Triple collaborative consistency with Mamba for semi-supervised 3D medical image segmentation

Current semi-supervised segmentation methods face the following challenges: (1) Cross-branch collaboration: Existing methods typically rely on single-branch pseudo-label generation or simple multi-view fusion strategies, failing to fully exploit the interaction between local details and global structures. This limitation leads to suboptimal performance in the boundary segmentation of complex anatomical structures. (2) Inefficiency in long-range modeling: While Transformer-based methods can capture global dependencies, they suffer from quadratic growth in computational complexity and the risk of overfitting when applied to high-resolution data (e.g., 3D medical images), making it difficult to balance efficiency and accuracy.

To address the above challenges, this article proposes a Tri-branch Collaborative Consistency model based on Mamba long-range modeling (TCC-Mamba), which aims to reduce reliance on annotations while improving segmentation accuracy in complex regions of medical images.

TCC-Mamba consists of a shared encoder and a tri-branch decoder. Specifically, a tri-branch collaborative supervision mechanism is introduced where three decoders form a closed-loop learning system through cross-pseudo-label supervision, enabling collaborative optimization and information sharing. Additionally, a geometric consistency loss function is incorporated to enhance boundary awareness. Furthermore, we integrate the SpatialTriMamba module, leveraging the efficient long-range dependency modeling of state-space models to achieve dynamic fusion of global context and local features, thereby improving segmentation accuracy for complex boundaries.

We conducted experiments on three public datasets: Left Atrium (LA), Pancreas CT, and ACDC, using 10%, 20%, and 30% labeled data. The results demonstrate that our method outperforms the six current advanced semi-supervised methods, achieving better segmentation performance.

The TCC-Mamba introduces novel methodologies in medical image segmentation tasks. This model combines the SpatialTriMamba module to capture long-range features and utilizes signed distance maps to enhance the use of geometric information, leading to exceptional results in handling complex anatomical structures. It provides an efficient and reliable solution for semi-supervised medical image segmentation.