CT-Match: A Unified Semi-Supervised Learning Framework for Ultrasound Breast Tumor Image Segmentation

Breast tumor segmentation in ultrasound imaging confronts inherent challenges including acoustic artifacts, intensity inhomogeneity, and limited expert annotations. While semi-supervised learning frameworks mitigate annotation scarcity through consistency constraints, existing approaches exhibit architectural homogeneity and constrained perturbation diversity. We propose Crossing Teaching Match, featuring: (1) A heterogeneous dual-branch framework fusing CNN's inductive biases with Transformer's global context via cross-architectural consistency constraints; (2) A hierarchical perturbation paradigm integrating feature-level strong augmentations and dual image-level stochastic transformations to maximize perturbation diversity. Experimental validation demonstrates superior generalizability across imaging protocols while substantially reducing annotation demands, offering practical clinical deployment potential in resource-constrained settings.