all the cores.

Even if this solution covers all the end-user requirements for speeding up the simulation, it is much better suited for accelerating a single instance of the Oncosimulator. With the next generation of processors, a larger number of cores will certainly be available, providing significant increase of the computing power. Thus, this is a necessity for the next version of programs, like Oncosimulator programs, which require high computational power to use this new opportunity.

III. APPLICATION TO THE ACGT ONCOSIMULATOR

As a first approach, we have made a cluster implementation on the INRIA Genouest bioinformatics platform where many instances of the Oncosimulator can be run simultaneously. Their submissions are done through a web browser at http://acgt.genouest.org.

The next step has been to integrate this work through the ACGT grid environment. The feedback from the ACGT end-users is encouraging since this implementation meets the need of simulating the tumor evolution according to a large set of parameters.

However, the execution time of each instance takes several minutes, leading to a non-interactive tool. Then, we first tried to develop a MPI version, but the code we developed was requiring too much synchronization points to get this version really efficient.


40

We are currently developing a multithreaded version targeted multicore processors with the objective to fall below an execution time of ten seconds.

The last point, which needs to be highlighted, is that all these technologies are not mutually exclusive. On the contrary! Grids are made of clusters which now include multicore processor nodes. Thus, parallelization improvement made on one of these technologies will have a direct impact on the whole ACGT project.

REFERENCES

[1] M. Creel and W.L. Goffe, Multi-core CPUs, Clusters and Grid Computing: a Tutorial, Society for Computational Economics, Computing in Economics and Finance 2005, 438.

[2] ACGT Project website, http://eu-acgt.org [3] The Globus Alliance website, http://www.globus.org


41

Applying Grid Computing Technologies to In Silico Oncology Theodoros Athanaileas, Andreas Menychtas, Dimitra Dionysiou, Georgios Stamatakos, Dimitra

Kaklamani, Theodora Varvarigou, Iakovos Venieris and Nikolaos Uzunoglu

Abstract—This paper presents the application of grid computing techniques to In Silico Oncology. Grid computing is a technology that has come forward as an efficient way for addressing the increasing demands of scientific applications in computational power. In the work presented herein, the simulation code for a radiotherapy model has been migrated to the EGEE grid infrastructure. A web based environment has been designed and developed, which provides advanced features and mechanisms that facilitate the grid-enabled execution of the simulations. Several scenarios of radiotherapy simulations have been performed on the EGEE grid and execution statistics are presented.

I. INTRODUCTION RID computing has emerged in the last years as a technology for large-scale, flexible and coordinated

resource sharing and in some cases with an orientation towards high performance computing [1]. Computational grids enable the sharing of a wide variety of geographically distributed resources across multiple organizations for solving large scale computational intensive and data intensive problems. In the context of the work presented in this paper, the simulation of a radiotherapy model has been enabled for execution in a grid infrastructure.

The radiotherapy model has been developed by the In Silico Oncology Group, NTUA, and is a four-dimensional simulation model of glioblastoma multiforme (GBM) [2, 3] response to radiotherapy. The simulation model may be used for evaluating and comparing different radiotherapeutic schemes while testing several critical parameters which present inter-patient variability, thus revealing the relative merits of various radiotherapeutic schedules in a patient-specific manner.

Manuscript received July 30, 2008. All authors are with the School of Electrical and Computer Engineering,

National Technical University of Athens, 9 Heroon Polytechneiou Str., 15773, Zografou, Athens, Greece; corresponding author: T. Athanaileas (phone number: 0030-210-7722287; email: [email protected])

This work has been performed in the context of the research project “Development and adaptation of an in silico oncology application in grid environment” (GRID-APP). The project is funded by the General Secretariat for Research and Technology, Ministry of Development, Greece (75%) and the European Regional Participation Fund (25%).

A simulation environment for performing radiotherapy parametric studies on the grid infrastructure provided by the EGEE project [4] has been developed. It builds on the gLite middleware [5] developed by EGEE and provides a web-based grid portal for enabling interactions with it in a simple and user-friendly way. The application – portal is enhanced with added functionality in order to simplify the job submission process and automate the interaction with the services of the grid infrastructure. In that way more users are able to access the computational resources while the administrators manage the application and monitor the operational status.

