A Few-Shot Medical Image Segmentation Network with Boundary Category Correction

Accurate medical image segmentation is the foundation of clinical imaging diagnosis and 3D image reconstruction. However, medical images often have low contrast between target objects, greatly affected by organ movement, and suffer from limited annotated samples. To address these issues, we propose a few-shot medical image segmentation network with boundary category correction named Boundary Category Correction Network (BCC-Net). Of overall medical few-shot learning framework, we first propose the Prior Mask Generation Module (PRGM) and Multi-scale Feature Fusion Module (MFFM). PRGM can better localize the query target, while MFFM can adaptively fuse the support set prototype, the prior mask and the query set features at different scales to solve the problem of the spatial inconsistency between the support set and the query set. To improve segmentation accuracy, we construct an additional base-learning branch, which, together with the meta-learning branch, forms the Boundary Category Correction Framework (BCCF). It corrects the boundary category of the meta-learning branch prediction mask by predicting the region of the base categories in the query set. Experiments are conducted on the mainstream ABD-MR and ABD-CT medical image segmentation public datasets. Comparative analysis and ablation experiments are performed with a variety of existing state-of-the-art few-shot segmentation methods. The results demonstrate that the effectiveness of the proposed method with significant enhance the segmentation performance on medical images.