Distribution-Aware Multi-Attention Tri-branch Networks with Feedforward Differential Features for semi-supervised medical image segmentation

Abstract

Semi-supervised learning (SSL) is a challenging yet significant subject. However, previous SSL methods usually directly transfer the knowledge learned from labeled data to unlabeled data, resulting in their limited abilities to fully leverage the distribution discrepancy between labeled and unlabeled data. To tackle this issue, this work proposes a novel SSL framework called Distribution-Aware Multi-Attention Tri-branch Networks with Feedforward Differential Features (DAMATN-FDF) for semi-supervised medical image segmentation. DAMATN-FDF consists of a shared encoder and a tri-branch decoder equipped with different attention mechanisms. To bridge the distributional gap between labeled and unlabeled data, we introduce two key modules: Distribution-Aware (DA) and Integrity Supervision and Uncertainty Minimization (IS- UM). The DA module is designed to learn distribution-aware features. The IS-UM module is designed to encourage the inter-branch consistency for regularization. Besides, Feedforward Differential Features (FDF) are introduced to enhance the knowledge transferring across different decoder branches. Extensive experiments are conducted on three typical datasets like LA, Pancreas CT and BraTS-2019 datasets. Experimental results demonstrate the effectiveness of the proposed DAMATN-FDF method, significantly improving the performance over state-of-the-art methods. Code is publicly available at https://github.com/MapleUnderTheMooon/DAMATN-FDF.