II. GRID COMPUTING

A. The EGEE infrastructure The grid-enabled environment presented in this paper is

built on the production grid deployed by the EGEE project. The EGEE (Enabling Grids for E-sciencE) project [4] is funded by the European Commission and aims to build on recent advances in grid technology in order to develop and deploy a service grid infrastructure which is available to scientists 24 hours-a-day. The EGEE infrastructure consists of computing resources distributed among 32 countries world-wide that belong to different virtual organizations according to research subject or geographical location.

The EGEE infrastructure is built on the gLite middleware [5], which provides mechanisms for security, user accounting, authentication and authorization, data access and workload management.

B. Grid – added value for In Silico Oncology Exploitation of grid technologies is beneficial to in silico

oncology for the following reasons: (a) exponential increase of required computational resources when considering a more dense discretization of the space-time (4D) grid of the biological problem, (b) heterogeneity of required data (imaging, histopathologic, genetic) with different preprocessing requirements and (c) large number of involved patients.

Exploitation of the vast resources provided by a grid may lead to a better understanding of the biological and clinical

G


42

behavior of cancer and especially solid tumours. Furthermore, computer simulation may be employed in order to optimize treatment of cancer, by conducting a number of simulations for different therapeutic schemes based on the individual data of a patient. Since simulations need to be conducted in clinically accepted computational time, exploiting grid computing is a very attractive solution, as the resources provided in a grid infrastructure may be efficiently used to reduce overall required execution time in a cost-effective and efficient manner.

III. THE GRID-ENABLED ENVIRONMENT

A. Architecture An educational grid-enabled environment for in silico

oncology has been developed, which may be used by doctors and researchers for evaluating and comparing the behaviour of different simulation codes (e.g. therapeutic schemes) and the ways they are affected by different input parameters. The development of the environment entailed code modifications, implementation of certain helper programs and wrapper scripts for facilitating the execution of the simulations on grid nodes, as well as creation of job description files, which are expressed according to a Job Description Language [6] and which are used by the gLite workload management system for job submission and job match making process.

The environment was designed and developed as a web

portal, following a multi-tier architectural approach. This approach defines different layers for the operations and functions of the application framework, simplifying the installation and the maintenance processes. The grid enabled In Silico oncology application belongs to the category of the applications that require enterprise functionality and usability at the same time, since it is targeted to people that are not

computer experts and are not familiar with grid technologies. Multi-tier architectures have all the characteristics for building this kind of applications.

The simulation application consists of four layers: the presentation layer (which includes all the functionality for the interaction with the end users and is presented to the users through java server pages – JSP [7]), the portal services layer (which includes all the functionality of the application, handles user and file management, and establishes the connection between the presentation and the gLite and database layers), the gLite layer (which includes all the functionality for the communication with the grid services and resources) and the database layer (which includes a database keeping all the data regarding the application and user management).

The application architecture is depicted in Figure 1.

B. Experiments

The grid-enabled environment has been used in order to perform a series of simulations corresponding to the various arms of the RTOG study 83-02 [7]. Six radiotherapeutic schemes have been considered; each scheme consists of 77 combinations of input parameters and for each parameter combination a job has been created and submitted for execution to the grid, thus each scheme consists of 77 independent jobs submitted simultaneously to the resources provided by the South Eastern Europe Virtual Organization (SEE-VO) [8] of the EGEE infrastructure.

Table I shows some statistics regarding the execution times of the simulations on the grid, as well as the speedup obtained.

TABLE I EXECUTION TIMES OF RADIOTHERAPY SIMULATIONS ON THE GRID

Mean Job Execution

Time

Overall Schema

Execution Time Schema # Speedup

1 ~32 mins ~58 mins ~42

2 ~31 mins ~59 mins ~40

3 ~34 mins ~59 mins ~44

4 ~38 mins ~72 mins ~41

5 ~36 mins ~66 mins ~42

6 ~21 mins ~47 mins ~34

Fig. 1. Application Architecture

IV. CONCLUSION In this paper, the application of grid technologies to the

domain of In Silico Oncology has been presented. The vast resources available in a grid enable the evaluation and comparison of different therapeutic schemes at clinically


43

acceptable time, while the access to these resources was considerably simplified through a web portal.

REFERENCES [1] I. Foster, C. Kesselman, S. Tuecke. “The Anatomy of the Grid:

Enabling Scalable Virtual Organizations.”, International Journal of Supercomputer Applications, vol. 15, No. 3, pp. 200-222, 2001.

