Visualization of Risk Structures for Interactive Planning ofImage Guided Radiofrequency Ablation of Liver Tumors

Christian Rieder, Michael Schwier, Andreas Weihusen, Stephan Zidowitzand Heinz-Otto Peitgen

Fraunhofer MEVIS, Institute for Medical Image Computing,Universitatsallee 29, D-28359 Bremen, Germany

ABSTRACT

Image guided radiofrequency ablation (RFA) is becoming a standard procedure as a minimally invasive methodfor tumor treatment in the clinical routine. The visualization of pathological tissue and potential risk structureslike vessels or important organs gives essential support in image guided pre-interventional RFA planning. In thiswork our aim is to present novel visualization techniques for interactive RFA planning to support the physicianwith spatial information of pathological structures as well as the finding of trajectories without harming vitallyimportant tissue. Furthermore, we illustrate three-dimensional applicator models of different manufacturescombined with corresponding ablation areas in homogenous tissue, as specified by the manufacturers, to enhancethe estimated amount of cell destruction caused by ablation. The visualization techniques are embedded ina workflow oriented application, designed for the use in the clinical routine. To allow a high-quality volumerendering we integrated a visualization method using the fuzzy c-means algorithm. This method automaticallydefines a transfer function for volume visualization of vessels without the need of a segmentation mask. However,insufficient visualization results of the displayed vessels caused by low data quality can be improved using localvessel segmentation in the vicinity of the lesion. We also provide an interactive segmentation technique of livertumors for the volumetric measurement and for the visualization of pathological tissue combined with anatomicalstructures. In order to support coagulation estimation with respect to the heat-sink effect of the cooling bloodflow which decreases thermal ablation, a numerical simulation of the heat distribution is provided.

Keywords: Image-Guided Therapy, Visualization, Therapy Planning, Radiofrequency-Ablation

1. DESCRIPTION OF PURPOSE

In the past few years, image-guided ablation therapies using thermal energy has been developed as a minimalinvasive alternative for the treatment of liver tumors.1 Particularly, the radiofrequency ablation (RFA) has takena significant part in the clinical routine because of its common technical procedure, low complication rate andlow cost. The principles of RFA are that electrodes are placed percutaneously at the center of the tumor. Ahigh-frequency electric field is induced into the tumor which causes a local resistive heating of the tissue byionic agitation. This leads to a coagulative necrosis as a result of irreversible protein denaturation of the cells.To treat a large target zone, repositioning of a single applicator or inducing multiple applicators are often usedprocedures.

Treatment is only of value if complete destruction of tumor cells is guaranteed. As available methods for on-line monitoring of the thermal destruction are insufficient, the success of a tumor treatment depends considerablyon the pre-interventional planning of the RFA where choice of adequate applicator trajectory is important toprevent harming healthy structures as vessels, ribbons and lungs.2 Also choice of appropriate applicator typeand exact positioning of electrodes is essential to achieve complete destruction of cancer cells with respect toheat-sink effect of vessels located in the immediate vicinity.3

Further author information: (Send correspondence to)Christian RiederE-mail: chistian.rieder@mevis.fraunhofer.deTelephone: +49 421 218 8194

In our previous work4 we described a workflow oriented software platform for image guided RFA orientedto radiological interventions which are mainly planned using slicing in 2D viewers. In this work we developedadvanced 3D visualization methods to additionally support physicians such as surgeons and gastroenterologistswith a spatial view of the scene including all relevant anatomical and computational results for RFA planning.In comparison to medical applications where RFA planning depends on huge amount of accurate segmentationprocedures,5 our goal is to visualize risk structures in a way that physicians are able to perform fast RFA planningin the clinical routine, e.g. without the need of time-consuming data manipulation.

2. METHODS

In this section we give a description of the algorithms and visualization methods we integrated in our RFAplanning application. These techniques were designed under consideration of the time and usability constraintsof the clinical routine i.e. robustness, intuitive interaction and little time effort.

