Boundary-Aware Network for Kidney Tumor Segmentation

Segmentation of the kidney and kidney tumors using computed tomography (CT) is a crucial step in related surgical procedures. Although many deep learning models have been constructed to solve this problem, most of them ignore the boundary information. In this paper, we propose a boundary-aware network (BA-Net) for kidney and kidney tumor segmentation. This model consists of a shared 3D encoder, a 3D boundary decoder, and a 3D segmentation decoder. In contrast to existing boundary-involved methods, we first introduce the skip connections from the boundary decoder to the segmentation decoder, incorporating the boundary prior as the attention that indicates the error-prone regions into the segmentation process, and then define the consistency loss to push both decoders towards producing the same result. Besides, we also use the strategies of multi-scale input and deep supervision to extract hierarchical structural information, which can alleviate the issues caused by variable tumor sizes. We evaluated the proposed BA-Net on the kidney tumor segmentation challenge (KiTS19) dataset. The results suggest that the boundary decoder and consistency loss used in our model are effective and the BA-Net is able to produce relatively accurate segmentation of the kidney and kidney tumors.