[2] G.S. Stamatakos et al. “In silico radiation oncology: combining novel simulation algorithms with current visualization techniques”. IEEE Proceedings, Special Issue on “Bioinformatics: Advances and Chalenges”, vol. 90, pp. 1764-1777, 2002.

[3] D.D. Dionysiou et al . “A Four Dimensional In Vivo Model of Tumour Response to Radiotherapy: Parametric Validation Considering Radiosensitivity, Genetic Profile and Fractionation.” J. theor. Biol., vol. 230, pp. 1-20, 2004.

[4] The EGEE Project, homepage: http://www.eu-egee.org/

[5] The gLite middleware, http://glite.web.cern.ch/glite/

[6] Job Description Language (JDL) Attributes Specification, https://edms.cern.ch/document/590869/1/

[7] M. Werner-Wasik et al. “Final report of a phase I/II trial of hyperfractionated and accelerated hyperfractionated radiation therapy with carmustine for adults with supratentorial malignant gliomas”, Cancer , vol. 77, pp. 1535-1543, 1996

[8] The South-Eastern Europe Virtual Organization (SEE-VO), http://www.egee-see.org/see-vo.php?language=en


44

Multi-Level Image Analysis for Extracting Pathophysiological Parameters Related to Cancer Modeling

Kostas Marias, Member, IEEE

Abstract— Recent research trends focus on how multiscale biomedical information can be modeled and transformed into knowledge, in order to lead to a less interfering but also more individualized diagnosis and therapy. In order to assess the clinical importance of models of human pathophysiology, a common problem is the extraction of the necessary parameters from different scales. This paper discusses some of important image analysis challenges for extracting more accurate and precise cancer-related biomedical information and parameters from different modalities and scales. This is a crucial step for initializing and validating individualized models of human function and disease.

I. INTRODUCTION N mportant research area in biomedical image analysis is the robust extraction of biomedical information from

different levels (from molecular to tissue/organ). This is also a crucial step for the computation of 4D maps of pathophysiological properties of tumors for model development and validation.

Medical Imaging has focused in providing anatomical information, mainly imaging human bones, dense tissue and arteries. Recent advances especially in PET and functional MRI allowed the study of various pathological processes via radio-labelled tracers (PET) or pharmaco-kinetic models in contrast enhanced MRI. The whole field of molecular medicine and molecular imaging is opening up new possibilities for targeted assessment of disease and disease mechanisms. Also, microarray imaging has created exiting possibilities for measuring gene differential expression and defining new disease biomarkers. However, the robust extraction of physiological parameters remains an open issue since computing them from measurements (e.g. pixel values), isn’t a trivial task. For example, the physics and the heterogeneous steps involved often lead to non-linearities resulting in qualitative rather than quantitative information.

Manuscript received September 1, 2008. Dr. K. Marias is a researcher at the Institute of Computer Science at

FORTH, Vassilika Vouton, GR-70013 Heraklion, Crete, Greece ([email protected])

K. Marias acknowledges support from EC Contra Cancrum Project ICT-2007-223979.

Also, each imaging technology inherently emphasizes certain aspects and provides less sensitivity to others. Last, even the most sophisticated MRI and CT techniques have only limited spatial resolution (typically an order of magnitude below millimeter). Given the local complexity and “multi-scale” nature of human physiology, any attempt to use imaging to capture detailed physiological processes will require significant post-processing to recover the desired pathophysiological parameters measured.

Concluding, from the imaging standpoint it is essential to consider the following points relevant to human pathophysiology modeling:

a) There is a need for a holistic understanding of pathophysiology and this clearly implies a multidisciplinary approach. To this end, molecular and genetic imaging offer unique opportunities to better understand pathophysiology in smaller scales and built multilevel models.

b) It is necessary to pre-process biomedical data at all possible scales (e.g. medical images, microarray scans) in order to extract all the information that is needed for a given model. This way, multiscale information extraction aims to ‘individualise’ a given model.

The next section focuses on specific issues regarding the previous points summarizing recent advances of our group.

