Pseudo-label guided selective mutual learning for semi-supervised 3D medical image segmentation

Abstract

Semi-supervised learning (SSL) have shown promising results in 3D medical image segmentation by utilizing both labeled and readily available unlabeled images. Most current SSL methods predict unlabeled data under different perturbations by employing subnetworks with same architecture. Despite their progress, the homogenization of subnetworks limits the diverse predictions on both labeled and unlabeled data, thereby making it difficult for subnetworks to correct each other and giving rise to confirmation bias issue. In this paper, we introduce an SSL framework termed pseudo-label guided selective mutual learning (PLSML), which incorporates two distinct subnetworks and selectively utilizes their derived pseudo-labels for mutual supervision to mitigate the above issue. Specifically, the discrepancies of pseudo-labels from two distinct subnetworks are used to select the regions within labeled images that are prone to missegmentation. We then introduce a mutual discrepancy correction (MDC) regularization to revisit these regions. Moreover, a selective mutual pseudo supervision (SMPS) regularization is introduced to estimate the reliability of pseudo-labels of unlabeled images, and selectively leverage the more reliable pseudo-labels in the two subnetworks to supervise the other one. The integration of MDC and SMPS regularizations facilitates inter-subnetwork mutual correction, consequently mitigating confirmation bias. Extensive experiments on two 3D medical image datasets demonstrate the superiority of our PLSML as compared to state-of-the-art SSL methods. The source code is available online at https://github.com/1pca0/PLSML.