2.1 Tumor and Vessel Segmentation

To support the physician with knowledge of local anatomy such as tumors and vessels, we integrated twosegmentation algorithms. The tumor segmentation is a morphology based region growing algorithm describedby Moltz et al.6 This algorithm requires only a stroke across the lesion from the user and has proven to achieverobust segmentation results with little time effort. The vessel segmentation is an automatic algorithm whichworks in a local region of interest around an coagulation zone and starts by mouse-clicking into one vessel.7

2.2 Electrode Placement and Affected Area Visualization in 2D

The radiofrequency applicators of different manufacturers are visualized by corresponding virtual models, whichcan be placed and moved within the scene in 2D as well as in 3D. The location of an applicator model isdetermined by a pair of spatial coordinates which define the position of the applicator electrode and the positionof the applicator handle. Those two spatial coordinates can separately or collective be positioned by the userto move the applicator within the scene. Also, the RF applicator can intuitively be moved in shaft direction tosimulate back-tracking ablation known from clinical routine.

Supplementary to the familiar 2D representation with its axial, sagittal and coronal views, we integratedtwo orthogonal multi-planar reformatted (MPR) views to allow fast exploration of tissue in the vicinity of theapplicator electrode. The first MPR view is oriented along the applicator axis and can be rotated 360 degreesaround the applicator. The second MPR view shows the scene plane normal to the applicator along shaftdirection and can be sliced from shaft to target. The advantage of the MPR views is that risk structures in thevicinity of the electrode can intuitively explored even if the applicator is not aligned to the main slicing axes.

The affected area of an applicator is the region of complete cell destruction in homogeneous tissue announcedby the manufacturer. This area should be large enough to enclose the tumor as well as a safety margin of 1 cm atall sides. On the other hand, the coagulation zone should be small enough to avoid ablation of risk structures.8

Thus, we create a three-dimensional model from the electrode parameters,9 which is an ellipsoid geometry, toallow a visual estimation whether the coagulation zone of the RF applicator is large enough for complete tumorablation. However, one should be aware of the heat sink effect caused by the cooling blood flow in vessels.

2.3 Three-dimensional Volume Rendering

Besides the 2D representations we described in Section 2.2, an additional three-dimensional volume renderingcan be chosen for a spatial view of the scene10 and all computational results which helps physicians like surgeonsor gastroenterologists planning RFA more intuitive than by slicing in 2D.

2.3.1 Anatomical Volume Rendering

The anatomy is visualized as three-dimensional direct volume rendering (DVR) using the possibilities of moderngraphics hardware.11 The goal of the anatomical visualization is to allow a fast recognition of potential riskstructures like important vasculature of the liver, which has to be preserved from harming, as well as importantlandmarks for optimal placement of RF applicators. Because of the limited available planning time in the clinicalroutine, often few minutes before intervention, accurate but time consuming segmentation procedures can notbe taken into account. Nevertheless, the difficulty in displaying anatomical structures such as vessels or lungs byvolume rendering is to set up an appropriate transfer function using simple interaction techniques like windowing,known from medical workstations.

To overcome this issue and to allow a fast exploration of the data set, we automatically calculate an appro-priate transfer function. Due to its robustness to noise and varying intensity distributions, we utilize the fuzzyc-means (FCM) clustering to determine adequate grey value thresholds to set up a transfer function for thevolume rendering that emphasizes liver vessel structures.12 To save computation time we apply the clusteringin a local region of interest (ROI) around the selected tumor. The input image for our method is the ROI ofa contrast enhanced venous phase CT image of the liver. The algorithm groups the voxels into four clusters.The clusters define the appropriate gray value thresholds to set up the transfer function we use for the volumerendering of the complete data set.