II. GENETIC IMAGE ANALYSIS DNA microarray technology has had a profound impact of

on biomedical research as it allows the concurrent observation of the expression of all known genes. This is of particular interest in cancer modeling since accurate molecular classification of the disease as well as temporal alterations can improve the accuracy of individualized models of cancer. However, microarray images consist mostly of low-intensity spots that are not well distinguishable from the background and it is often the case that important information maybe missed or even altered in the analysis/expression quantification process. These low-intensity spots are affected by inherent additive and multiplicative noise components. At the same time, due to the heterogeneous steps involved the results aren’t always reproducible and the measured expression can vary

A

mailto:[email protected]


45

depending on the experimental conditions. Our group has worked in developing methods for obtaining more robust result in microarray image analysis. More specifically:

- Improved Microarray Spot Segmentation by Combining

two Information Channels In microarray imaging technologies, several non-linearities

in the experimental process render the measured expression values prone to variability and often, to poor reproducibility. Accurate segmentation of the true signal is a very important task, not least because a single value per spot needs to be derived for further knowledge discovery analysis. We have presented [1] a fully automatic segmentation method for improving the spot segmentation result. The method doesn’t make any assumptions concerning the number of classes present in each image spot, and it isn’t driven only by the most intense features, since it takes into account the underlying “hybridization ground truth” derived from both information channels of the spotted arrays. Our method was compared to widely used, state-of-the-art segmentation methods in microarray image analysis in a study of a metabolic disorder in yeast, where replicates of reporters are present. Initial results indicate that our method yields more reproducible log ratio measurements across replicates.

- Microarray Image Denoising Using a Two-Stage

Multiresolution Technique DNA microarrays have demonstrated an excellent potential

in correlating specific gene expression profiles to specific conditions (e.g., disease). However, they are affected by inherent noise. We have presented a paper introducing a two-stage approach for noise removal that processes the additive as well as the multiplicative noise component [2]. The proposed approach first decomposes the signal by a multiresolution transform and then accounts for both the multiscale correlation of the subband decompositions and their heavy-tailed statistics. Real microarray images have been processed by the proposed method and its improved performance was assessed through quantitative measures and qualitative visual evaluation. Our results suggest that the method enhances the dynamic range of existing microarray imaging technology, which is very important in order to identify the most significant genes with increased accuracy and robustness.

III. DATA FUSION IN MOLECULAR AND TISSUE LEVEL IMAGING

Part of the objective of image fusion is to better enable the communication and comparison of pathological indicators

between temporal or multi-modal data. This has been traditionally relevant for comparing tomographic data over time (e.g. CT, MRI), in order to assess changes in e.g. tumour size in a geometrically consistent fashion but also to address changes in relevant physiological parameters (e.g. angiogenic leakage permeability can be assessed with CE-MRI). Data fusion is also an important step in molecular imaging studies and our group has developed novel methods for assessing molecular activity over time and improve the accuracy of monitoring fluorophore distribution [3,4]. In [4], we reported a novel automatic method for aligning 3D temporal data of small animals, based on the robust detection of surface anatomical features. It uses a dynamic time warping algorithm to define regions of similar curvature and since it is based on 3D surface landmarks the method can be used for mono-modal or multi-modal fusion (Fig. 1 shows results for the multi-modal case between CT and Fluorescence Molecular Tomography - FMT), in order to account for differences in the positioning and compression of small animals. This can be of particular interest in applications that need several imaging sessions, such as cell trafficking, tumor growth factor, and others, potentially improving the resolution and quantization.

Fig. 1. 3D multi-modal registration and data fusion (CT-FMT) for assessing molecular activity over time

IV. CONCLUSION The extraction of useful anatomical and physiological information from biomedical data is important for initializing (e.g. voxel classification as ‘proliferating’, ‘necrotic’, etc.), improving and validating models of pathophysiology.


46

However, this isn’t a trivial task due to the complex physical interactions involved in each acquisition as well as several systematic and random errors involved in each process. This paper discussed some relevant work of our group in a multi-level context.

REFERENCES [1] Th. Margaritis, K. Marias, D. Kafetzopoulos, “Improved Microarray

Spot Segmentation by Combining two Information Channels”, 2006 IEEE Engineering in Medicine and Biology Society (EMBS) Annual International Conference, IEEE.

[2] H. Stefanou, T. Margaritis, D. Kafetzopoulos, K. Marias, P. Tsakalides: Microarray Image Denoising Using a Two-Stage Multiresolution Technique. BIBM 2007: 383-389

[3] K. Marias, J. Ripoll, H. Meyer, V. Ntziachristos, S. Orphanoudakis, “Image Analysis for Assessing Molecular Activity Changes in Time-Dependent Geometries”, IEEE Transactions on Medical Imaging, Special issue on Molecular Imaging, Volume 24(7), July 2005.

