Page - 104 - in Joint Austrian Computer Vision and Robotics Workshop 2020

to 0.977mm frontal/sagittal and 3mm longitudi- nal. Weused theActive-Contour/Snake-Modeof the Software ITK-Snap [25] to semi-automatically cre- ate the initial ground truth segmentation of the aor- tic blood lumen from below the heart to the second iliac bifurcation. The stent-graft segmentation was further added by applying a threshold to a region of interest around the blood lumen. The segmentation was then revised using the Paintbrush Mode of ITK- Snap. Figure 3 shows examples of the final ground truth segmentations used for training and validation. The dataset was split into 5-folds using a grouping criterion on the patient number to avoid having mul- tiple scans of the same patients assigned to different folds. 4.Method We use a two step approach in our segmentation method that is outlined in Figure 2. First we extract theaorticcenterlinesfromacoarseblood-lumenseg- mentation and subsequently use them to extract high resolution patches along the entire span of the aorta. In the second step we segment the blood lumen and the stent-graft wire frame for each patch and merge theresults toafinalsegmentation. Theentiresetupis tuned to work with an NVIDIA GeForce 1080 Ti (11 GBRAM). 4.1.CenterlineExtraction We use a full-image segmentation modelM1 to create a low resolution segmentation of the aortic blood-lumen. We resample the scans and ground truth to a voxel spacing of 1mm frontal/sagittal and 3mm longitudinal and crop them to a large re- gion of interest of 192×192 voxels and 128 slices (i.e., aphysicalextentof192mm frontal/sagittaland 384mm longitudinal). The largest connected region of blood lumen voxels in the resulting segmentation is then selected and skeletonized using homotopic thinning[15]. Using thepython librarySkan [20]we extract thecenterlinegraphfromtheskeletonizedim- ages, which is essential for the patchwise segmenta- tionstep. Figure4outlinestheintermediateresultsof the centerlineextractionstepandanexamplepatch. 4.2.PatchwiseSegmentation A patchwise segmentation modelM2 is used to segment the aortic blood lumen and the stent-graft wire frame in high resolution patches. We resam- ple the scans and ground truth to a voxel spacing (a) (b) (c) (d) Figure4.Extractingpatchesalongthesegmentedaorta lu- men: (a) coarse low resolution segmentation of the aor- tic blood lumen, (b) centerlines approximated via skele- tonization of the blood-lumen, (c) ground-truth segmen- tation for a patch sampled along the centerlines, (d) axial slice of the same patch (scan overlayed with the ground- truth segmentation). of 0.35mm frontal/sagittal and 0.75mm longitu- dinal before extracting patches of size 160× 160 voxels and 128 slices. Choosing a high resolution (i.e., small voxel spacing) significantly reduces the amount of distortion introduced by resampling, es- peciallyconsidering thevaryingvoxel spacing in the dataset. However, this results in a rather small phys- ical extent of56mm frontal/sagittal and96mm lon- gitudinal thatweseektouseasefficientlyaspossible by centering the patches at equally distributed loca- tions along the entire centerline graph. In our exper- iments 100 patches per scan proved more than suf- ficient to cover the aorta and introduce a significant overlapbetweenthepatches. Thepatchesaremerged into a final segmentation using a Gaussian-weighted kernel that attenuatesvoxels at thepatchboundaries, where the segmentation results are less reliable. 5. Implementation In this section we discuss the implementation de- tails, i.e., the operations used for preprocessing the dataset, themodelarchitectureandconfigurationand the training routine. 5.1.Preprocessing Preprocessing of the dataset consists, in addition to the resampling mentioned in Section 4, of clip- ping and normalization. As the voxel spacing varies 104