Thus, our method allows to explore liver vasculature without the need of segmented vessels in a visual qualitywhich is close to state-of-the-art visualization techniques for segmented vasculature as convolution surfaces.13

Compared with traditional adjustment of rendering values (see Figure 1 a) based on manual windowing, the visualquality of our automatic method is as good as quality of the manual technique, often on top of it – particularlywith regard to the visual separation of different structures – without the need of substantial interaction (seeFigure 1 b).

(a) (b)

Figure 1. In (a) the contrast of the volume rendering is manually set by the user; (b) shows the rendering with automaticallydetermined transfer function. Pulmonary structures and other soft tissue is visible, too.

2.3.2 Smooth Volume Rendering of Segmentation Results

In the two-dimensional visualization, we blend labeled segmentation results within the anatomical context. Wealso integrated the tumor mask into the three-dimensional volume rendering14 to allow exploration of the spatialrelations around lesions. For high-quality visualization results, each tumor mask is smoothed using a diffusionfilter. To enhance the spatial perception of tumors, volume shading is enabled. For that, gradients, whichare needed for volume shading computation, have to be computed using a Sobel filter as a pre-processing step.Figure 2 shows the visualization of un-smoothed, un-shaded and smoothed-shaded segmentation mask.

(a) (b) (c)

Figure 2. In (a) the tumor is rendered as a simple mask. In (b) the tumor is visualized using volume shading. Image (c)shows the shaded volume rendering of the smoothed segmentation mask.

2.3.3 Volume Visualization of Affected Areas

To visually estimate the amount of cell destruction caused by ablation, we integrated the rendering of affectedareas, specified as ellipsoids by the manufacturers, in our volume visualization method. We change the coloringof lesions and healthy risk structures such as vessels, located in the affected area in order to allow the recognitionof the area of destruction. Also, silhouettes are drawn at the areas surface to enhance the boundary of tissuelocated inside and outside of the affected area (see Figure 3).

Figure 3. An RF applicator is positioned into a tumor. The color of the tumor is yellow if it is located inside the affectedarea and blue outside. Silhouettes are drawn at the boundary to emphasize tumor regions which will not be burned byablation.

Technically we compute in the volume rendering shader for every voxel if the voxel position xxx is inside oroutside of the ellipsoid (see Figure 4). The result of following equation is negative if xxx is inside the ellipsoid andpositive outside:

f(xxx) = |xxx − xxx0 +rmin − rmax

rmaxvvv(vvv · (xxx − xxx0))|2 − r2

min

where xxx0 denotes the middle point, rmin is the minimal radius of the ellipsoid and rmax is the maximal radiusalong orientation vvv.

In clinical routine, multiple applicator placement is becoming a standard procedure for ablation of largelesions. To support multiple affected areas we compute the union of all visible affected areas, which means thatevery voxel has to be tested if its location is inside the ellipsoids. For that purpose we developed a dynamicshader framework for direct volume rendering to enable manipulation of shader code during application runtime.Consequently, applicators as well was affected areas can be interactively created, removed or manipulated. Theindividual applicator parameters are immediately forwarded into the shader and the affected area ellipsoids arecomputed.

rmin

rmaxvxoRF applicator

a�ected area

Figure 4. An illustrative description of an RF applicator and corresponding affected area.

2.4 Numerical Simulation of Affected Areas

Due to the patients individual anatomy where tumors can be located in the vicinity of vessels, the cooling of theblood flow should be considered in RFA planning.15 As a consequence, an ellipsoid coagulation zone of constantdiameter can not be expected.16 Since the manufacturers description of the affected area does not incorporateheat sink effects, we integrated a numerical simulation. The computation of the heat distribution is based onFEM17 and incorporates the characteristics of the chosen RF applicator type as well as the neighboring vascularstructure.

We visualize the resulting heat distribution in 2D using a color coded coagulation temperature map (seeFigure 5 a). Tissue, which is located in the zone of immediate cell destroyed is colored red. Coagulation zonesof incomplete destruction are colored orange and yellow. In 3D, we visualize the computed coagulation maskusing geometric surfaces. Those surfaces (see Figure 5 b) are integrated into the volume rendering. Thus, thephysician is able to interpret the simulation result which could lead to repositioning of RF applicators becauseof heat sink effects, which influence the coagulation zone.

