BSNet: a boundary-aware medical image segmentation network

Accurate segmentation of medical images can provide foundations for clinical and disease diagnosis. Inaccurate segmentation boundaries often result from limited contextual information and insufficient discriminating feature maps after consecutive pooling and upsampling operations in most existing methods. In this paper, we present a novel boundary-aware medical image segmentation network (BSNet) for resolving the multi-objective segmentation problem. We exploit a backbone network to extract multi-scale feature representations and design an adaptive contrast boundary-aware module (ACB), which uses the method of combining nonlinear filters with deep learning to extract high-quality boundary maps. We then build a feature fusion (FF) module to fuse multi-scale features with boundary maps, providing decoder with rich multi-scale features enhanced with boundary information, and facilitating cross-channel interactions. To further enhance the uncertain regions of the boundaries, we utilize the boundary spatial enhancement (BSE) module to learn the feature map of boundary locations with the assistance of the Sobel operator. We conducted experiments with three challenging public datasets to evaluate the effectiveness of BSNet. Simulation results on various datasets show that the present model outperforms state-of-the-art segmentation methods, obtaining up to 2.73% improvement in Dice coefficient (DICE) score. BSNet opens new ways of designing better boundary-aware segmentation network.Please confirm the corresponding author is correctly identified.No problem.