BGBF-Net: Boundary-Guided Buffer Feedback Network for Liver Tumor Segmentation

Medical images such as CT can provide important reference value for doctors to diagnose diseases. Identifying and segmenting lesions from medical images is crucial for its diagnosis and treatment. However, unlike other segmentation tasks, medical image has the characteristics of blurred boundaries and variable lesions sizes, which poses challenges to medical image segmentation. In this paper, we propose a Boundary-Guided Buffer Feedback Network(BGBF-Net), using the boundary guidance module to combine the low-level feature map rich in boundary information and the high-level semantic segmentation feature map generated by the encoder module, and output the features that focus on the boundary, which is used to enhance the attention of the decoder to boundary features. The buffer feedback module is used to strengthen the network's supervision of the decoder while speeding up convergence of the model. We apply the proposed BGBF-Net on the LiTS dataset. Comprehensive results show that the BGBF-Net improves by 2.36% compared to other methods in terms of Dice.