(a) (b)

Figure 5. In (a) the simulation result is visualized using a color coded coagulation mask; image (b) shows the correspondingthree-dimensional rendering with the computed coagulation mask using WEM surfaces.

2.5 Object Navigation

To allow intuitive handling of multiple objects as applicators, affected areas, annotations, measurements, lesionsand segmented vessels we integrated a data navigation tool in our application. For clear representation, objects

can be selected, hidden and deleted using the navigation tool. The 2D viewers show the corresponding slice ofthe selected object to allow the physician a fast exploration of the objects location. By this, the physician isable to optimize the placement of multiple applicators in a intuitive manner. Figure 9 shows the GUI of theapplication. A single Lesion is selected in the Object Navigator and displayed orange in the 2D viewers. The 3Drendering shows the applicator with affected area around the green tumor.

3. RESULTSClinical partners use the application for scientific projects and clinical evaluations. They regard the methodsas helpful and intuitive for RFA planning. The volume rendering method with automatic determined transferfunction enables a 3D visualization of multiple anatomical structures like bones, vessels and lungs without theneed of expansive segmentation tasks and thus enables an immediate spatial view of the planning situation.Particularly, pulmonary structures can be distinguished which is important to prevent harming these structuresduring electrode placement.

To assess the value of our anatomical volume rendering technique, we compared the visual quality of severalvolume rendered vessel systems with corresponding convolution surface visualizations of manually segmentedvasculature. We define a good visual quality as a visualization of all vessels greater than two millimeter diameter.To our knowledge, vessels smaller than two millimeter are no substantial risk structures for radiofrequencyablation. The manual segmentation, which is part of an extensive vasculature analysis for oncological surgeryplanning, was done by medical technical assistants with high amount of expertise and is thus our ground-truth.The time effort for the manual segmentation process ranges from 10 minutes for image data with good qualityto 40 minutes for image data with bad quality or strong anatomical deformations. Depending on the CT imagequality as noise, contrast and resolution, with our method it is possible to achieve a visual quality which isclose to the convolution surfaces visualization for segmented vasculature (compare Figure 6). In summary, thiscomparison shows that the rate of accuracy depends on the image data quality whereas bad image quality causeshigher segmentation process time. Nonetheless, vessels with minor visual quality supports the physician withimportant anatomical information for radiofrequency planning if accurate segmentation masks of vessels are notavailable.

(a) (b)

Figure 6. Image (a) shows the combined visualization of bone structures using volume rendering and of vessels usingconvolution surfaces. In (b) the same data set is visualized with our automatic rendering technique.

The visual quality of our automatic transfer function calculation depends mainly on two parameters: thecontrast between vessel structures and liver parenchyma as well as the noise in the CT image. On the one hand,

if the contrast is very high, a good visualization is achievable with high noise in the image. At the other hand, ifthe noise in the image is negligible low a good visual quality is also possible with low contrast. The explanationof that effect is that a small intensity difference between liver parenchyma and vessel structures with respectto the image noise level results in intensity clustering of voxels without reasonable object classification. If thecontrast between liver and vessels is high, i.e. substantial difference of the mean intensities, a good visual qualityis possible in defiance of strong noise, i.e. a huge standard deviation. Figure 7 shows the mean and standarddeviation values for liver and vessels taken in a ROI image (see Figure 8 a) where gaussian noise was added. Thevertical axes shows the Hounsfield Units (HU), the horizontal axes the σ values of the gaussian noise (with meanvalue = 1). At a gaussian noise level of σ = 17 the standard deviation of liver parenchyma and vessel structuresbegin to overlap and at a noise level of σ = 44 the standard deviations cross the mean values. Whereas thevisualization of the ROI at σ = 17 is minor different (see Figure 8 b), the visual quality at σ = 44 is significantlylower because visualization artifacts appear which are mainly undesired liver parenchyma instead of vessels (seeFigure 8 c).

mean vessel

mean liver

std. deviation

std. deviation

24 HU

108 HU

192 HU

275 HU

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 950 100

Figure 7. The mean and standard deviation of liver parenchyma as well as vessel structures subject to the standarddeviation σ of the added gaussian noise. At a noise level of σ = 17 the standard deviation of liver parenchyma and vesselstructures begin to overlap (see Figure 8 b) and at σ = 44 the standard deviations cross the mean values (see Figure 8 c).

