Designing Ensembles for Uncertainty-aware Vessel Segmentations from MRI Data

Ristovski Gordan, Jawad Muhammad, K. Hahn Horst, Linsen Lars

Zusammenfassung

Medical visualization is concerned with the visual representationand analysis of medical data. Acquiring patient-speci c images is the startingpoint towards examinations and diagnoses, but these images are not free ofartifacts which introduce some error to the data. Moreover, for many medicalapplications these data need to be pushed through a pipeline of data processingsteps before they are visualized, rendered, and presented to the expert for in-spection. Each of these steps in the pipeline is based on some assumptions thata ect and inuence the result. Before deciding whether and how to presentthis uncertainty to the medical experts, the overall uncertainty is typicallycaptured at the segmentation step. To do this, one can either apply a prob-abilistic segmentation approach, or can apply many segmentation approachesand combine them to get to a probabilistic segmentation instead. However,the di erent segmentation algorithms give di erent results. Even within thesame approach, one may get rather di erent results based on the choice ofthe parameters the algorithm depends on. Therefore, it is of importance toinvestigate whether we can use the many varied results and combine them inorder to improve the segmentation accuracy and at the same time reduce theoverall segmentation uncertainty. In this work we present a pipeline for pre-processing vessel imaging data so that the vessels and their surrounding arevery roughly isolated from the rest of the image. We then proceed to applymany di erent hard segmentation algorithms as well as fuzzy segmentationalgorithms with di erent parameter settings and analyze which of these aresuitable for segmenting vessel data and what e ect these parameters have onthe segmentation accuracy and uncertainty. Finally, we combine the di er-ent segmentation outputs by using various probabilistic ensembles to use theknowledge gained for better nal result.