Abstract
Deep learning-based medical image segmentation techniques have shownpromising results when evaluated based on conventional metrics such as the Dicescore or Intersection-over-Union. However, these fully automatic methods oftenfail to meet clinically acceptable accuracy, especially when topologicalconstraints should be observed, e.g., continuous boundaries or closed surfaces.In medical image segmentation, the correctness of a segmentation in terms ofthe required topological genus sometimes is even more important than thepixel-wise accuracy. Existing topology-aware approaches commonly estimate andconstrain the topological structure via the concept of persistent homology(PH). However, these methods are difficult to implement for high dimensionaldata due to their polynomial computational complexity. To overcome thisproblem, we propose a novel and fast approach for topology-aware segmentationbased on the Euler Characteristic ($\chi$). First, we propose a fastformulation for $\chi$ computation in both 2D and 3D. The scalar $\chi$ errorbetween the prediction and ground-truth serves as the topological evaluationmetric. Then we estimate the spatial topology correctness of any segmentationnetwork via a so-called topological violation map, i.e., a detailed map thathighlights regions with $\chi$ errors. Finally, the segmentation results fromthe arbitrary network are refined based on the topological violation maps by atopology-aware correction network. Our experiments are conducted on both 2D and3D datasets and show that our method can significantly improve topologicalcorrectness while preserving pixel-wise segmentation accuracy.