university of toronto - radiomics for oncology - 2017
TRANSCRIPT
Radiomics for Oncology
Andre DekkerDepartment of Radiation Oncology (MAASTRO)GROW - Maastricht University Medical Centre +Maastricht, The Netherlands
SLIDES AVAILABLE ON SLIDESHARE (slideshare.net/AndreDekker)
2
Disclosures• Research collaborations incl. funding and speaker honoraria
– Varian (VATE, SAGE, ROO, chinaCAT, euroCAT), Siemens (euroCAT), Sohard (SeDI, CloudAtlas), Mirada Medical (CloudAtlas), Philips (EURECA, TraIT, SWIFT-RT, BIONIC), Xerox (EURECA), De Praktijkindex (DLRA), ptTheragnostic (DART, Strategy), CZ (My Best Treatment), OncoRadiomics
• Public research funding– Radiomics (USA-NIH/U01CA143062), euroCAT(EU-Interreg), duCAT&Strategy (NL-
STW), EURECA (EU-FP7), SeDI & CloudAtlas & DART (EU-EUROSTARS), TraIT (NL-CTMM), DLRA (NL-NVRO), BIONIC (NWO)
• Spin-offs and commercial ventures– MAASTRO Innovations B.V. (CSO)– Various patents on medical machine learning
3
LectureLearning objectives, after this lecture you should be able to • Formulate what the rationale of Radiomics is and how it might
contribute to personalized medicine• Name the major workflow steps to use Radiomics to get from image
data to decision support• Appraise papers that describe Radiomics research incl. how the
authors handle the many Radiomics challenges• Name a few future directions for Radiomics
Part 1: Rationale (Predictions, Big Data, Radiomics) – 15 minsPart 2: Radiomics workflow & challenges – 25 minsPart 3: New directions in Radiomics – 15 mins
Part 1 - Rationale
5
Can we predict a tulip’s color by looking at the bulb?
http://www.amystewart.com
6
Predicting the color of a tulip - AUC
1.00AUC
0.72
0.50
7
Predicting the survival of NSCLC patients
AUC1.00
AUC0.50
AUC0.72
8
Prediction by MDs?
NSCLC2 year survival30 patients8 MDsRetrospectiveAUC: 0.57
NSCLC2 year survival158 patients5 MDsProspectiveAUC: 0.56
Oberije et al. Kruger et al. 1999
Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence leads to inflated self-assessments. J Pers Soc Psych
9
The problem of Big Data – The doctor is drowning
• Explosion of data• Explosion of decisions• Explosion of
‘evidence’*• 3 % in trials, bias• Sharp knife
*2010: 1574 & 1354 articles on lung cancer & radiotherapy = 7.5 per dayHalf-life of knowledge estimated at 7 years (in young students) J Clin Oncol 2010;28:4268
JMI 2012 Friedman, RigbyBMJ Clinical Evidence
We cannot predict outcomes of individual treatments
10
The potential of Big Data - Rapid Learning Health Care
In [..] rapid-learning [..] data routinely generated through patient care and clinical research feed into an ever-growing [..] set of coordinated databases. J Clin Oncol 2010;28:4268
[..] rapid learning [..] where we can learn from each patient to guide practice, is [..] crucial to guide rational health policy and to contain costs [..].Lancet Oncol 2011;12:933
Examples: DLRA, NROR, CAT (www.eurocat.info) ASCO’s CancerLinQ
11
Cancer Data?
Oncology2005-2015140M patients0.1-10GB per patient14-1400PB80% unstructured100k hospitals
12
Images are not picture, they are data
Gillies et al., Radiology 2016;278(2). Larue, et al., Br J Radiol 2017
13
Nature selects for phenotype
Lambin et al., Eur J Cancer. 2012 Mar;48(4):441-6
14
Radiomics vs Radiongenomics• Radiomics
– High throughput quantitative analysis of standard of care imaging to characterize tumours and normal tissues to improve cancer diagnosis, prognosis, prediction and response to therapy.
• Radiogenomics– The link between Radiomics and Genomics (i.e. how the imaging phenotype
and genotype are related)– The interaction between Radiotherapy and Genomics (genetic risk factors for
radiation toxicities?)
15
Animation• https://www.youtube.com/watch?v=Vf0F7q8vaS4
Part 2 - Radiomics workflow & challenges
17
Radiomics WorkflowLambin, Walsh et al., Nat Rev Clin Oncol (in-press)
Larue, et al., Br J Radiol 2017
18
Guide
19
Feature Extraction – Imaging Protocols
Oliver et al. , Translational Oncology (2015) 8, 524–534
20
Guide
21
Feature Extraction – Robust Segmentation
Parmar et al., PLoS One. 2014; 9(7): e102107.
22
Feature Extraction – Robust Segmentation
Parmar et al., PLoS One. 2014; 9(7): e102107.