[4] “3D Multi-Modal Registration for Assessing Molecular Activity Changes in Time-Dependent Geometries”, IEEE EMBS, 2008, Vancouver,BC, Canada


47

LATE ARRIVED PAPER


48

Parameter Estimation for Reaction-Diffusion Tumor Growth Models from Time Series of Images

Ender Konukoglu, Olivier Clatz, Hervé Delingette and Nicholas Ayache

Abstract—Reaction-diffusion models have been proposed and used widely in the literature to describe the evolution of brain tumors in the macroscale. These models have been successfully applied to medical images creating the integration of the clinical information into mathematical models. In this work we try to build the integration in the other direction through adapting these models to specific patient cases by estimating the personal parameters. Here, we propose a parameter estimation methodology for reaction-diffusion models using Magnetic Resonance images (MRI) of the same patient taken at successive time instances. We show preliminary results on synthetic data showing the feasibility of the method.1

I. INTRODUCTION

BRAIN tumors that start from glial cells, gliomas, form the major class of primary intracranial cancer. During the last 20 years there have been vast amount of research on mathematical descriptions of the growth dynamics of gliomas both in microscopic scale and in macroscopic scale [6], [7]. Specific attention has been given to one class of macroscopic models, the reaction-diffusion models, in the attempt to link tumor growth models to medical images [1], [9]. These models describe the evolution of the pathology via proliferation and infiltration of tumor cells. Their formulation consists reaction-diffusion type partial differential equations (PDEs) [4]. The system

Manuscript received July 4, 2008. Ender Konukoglu (corresponding author) is with the Asclepios

Project at INRIA Sophia Antipolis, 2004 Route des Lucioles 06902 France. (email: [email protected])

Olivier Clatz is with the Asclepios Project at INRIA Sophia Antipolis, 2004 Route des Lucioles 06902 France. (email: [email protected])

Hervé Delingette is with the Asclepios Project at INRIA Sophia Antipolis, 2004 Route des Lucioles 06902 France. (email: [email protected])

Nicholas Ayache is with the Asclepios Project at INRIA Sophia Antipolis, 2004 Route des Lucioles 06902 France. (email: [email protected]).

E.Konukoglou acknowledges support by the Health-e-Child and the CompuTumor projects.

( ) 0),1()( =⋅∇−+∇⋅∇=∂∂

Ω∂nuDuuuxDtu rρ (1)

⎩⎨⎧

∈∈

=matterwhitexDd

mattergrayxIdxD

waterw

g

,,

)( (2)

is an example of such reaction-diffusion models, where u is the tumor cell density, D is a local anisotropic diffusion tensor derived from the water diffusion Dwater (which can be obtained through DT-MRI), ρ is the proliferation rate, Ω is the brain domain and ∂Ω represents the boundaries of the brain [1].

In order to adapt the general model to patient specific cases the parameters should be estimated so that the model best fits the evolution of the tumor observed in the time series of images. The difficulty in this estimation is due to the sparsity of the information available. The reaction-diffusion models describe the temporal evolution of tumor cell densities however, in the images we only observe the evolution of the tumor delineation which is assumed to correspond to an iso-density contour [9]. Several works have used and proposed parameter estimation [3] however, the problem of estimating has not been analyzed thoroughly yet. Recently Swanson et al. in [8] have proposed such an analysis for the petri-dish case using analytical results obtained for the specific case.

In this work, we propose a parameter estimation method for reaction-diffusion tumor growth models using time series of medical images. The method is based on the evolution of the tumor delineation rather than tumor cell densities and in this respect it is consistent with the observations. We then evaluate the methodology using synthetic images created by reaction-diffusion model in order to understand our performance in retrieving the parameters of the model.

II. METHODS The parameter estimation methodology and the

estimated parameters depends naturally on the exact formulation of the underlying reaction-diffusion model. In this work we focus on the specific formulation proposed in [1] as given by Equations 1 and 2. However, due to the similarities of reaction-diffusion models the ideas we present here can be carried over to other formulations.

