Hierarchical Reliability-Guided Mutual Learning for Semi-Supervised Ultrasound Image Segmentation

The high cost of manually annotating ultrasound images has motivated semi-supervised learning approaches that leverage abundant unlabeled data. Mutual learning, which integrates consistency regularization and pseudo labeling through cross-network supervision, provides a promising solution for ultrasound image segmentation. However, the inherent characteristics of ultrasound images, such as speckle noise and artifacts, may lead to structurally inaccurate and semantically ambiguous pseudo labels, resulting in unreliable supervision and error propagation during iterative training. We propose Hierarchical Reliability-guided Mutual Learning (HRML), a dual-network framework that improves both pseudo label generation and reliability-aware pseudo label selection. HRML incorporates a Detail Refinement Module (DRM) to enhance fine-grained local structural representations and a Semantic Decoupling Module (SDM) to encourage semantically clearer high-level representations, thereby improving pseudo label quality. It further introduces a Multi-Dimensional Dynamic Reliability Weighting (MDRW) mechanism, which jointly evaluates the historical stability and semantic consistency of pseudo labels to suppress unreliable supervision and mitigate error propagation during mutual learning. Extensive experiments on four public ultrasound datasets confirm the superior performance of HRML over state-of-the-art fully supervised and semi-supervised methods.