DT-MSMamba: dual-teacher consensus with multi-scale Mamba-CNN interaction for semi-supervised 3D medical image segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Although accurate semi-supervised 3D medical image segmentation is vital for clinical diagnostics, existing approaches face two critical bottlenecks. On the one hand, most traditional networks struggle to simultaneously capture global contextual information and preserve fine-grained edge details in their architecture. On the other hand, conventional single-teacher learning paradigms remain highly susceptible to ‘confirmation bias’ driven by noisy pseudo-labels. To address these issues, we propose a novel semi-supervised framework called DT-MSMamba, which combines a dual-path backbone network with a robust dual-teacher consensus strategy. In the proposed method, to seamlessly fuse representations from local CNN branches and global Mamba branches, we design a novel Bi-path spectral-response cross-attention module, which ensures structural consistency at complex tissue boundaries and fully leverages the synergistic advantages of the dual-branch architecture. Furthermore, to overcome the ambiguity of tissue boundaries caused by pseudo-labels, we propose an entropy-weighted dual-teacher consensus method which incorporates an entropy-weighted soft fusion mechanism, assigning higher weights to high-confidence regions to mitigate error accumulation. Finally, to mitigate confirmation bias, we design a unified multi-constraint loss to achieve high-fidelity semantic consensus through uncertainty-rectified deep supervision and cross-architecture probability alignment, ultimately facilitating robust and clear boundary delineation. Extensive experiments on the pancreas-CT, LA, and BraTS2019 datasets show that when 20% of the datasets are labeled, the Dice scores are 84.15%, 92.30%, and 88.24%, respectively. The proposed method represents a promising solution for more accurate and informative 3D medical image segmentation and clinical research.