Unified Feature Consistency of Under-Performing Pixels and Valid Regions for Semi-Supervised Medical Image Segmentation

Abstract

Existing semi-supervised medical image segmentation methods based on the teacher-student model often employ unweighted pixel-level consistency loss, neglecting the varying difficulties of different pixels and resulting in significant deficits in segmenting challenging regions. Additionally, consistency learning often excludes pixels with high uncertainty, which destroys the semantic integrity of a medical image. To address these issues, we propose a novel unified feature consistency (UFC) of under-performing pixels (UPPs) and valid regions for semi-supervised medical image segmentation: 1) high-performing pixels (HPPs) and UPPs are distinguished by confidence differences between the student and teacher models, and then UPPs are mapped into a latent feature space to improve consistency learning effect (UPPFC); 2) in order to obtain richer semantic information from a medical image, vectors of valid regions are selected from both image- and patch-level class feature vectors by using the output probabilities of the teacher model; and 3) these vectors are mapped into the latent feature space for class feature consistency (CFC) learning as a supplement to UPPFC which only focuses on challenging regions for pixel-level consistency learning, thereby enhancing the model’s ability to learn structured semantic information from images themselves. Experimental results demonstrate that the proposed UFC achieves sufficient learning for challenging regions and retains the semantic integrity of medical images. Encouragingly, our proposed UFC provides better-segmentation results than the current state-of-the-art methods on three publicly available datasets.