Seite - 106 - in Joint Austrian Computer Vision and Robotics Workshop 2020

the experiment on the first fold and compared the results to those of our method for the same fold: Although the blood-lumen segmentation is with a DSC of 0.963 slightly better than our method, yielding 0.951 for the same fold, the more complex stent-graft segmentationdoesnotcomparewell,with aDSCof0.785versusour method’s scoreof0.852. 7.Discussion The strength of our method is the centerline- guided segmentation method using the aortic center- lines to optimize the patch locations during training andinference. Whileourmethodyieldsbetter results than a comparable model using a traditional sam- pling setup, it also reduces the computational cost significantly. For traditional patching using a grid of overlapping tiles (when not allowing the patches to contain regions outside the image) the number of patchescalculatesas follows: npatches= nd∏ d=1 d |I|d−θd|P|d−θd e (1) wherend is the number of dimensions, |I|d the size of an image, |P|d the size of a patch and θd the overlap in dimension d. For our setup and a cho- sen overlap of 32 voxels this results in 343 patches on average per scan (|I| = (990,990,678), |P| = (160,160,128), θd = (32,32,32)). Increasing this overlap to improve the model’s performance quickly raises this number, e.g., an overlap of half the patch size (as used by Isensee et al. [10]) would result in 1440 patches on average per scan (θd = (80,80,64)). Themajorityof thepatchesare irrelevant for the result, as they do not intersect with the target structure. By using the aorta centerline in- formation, our method is able to greatly reduce the number of patches, while also optimizing their con- tent for training and inference. As a result, we can target a smaller voxel spacing (which effectively re- duces the physical extent of the patches) without the disadvantages of excessive computational costs and poormodelperformance. Conclusions We presented a novel centerline-guided method for fullyautomatedsegmentationof theaorticblood- lumen and the stent graft wire frame in abdominal CT-Ascans. Usingourmethod,both trainingand in- ference can be conducted more efficiently. The eval- uated DSC of 0.961 for the blood lumen and 0.841 for the stent graft wire frame suggest results that are suitable for medical analysis. In the future, we plan to use the results of our method for the analysis of risk factors for post-EVAR patients. Futhermore, we plantoextendtheuseofourmethodtoothermedical segmentation tasks. Acknowledgements This work was funded by the FFG (Austrian Re- search Promotion Agency) under the grants 851461 (EndoPredictor), 872604 (MEDUSA) and 867536 (vizARd). This project was supported by the strate- gic economic and research programme ”Innovatives OO¨ 2020” of the province of Upper Austria. RISC Software GmbH is a Member of UAR (Upper Aus- trianResearch) InnovationNetwork. References [1] S. Aggarwal, A. Qamar, V. Sharma, and A. Sharma. Abdominal aortic aneurysm: A comprehensive review. Experimental & Clinical Cardiology, 16(1):11–15,2011. [2] M. Amanuma, R. H. Mohiaddin, M. Hasegawa, A. Heshiki, and D. B. Longmore. Abdominal aorta: characterisation of blood flow and measurement of its regional distribution by cine magnetic resonance phase-shift velocity mapping. European Radiology, 2(6):559–564, Dec.1992. [3] A. W. Beck, A. Sedrakyan, J. Mao, M. Venermo, R.Faizer,S.Debus,C.-A.Behrendt,S.Scali,M.Al- treuther, M. Schermerhorn, B. Beiles, Z. Szeberin, N. Eldrup, G. Danielsson, I. Thomson, P. Wigger, M. Bjo¨rck, J. L. Cronenwett, and K. Mani. Varia- tions in abdominal aortic aneurysm care: A report from the international consortium of vascular reg- istries. Circulation, 134(24):1948–1958,2016. [4] J. Egger, B. Freisleben, R. Setser, R. Renapuraar, C.Biermann,andT.O’Donnell. Aortasegmentation for stent simulation. Computing Research Reposi- tory -CORR, 2011. [5] D. G. Ellis. 3DUnetCNN. github.com/ ellisdg/3DUnetCNN, 2018. Accessed: 2020- 02-11. [6] M.Freiman,S. J.Esses,L. Joskowicz, andJ.Sosna. An iterative model-constrained graph-cut algorithm for abdominal aortic aneurysm thrombus segmen- tation. In 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, pages 672–675,2010. [7] D. P. J. Howard, A. Banerjee, J. F. Fairhead, A. Handa, L. E. Silver, and P. M. Rothwell. Age- specific incidence, risk factorsandoutcomeofacute 106