(a) (b) (c)

Figure 8. Image (a) shows a ROI with automatic determined transfer function. In (b) a gaussian noise (mean = 1, σ = 17)was applied onto the image, the image quality begins to decrease. Image (c) shows the resulting rendering with gaussiannoise of mean = 1 and σ = 44 where the calculated transfer function overlaps vessel structures as well as liver parenchyma.

Additionally we performed user experiments for a feasibility study to evaluate how our visualization methods

could influence and improve the planning of RF ablations. Using manual windowing, medical experts had tojustify the volume rendering to achieve a clear view to vascular structures as well as bones. We measuredthe required time effort, calculated the mean value and compared it with the mean computation time of ourautomatic method. On the one hand, we could prove our assumption that manual windowing of the transferfunction with similar visualization results is more difficult and time consuming. On the other hand, in 22 of 31test cases, we could not observe significant differences to our automatic transfer function determination focusedon vascular structures. In 9 cases, the users were able to visualize smaller vessels. Because of that, we also addedmanual windowing to adjust the calculated transfer function afterwards, if needed.

Furthermore, most of the respondents judged that the proposed technique for the affected area visualizationprovide an immediate and spatial recognition of untreated tumor tissue. Compared with the 2D visualization,needed repositioning of the electrodes was earlier recognized. According to the 3D visualization, a requiredadaption of the access path as well as the rapid detection of risk structures along the trajectory is intuitive toachieve.

Figure 9. This screen shot shows the GUI of our RFA planning software with enabled volume rendering. The RF applicatoris positioned close to the selected lesion but the affected area surrounds not completely the lesion.

4. CONCLUSIONS

In this work we described visualization methods for a workflow optimized application, which assists physiciansin pre-interventional RFA planning. We integrated segmentation algorithms to support knowledge of localanatomy. To support electrode placement we developed different RF applicator models with correspondingaffected areas. Among the 2D viewers, the three-dimensional volume rendering provides an intuitive placementof RF applicators to ensure complete destruction of tumor cells as well as preserving risk structures consideringthe patients anatomy.

Future work will focus on the transfer of RFA planning data into the intervention and the support of ultrasonicnavigated devices. Also we make plans of an intervention monitoring application.

ACKNOWLEDGMENTS

This work was funded by the Federal Ministry of Education and Research (SOMIT-FUSION project FKZ01IBE03C). We would like to thank the people at Fraunhofer MEVIS as well as our clinical partners for theircontribution to our work.

REFERENCES[1] Garrean, S., Hering, J., Saied, A., Helton, W., and Espat, N., “Radiofrequency ablation of primary and

metastatic liver tumors: a critical review of the literature.,” Am J Surg 195, 508–520 (Jan 2008).[2] GANGI, A., “Trajectory optimization for the planning of percutaneous radiofrequency ablation of hepatic

tumors,” Comp. Aided Surgery 12(2), 82–90 (2007).[3] Villard, C., Soler, L., and Gangi, A., “Radiofrequency ablation of hepatic tumors: simulation, planning,

and contribution of virtual reality and haptics,” Comp. Methods in Biomechanics & Biomedical Eng. 8(4),215–227 (2005).

[4] Weihusen, A., Ritter, F., Kroger, T., and Preusser, T., “Workflow oriented software support for imageguided radiofrequency ablation of focal liver malignancies,” Proceedings of SPIE 6509, 650919 (Jan 2007).

