Title: Quality-Guided Semi-Supervised Learning for Medical Image Segmentation

Abstract:

Developing precise models for medical image segmentation is traditionally hindered by the high cost and time intensity associated with acquiring large volumes of densely annotated data. Semi-supervised learning (SSL) offers a solution by leveraging both scarce labeled samples and plentiful unlabeled data. Nevertheless, contemporary SSL techniques predominantly depend on pseudolabels for unlabeled instances, evaluating their trustworthiness via model confidence or uncertainty metrics. These metrics are inherently self-referential and fail to provide explicit grounding in actual segmentation quality. To address this limitation, we introduce a quality-guided SSL framework that employs a specialized network to estimate segmentation quality based on image-mask pairs. This predictor is trained using masks of varying quality, which are generated through synthetic corruptions and combined with imperfect outputs from partially trained segmentation models, thereby capturing the realistic error patterns typical of the training process. We incorporate the quality predictor into SSL via two distinct mechanisms: a quality-aware regularization loss and a pseudolabel sample reweighting scheme grounded in quality metrics. Our results indicate that this approach functions as a seamless enhancement to existing SSL frameworks. Comprehensive experiments conducted across five datasets and multiple architectures reveal consistent performance gains over rival SSL methods, thereby pushing the boundaries of semi-supervised medical image segmentation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC