Incremental Boundary Refinement using Self Axial Reverse Attention and Uncertainty-aware Gate for Colon Polyp Segmentation

Sang Dinh Viet, Quang Do Duy

Abstract: Medical image segmentation, particularly polyp segmentation, is a pervasive and widely applicable problem. Numerous recent works have been introduced to solve the problem in various ways. In practice, however, polyps can be extremely difficult to segment when considering the boundary region, where the information is specific and complicated. This paper proposes an Incremental Boundary Refinement method that focuses on learning and refining the boundary from both inside and outside the polyp region. In contrast to lego-like approaches that design deep neural architectures by stacking well-known existing blocks to improve performance, we perform a deep insight analysis of a state-of-the-art network known as SSFormer and then propose novel effective modules to address its weaknesses. Experimental results on five benchmark datasets indicate that our method successfully refines boundary information under various challenging conditions while archiving 80% mDice in the most complex set and more than 90% mDice in the other sets. These results surpass several existing colon polyp segmentation methods.