[5] McCreedy, E., Cheng, R., Hemler, P., Viswanathan, A., Wood, B., and McAuliffe, M., “Radio frequencyablation registration, segmentation, and fusion tool,” Information Technology in Biomedicine, IEEE Trans-actions on 10, 490 – 496 (Jul 2006).

[6] Moltz, J. H., Bornemann, L., Kuhnigk, J.-M., Dicken, V., Peitgen, E., Meier, S., Bolte, H., Fabel, M.,Bauknecht, H.-C., Hittinger, M., Kießling, A., Pusken, M., and Peitgen, H.-O., “Advanced segmentationtechniques for lung nodules, liver metastases, and enlarged lymph nodes in CT scans,” IEEE J. Sel. TopicsSignal Process. (2009). to be published.

[7] Zidowitz, S., Drexl, H., Kroger, T., Preusser, T., Ritter, F., Weihusen, A., and Peitgen, H.-O., “BaysianVessel Extraction for Planning of RF-Ablation,” Bildverarbeitung fur die Medizin , 187–191 (2007).

[8] Wood, T., Rose, D., Chung, M., Allegra, D., Foshag, L., and Bilchik, A., “Radiofrequency ablation of 231unresectable hepatic tumors: Indications, limitations, and complications,” Annals of Surgical Oncology 7(8),593–600 (2000).

[9] Weihusen, A., Schumann, C., Bieberstein, J., Peitgen, H.-O., Niethammer, M., and Haras, G., “Softwareassisted treatment planning of image guided rf ablations,” in Proceedings of CURAC 3, 69–70 (2008).

[10] Antoch, G., Kuehl, H., Vogt, F. M., Debatin, J. F., and Stattaus, J., “Value of ct volume imaging foroptimal placement of radiofrequency ablation probes in liver lesions,” JVIR 13, 1155–61 (Nov 2002).

[11] Westermann, R. and Ertl, T., “Efficiently Using Graphics Hardware in Volume Rendering Applications,”Computer Graphics (SIGGRAPH ’98) 32(4), 169–179 (1998).

[12] Schwier, M., Dicken, V., and Peitgen, H.-O., “3D Visualization of Vascular Structures around Liver Tumorsusing Fuzzy Clustering,” in Proceedings of CARS 3, 403–404 (2008).

[13] Oeltze, S. and Preim, B., “Visualization of Anatomic Tree Structures with Convolution Surfaces.,” in Proc.Joint IEEE/EG Symposium on Visualization. Eurographics Association , 311–320 (2004).

[14] Hadwiger, M., Berger, C., and Hauser, H., “High-quality two-level volume rendering of segmented data setson consumer graphics hardware,” in Proceedings of IEEE Visualization , 301–308 (Oct. 2003).

[15] Mulier, S., Ni, Y., Frich, L., Burdio, F., Denys, A. L., Wispelaere, J.-F. D., Dupas, B., Habib, N., Hoey,M., Jansen, M. C., Lacrosse, M., Leveillee, R., Miao, Y., Mulier, P., Mutter, D., Ng, K. K., Santambrogio,R., Stippel, D., Tamaki, K., van Gulik, T. M., Marchal, G., and Michel, L., “Experimental and clinicalradiofrequency ablation: proposal for standardized description of coagulation size and geometry,” Ann SurgOncol 14, 1381–96 (Apr 2007).

[16] Lobik, L., Leveillee, R. J., and Hoey, M. F., “Geometry and temperature distribution during radiofrequencytissue ablation: an experimental ex vivo model,” J Endourol 19, 242–7 (Mar 2005).

[17] Kroeger, T., Altrogge, I., Preusser, T., Pereira, P., Schmidt, D., Weihusen, A., and Peitgen, H., “Numericalsimulation of radio frequency ablation with state dependent material parameters in three space dimensions,”Proceedings of MICCAI 4191, 380–388 (2006).