In order to solve the parameter estimation problem we need a formulation consistent with the images in which the evolution of the tumor delineation will be mathematically described instead of the evolution of tumor cell densities. The asymptotic properties of the reaction-diffusion equations under certain conditions (specifically in the infinite cylinder and with constant coefficients) permits us


49

to construct such a formulation. Using traveling wave solutions of Equation 1 and generalizing to lower times and curved surfaces we can obtain the following traveling time formulation for the tumor delineation [2], [11].

1''2

34=∇∇

⎭⎬⎫

⎩⎨⎧

∇∇∇

⋅∇−− TDT

TDTTD

TTρ

ρ (3)

Γ∈∀= xTxT 0)( (4) where T(x) is the function representing the time the tumor delineation will pass from the point x, Equation 4 represents the Dirichlet type boundary condition stating the initial tumor delineation we start evolving from and T0 is the time elapsed since the tumor has started diffusing until the acquisition of the first image. The value of T0 is not available in the images therefore it is regarded as another parameter of the model to be estimated for.

In the reaction-diffusion model given by Equations 1 and 2 we have three different parameters, dw, dg and ρ. In addition to these, the traveling time formulation has an extra parameter T0. In this work we try to optimize the parameters such that the evolution we simulate using the traveling time formulation best fits the real evolution we observe in the images. For an image set of N images taken at time intervals Δt1, ..., ΔtN−1 we can formulize the optimization criteria as

( )∑−

−+ΓΓ=1

1

20minminmin

2 )()ˆ,(N

ii TTTvdistC (5)

000),,,( )(,)(ˆ0

TTtTxTx iTddi gw=ΓΔ+==Γ ρ

(6)

where dist() is the symmetric distance between two surfaces, is the surface enclosing the tumor delineation

in the image i and is the tumor delineation simulated by the travelling time formulation for image i. The second part of C integrates the size of the first delineation in the optimization by inquiring whether it would have been possible to obtain at T0 using the traveling time formulation if we had started from the time the tumor had started diffusing. For this purpose, running the traveling time formulation backwards in time we get a minimum T value Tmin, the minimum point (or a set of points) xmin and the minimum speed vmin. In the ideal case T0 estimate should equal to Tmin. The minimization of C is a multidimensional optimization problem and for this we have chosen to use the algorithm proposed in [5] because it does not required explicit derivatives of the objective function.

iΓ

iΓ

0Γ

III. RESULTS We performed theoretical evaluation of the proposed

method using synthetic tumors. We have simulated 180 tumors with 60 different parameter sets at 3 different locations using the reaction-diffusion model. The different parameter sets were constructed using different combinations of dw = (0.025, 0.05, 0.1, 0.25, 0.5), dg = (0.005, 0.01, 0.025) and ρ = (0.009, 0.012, 0.018, 0.024).

In order to create synthetic images we assume medical images can only visualize the tumor in a voxel when the number of tumor cells is more than 40% of the maximum tumor cell capacity brain parenchyma can handle [10]. For each tumor we set the detection and the first image acquisition at the moment when the visible tumor reaches 1.5cm in diameter. After the first one we take 3 more images taken at intervals of 200 days. Using these 4 images and the time intervals between these images we estimate the parameters of the synthetic tumors (dw, dg, ρ, T0). In our experiments we have found out that the problem of estimating all the 4 parameters does not have a unique solution under the circumstances constrained by medical images. Different parameter sets produces very similar evolutions of the tumor delineation as given in Figure 1.

Fig.1. The Reaction-diffusion model produces very similar evolutions of the tumor delineation with different parameter sets. In the table we give two parameter sets. And in the image we show the evolution of the delineation of the synthetic tumors grown using these parameters. The contours are shown for days 200, 400, 600 and 800.

Instead of the 4 parameter estimation we focus on the case where the proliferation rate is fixed to its real value. In doing so we assume that the proliferation rate ρ can be found through microscopic analysis on biopsy results. We summarize our findings in Figure 1 where we plot the real parameters (big markers) and the corresponding estimated parameters (small markers). Although the parameter space is 3 dimensional we only show projections of this space onto the relevant 2 parameter space, (dw, dg). We observe that we are able to estimate dw successfully for all tumors. We notice the slight bias in the grey matter diffusion which we believe is due to the difference in numerical schemes we use to solve the reaction-diffusion equation and the traveling time formulation. Also we observe that estimation of dg is harder due to the fact that the gliomas mostly evolve in the white matter therefore there are more information available for estimating dw. We also tried to fix the proliferation rate to a different value than its actual value. In this case our experiments have shown that although the estimated value of the diffusion rate changes the product dρ remain constant. This shows that we are able to identify the speed of growth of tumors uniquely using medical images. In Figure 2(b) we show the dρ products found in two different experiments where ρ is fixed to 0.012 and 0.015. The products remain on the line

Red Green dw 0.273 0.153 dg 0.024 0.014 ρ 0.012 0.0185


50

y=x showing that the speed of growth of a tumor is uniquely identifiable through medical images.

