Adaptive consistency representation learning for semi-supervised 3D medical image segmentation.
Abstract
Three-dimensional medical image segmentation plays a critical role in clinical diagnosis and surgical planning. Semi-supervised learning provides an effective way to reduce annotation requirements; however, heterogeneous feature perturbations may induce predictions with different levels of local disagreement, making uniformly weighted consistency supervision sub-optimal. This work presents an Adaptive Consistency Representation Learning (ACRL) framework for semi-supervised 3D medical image segmentation. ACRL constructs heterogeneous feature-perturbed auxiliary views and adaptively modulates the voxel-wise consistency strength according to the disagreement between each auxiliary prediction and the feature-unperturbed main prediction. A Multi-scale Feature Aggregation module and a Bottleneck Attention Module are incorporated to enhance structural representation. Experiments on the LA, Pancreas-CT, and BraTS19 datasets achieve Dice scores of 90.54%, 81.85%, and 87.51%, respectively, under the 10% labeled setting. Across the evaluated settings, ACRL shows competitive overlap accuracy and favorable boundary-distance performance relative to representative conventional and foundation-model-assisted semi-supervised baselines. Controlled ablations further show that reliability-adaptive weighting provides additional benefits over uniformly weighted multi-branch consistency.