Anatomically-Informed Dynamic Weighting for Robust Semi-Supervised Fetal MRI Segmentation

Semi-supervised learning (SSL) has proved to be an effective tool for medical image segmentation, as it leverages the use of unlabeled data. However, current SSL methods rely on fixed pseudo-label selection strategies and treat selected predictions with uniform importance, disregarding their relative confidence levels and potential quality differences. Furthermore, these methods fail to incorporate readily available anatomical information from labeled data, such as organ volume distributions and connectivity patterns. We present a new dynamic weighting method that addresses these limitations. It consists of three key components: 1) anatomical prior scoring to quantify deviations from expected anatomical characteristics inferred from the labeled data; 2) batch-wise pseudo-label weighting that uses these anatomical measures to dynamically adjust training emphasis; and 3) anatomically-informed component filtering that refines segmentation outputs by filtering results based on connectivity priors. These components work synergistically to measure anatomical deviations, maintain training stability, and ensure anatomically consistent predictions. We integrate our method within the Uncertainty-guided Collaborative Mean-Teacher (UCMT) framework. Evaluation on the segmentation of the liver and lungs in fetal MRI scans using five labeled scans and multiple splits of 150 unlabeled scans shows that our method consistently outperforms the baseline UCMT: fetal liver and lungs segmentation Dice scores increased by 11.6% (0.69 to 0.77) and by 2.5% (0.78 to 0.80), respectively. This indicates that the dynamic weighting approach, guided by anatomical knowledge, minimizes both annotation requirements and the need to manually verify unlabeled data.