SM2C – Boost the semi-supervised segmentation for medical image by using meta pseudo labels and mixed images

Abstract

Recently, semi-supervised learning methods have effectively leveraged unlabeled data to address the scarcity of annotated medical images. However, unlike common object datasets, the limited medical image resources often lead to overfitting due to significant shape variations of specific organs across cases or even within different sections of the same case. The intricate shapes of organs and lesions in medical images introduce additional complexity in auto-diagnosis, hindering the generalization of networks. To address this challenge, we propose a novel method, Scaling-up Mix with Multi-Class (SM${}^2$C), to synthesize organs and lesions with diverse shapes for clinical diagnosis. Integrated into a teacher–student framework, SM${}^2$C enhances the reliability of pseudo-labels generated by the teacher network, thereby improving the generalization of the student network. This method employs three key strategies: scaling up image size, multi-class mixing, and object shape jittering. We conduct ablation studies to validate the SM${}^2$C design, demonstrating its effectiveness in diversifying segmentation object shapes. In detail, multi-class mixing preserves inter-class balance, object shape jittering generates the various shapes that may appear in clinical diagnosis, and scaling up image size enriches context while enhancing robustness. Furthermore, Extensive experiments on three benchmark medical segmentation datasets further show solid gains compared with other state-of-the-art methods.