Approaches1. Perform semi-automatic segmentation2. Remove features which are too sensitive to the exact
segmentation
Larue, et al., Br J Radiol 2017
23
Key points until now• They key to Radiomics is not to be perfect but to be
consistent and adhere to (other people’s) standards• Radiomics on the state of the art imaging does not
makes sense, focus on clinical standard of care
• Radiomics until now works (much) better in Radiotherapy than in Radiology
24
Guide
25
Feature Extraction - Software
Non-texture-based features: Histogram, GeometryTexture-based features: GLCM, GLRLM
Sample capacity: 31 51 33
Correlation Coefficients Distribution
correlation coefficient range
Fudan University Cancer Hospital (unpublished)
26
Feature Extraction – Phantom / Ontology
27
Test-retest feature stability• Rectal cancer clinical test-retest data from Fudan (Shanghai)
– n=40, different scanners, tube currents, recon parameters– Time between scans 5-19 days (median 8)
• Lung cancer coffee-break test-retest from NCI (RIDER)– N=35, same scanner, same recon– Time between scans 10 minutes
• Hypotheses – Similar features are reproducible in the clinical scenario as in the “coffee-
break” scenario – Features found to be robust in one tumor site are also robust in another
tumor site.• Compare ICC between Lung (RIDER, coffee-break) & Rectum (Fudan,
clinical)
28
Rectum clinical vs. Lung coffee-break
29
Guide
30
Combining with clinical
Aerts, JAMA Oncol 2016
31
Dimensionality reduction - Archetypes
Gillies et al., Radiology 2016;278(2).
219 features in 235 patients
Aerts et al., Nature Communications 5, 4006
32
Guide
33
Our modelling approach
34
How much data do you need?• Rule of thumb. Min. 10 events per input feature
• 200 NSCLC patients• 25% survival at two years• 50 events
• 10 input features• Less features is generally better Source: vitalflux.com (2017)
35
Source: Jason Brownlee (2013)
Machine Learning
36
Considerations for machine learning• Discrimination (AUC)• Calibration (Brier)
• Interpretability (black box vs. transparent)
• Can it handle low data quality (of training and validation)?
37
Choose alreadySimple and quick, but need complete data• Logistic regression• Support Vector Machines
Intuitive and can handle missing data• Bayesian Networks
Review pending
38
TRIPOD
https://www.tripod-statement.org/
39
So, Radiomics needs a lot of training data….
Aerts et al., Nature Communications 5, 4006
40
…. and a lot of validation data
Aerts et al., Nature Communications 5, 4006
Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis
41
Radiomics – End result
Part 3: New directions
43
Radiomics – Preclinical
44
Radiomics – PET
45
Radiomics - MRI• Rectal cancer - Chemoradiation• Pathological response• Training n=173, Validation n=25• AUC 0.79 (validation)
1) MR GTV delineation
2) GTV ROI extraction
3) LoG filter application according different s
0.3 0.5 1.0 2.0 3.0 4.0
4) Data analysis
|||cT
234
Points
| | |cN
||||||||||||||SKE0485
−0.6−0.4−0.200.20.40.6
| | | | | | | | | |ENT0344
1.6 1.8 2 2.2 2.4
| | | | | | | || | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |Total Points
320 330 340 350 360 370 380 390
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
TRG1Probability
0 1 2
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190
46
Radiomics outside of oncology
47
Radiomics – Delta Radiomics
48
CBCT and CT interchangeable? • 132 patients with stage I-IV non-small cell lung cancer
(NSCLC) treated with curative intent• Total of 543 radiomic features
49
Kaplan-Meier curves Correction with slope of linear regression
p = 0.0054 (pCT) and p = 0.00099 (CBCT-FX1)
50
Radiogenomics – Virtual Biopsy
Wu et al., Front Oncol 2016
51
Distributed Radiomics
52
Rapid Learning Health Care
53
Conclusion• We are still in the very early phase• A lot of underpowered, exploratory
papers out there • A lot of dials to control (medical
physics needs to get involved)• Prospective validation as a
decision support system is needed• We all can help by collection of
highly standardized images in our clinics
• But the promise is HUGE
1 2 3 4 50
20406080
100120140160
Pubmed RadiomicsRadiomics
54
Acknowledgements• MAASTRO Clinic, Maastricht, The Netherlands
– Philippe Lambin, Ralph Leijenaar,….• Moffitt Cancer Center, Tampa, FL, USA
– Bob Gillies, Bob Gatenby,…• Dana-Farber Cancer Institute, Brigham and Women’s Hospital, Harvard Medical
School, Boston– Hugo Aerts, Emmanuel Rios Velazquez, …
• Radboud University Medical Center, Nijmegen, The Netherlands• VU University Medical Center, Amsterdam, The Netherlands
More info on: www.radiomics.org
Thank you for your attention
Andre DekkerDepartment of Radiation Oncology (MAASTRO)GROW - Maastricht University Medical Centre +Maastricht, The Netherlands