Conflict-aware semi-supervised mutual learning for medical image segmentation

Semi-supervised learning (SSL) has shown promising performance in medical image segmentation by effectively utilizing extensive unlabeled images. However, inaccurate predictions of unlabeled images can significantly impair the segmentation performance of SSL models. Furthermore, most current SSL methods lack mechanisms to handle cognitive bias, causing the model easily overfit on inaccurate predictions and making self-correction challenging. In this work, we propose a conflict-aware semi-supervised mutual learning framework (CSSML), which integrates two different subnetworks and selectively utilizes conflicting pseudo-labels for mutual supervision to address these challenges. Specifically, we introduce two subnetworks with different architecture incorporating a conflict-aware distinct feature learning (CDFL) regularization to avoid the homogenization of subnetworks while promoting diversified predictions. To handle potential inaccurate predictions, we introduce a geometry-aware mutual pseudo supervision (GMPS) regularization to determine the reliability of conflicting pseudo-labels of unlabeled images, and selectively leverage the more reliable pseudo-labels in the two subnetworks to supervise the other one. The synergistic learning between CDFL and GMPS regularizations during the training process facilitates each subnetwork to selectively incorporates reliable knowledge from the other subnetwork, thereby helping the model overcome cognitive bias. Extensive experiments on three public medical image datasets demonstrate that the proposed CSSML achieves an average of 80.65% DSC, 87.83% Precision, and 14.48mm 95HD using only 20% labeled data, highlight-ing its superior performance. The code is available at: https://github.com/Mwnic-AI/CSSML.