Pay More Attention to Discontinuity for Medical Image Segmentation.

Medical image segmentation is one of the most important tasks for computer aided diagnosis in medical image analysis. Thanks to deep learning, great progress has been made recently. Yet, most existing segmentation methods still struggle at discontinuity positions (including region boundary and discontinuity within regions), especially when generalized to unseen datasets. In particular, discontinuity within regions and being close to the real region contours may cause wrong boundary delineation. In this paper, different from existing methods that focus only on alleviating the discontinuity issue on region boundary, we propose to pay more attention to all discontinuity including the discontinuity within regions. Specifically, we leverage a simple edge detector to locate all the discontinuity and apply additional supervision on these areas. Extensive experiments on cardiac, prostate, and liver segmentation tasks demonstrate that such a simple approach effectively mitigates the inaccurate segmentation due to discontinuity and achieves noticeable improvements over some state-of-the-art methods.