(a)

(b)

Fig. 2 (a) shows the estimated diffusion rates when the proliferation rate is set to its real value. We see that the diffusion rates are estimated successfully using MR images. (b) shows that if we fix the proliferation rate to something else than its real value we find a different diffusion rate. However, the product dρ remains constant showing that we can uniquely identify the speed of growth of gliomas from medical images.

IV. CONCLUSIONS We have proposed a parameter estimation

methodology for the reaction-diffusion models using time series of medical images. The methodology formulates the evolution of tumor delineation thus staying consistent with the information available in the images. We have performed theoretical evaluation using synthetic images created by reaction-diffusion model and tried to retrieve the exact parameters using the parameter estimation method. We have shown that finding all parameters of the reaction-diffusion model does not have a unique solution based on the constraints given by the medical images. Further we have shown that if we fix the proliferation rate we can estimate for the diffusion rates correctly. Moreover, regardless of the value we fix ρ to we can uniquely identify the speed of growth of gliomas using medical images. This is especially important for clinical applications as this would give us a tool for quantification and characterization of the tumor. We demonstrated what could be estimated from the images in a very controlled study and the drawbacks of the method arising from sparsity of the information available. The future work is to enlarge the

theoretical analysis, apply this methodology to real images and perform experimental evaluation.

REFERENCES [1] O. Clatz, M. Sermesant, P. Bondiau, H. Delingette, S. Warfield, G.

Malandain, and N. Ayache, “Realistic simulation of the 3d growth of brain tumors in mr images coupling diffusion with biomechanical deformation,” IEEE T.M.I., vol. 24, no. 10, 2005.

[2] J. Keener and J. Sneyd, Mathematical physiology. Springer, 1998. [3] E. Konukoglu, O. Clatz, P. Bondiau, M. Sermesant, H. Delingette,

and N. Ayache, “Towards an identification of tumor growth parameters from time series of images,” in Lec. Notes Comp. Sci. 4791. MICCAI 2007, 2007, pp. 549–556.

[4] J. Murray, Mathematical Biology. Springer-Verlag, 2002. [5] M. Powell, “Uobyqa: unconstrained optimization by quadratic

approximation,” Math. Program. Ser. B, vol. 92, 2002. [6] S. Sanga, H. Frieboes, X. Zheng, R. Gatenby, E. Bearer, and V.

Cristini, “Predictive oncology: A review of multidisciplinary, multiscale in silico modeling linking phenotype, morphology and growth,” NeuroImage, vol. 37, pp. S120–S134, 2007.

[7] G. Stamatakos, V. Antipas, and N. Uzunoglu, “A spatiotemporal, patient individualized simulation model of solid tumor response to chemotherapy in vivo: The paradigm of glioblastoma multiforme treated by temozolomide,” IEEE Tran. Bio. Med. Eng., vol. 53, no. 8, pp. 1467–1477, 2006.

[8] K. Swanson, “Quantifying glioma cell growth and invasion in vitro,” Math. Comp. Model., vol. 47, pp. 638–648, 2008.

[9] K. Swanson, E. Alvord, and J. Murray, “Virtual brain tumours (gliomas) enhance the reality of medical imaging and highlight inadequacies of current therapy,” British Journal of Cancer, vol. 86, 2002.

[10] P. Tracqui, G. Cruywagen, D. Woodward, G. Bartoo, J. Murray, and E. Alvord, “A mathematical model of glioma growth: the effect of chemotherapy on spatio-temporal growth,” Cell Proliferation, vol. 28, no. 1, 1995.

[11] W. S. U. Ebert, “Front propagation into unstable states: universal algebraic convergence towards uniformly translating pulled fronts,” Physica D: Nonlinear Phenomena, vol. 146, 2000.

a proposed platform for intelligent identification of organs from medical images

Documents