Page - 103 - in Joint Austrian Computer Vision and Robotics Workshop 2020

low-resolution data blood-lumen mask annotated dataset centerline information blood-lumen and stent-graft patcheshigh-resolution data high-resolution segmentation train train merge extract patch locations M1 M2 Figure 2. Outline of our method: The top branch shows the centerline extraction step (modelM1) and the bottom branch thepatchwise segmentationstep (modelM2). challenges arise due to considerable imaging arti- facts caused by the stent-graft wire frame and the distinctboundariesbetweenblood lumenand throm- bus. While there are a number of publications on the segmentation of the abdominal aorta, very few have focusedonstent segmentation. Kleinetal. [12]used agraph-basedmethodtocreateageometricmodelof thestent-graft, disregarding theaortaentirely. To the bestofourknowledge, there isnot a singleapproach segmenting both structures simultaneously. For the segmentation of the abdominal aorta, traditional ap- proaches includegraph-basedmethods [6,4,23]and deformable-models [13, 14] which require user in- teraction to varying degrees and have predominantly been evaluated on pre-operative scans. A common problem with graph- and deformable-model-based approaches is the introduction of many parameters optimized for the respective dataset, limiting the ro- bustness and applicability of the methods in clinical settings [17]. With the introduction of the convolu- tionalneuralnetwork(CNN)thefieldofmedical im- age analysis changed significantly. Today the U-Net [21] and its 3D equivalent [8] are the most widely models used for medical image segmentation. Both models have been applied to the task of the abdom- inal aorta segmentation, Zheng et al. [26] reporting a Dice similarity coefficient (DSC) of 0.82 for the aneurysm thrombus and Li et al. [16] reporting a DSC of0.92 for the aorta blood lumen. For the seg- mentation of blood lumen and stent graft wire frame we will therefore likewise rely on the (3D) U-Net architecture. The distinguishing challenge to other segmentation tasks is inourcase thefinestructureof the stent-graft, with a diameter as small as 0.4mm [24],whichrequiresanexceptionallyhighresolution for accurate reconstruction, pushing the limitations ofmodernhardware. 3.Dataset Our dataset consists of 76 abdominal CTA scans of 36 patients treated with EVAR that we received fromtheKeplerUniversityHospitalLinz. Eachscan consists of 155 to 873 axial slices with 512× 512 voxels. There are large differences in the resolu- tion with a minimum voxel spacing ranging from 0.404mm frontal/sagittal and 0.8mm longitudinal (a) (b) Figure 3. Examples of two ground truth segmentations: Medtronic Endurant (a) and Anaconda (b). In total the dataset contains5 different types of stent-grafts. 103