SynMatch : a synergistic system for multi-level semi-supervised medical image segmentation

Abstract

Semi-supervised learning (SSL) is critical for medical image segmentation where annotated data is scarce, yet existing methods often struggle with an inherent trade-off between pseudo-label quality and diversity. To address this challenge, we propose SynMatch, a synergistic framework designed for multi-level learning. The framework introduces two key innovations: (1) a Dynamic Gating Mechanism (DGM) that adaptively balances cross-teaching and consistency regularization for robust semantic learning, and (2) a Synergistic Consensus-Gated Boundary Loss (SCBL) that leverages cross-model agreement for precise boundary refinement. By systematically integrating solutions for both semantic and boundary-level challenges, SynMatch provides an effective approach to SSL segmentation. Extensive experiments on five public benchmarks, spanning diverse modalities (MRI, CT, and fundus photography), demonstrate that our framework achieves competitive performance, significantly improving segmentation accuracy in data-scarce settings (e.g., a + 19.07% absolute Dice improvement over UniMatch on PROMISE12). Furthermore, our analysis reveals important insights into the interplay between our boundary module and architectural diversity. We demonstrate that while homogeneous architectures provide exceptional stability on uniform data, the synergy between heterogeneous configurations and our SCBL module proves crucial for achieving robust boundary precision on high-variability clinical datasets. These findings highlight SynMatch's effectiveness for data-efficient learning and demonstrate that architectural choice represents a significant design consideration for SSL frameworks, particularly in challenging clinical scenarios with diverse data characteristics.