Semi-Supervised Image Segmentation via Heterogeneous Dual-Teacher Mechanism and Multi-Directional Feature Modeling
TL;DR
This work proposes DTCL-Swin-UMamba, a novel Swin-UMamba backbone that fuses global context from Swin Transformer with linear-complexity long-range dependency learning from Mamba to address label scarcity and anisotropic structural distortion from unidirectional modeling in semi-supervised cardiac MRI segmentation.
Abstract
To address label scarcity and anisotropic structural distortion from unidirectional modeling in semi-supervised cardiac MRI segmentation, we propose DTCL-Swin-UMamba. It features a novel Swin-UMamba backbone and a Multi-directional Vision-based Spatial Synthesis Module (M-VSS). To mitigate the noise sensitivity of traditional Mamba, the weight-shared M-VSS executes eight-directional parallel spatial scanning to build isotropic feature representations, enhancing thin-wall and direction-dependent structure segmentation. The Swin-UMamba backbone fuses global context from Swin Transformer with linear-complexity long-range dependency learning from Mamba. Furthermore, a heterogeneous dual-teacher training strategy—comprising an EMA-updated Swin Transformer teacher and a frozen MS-UNet teacher—provides complementary global and local structural supervision for the student network. Evaluated on the public ACDC and LA datasets, our framework achieves state-of-the-art performance. On ACDC with only 10% labeled images, it achieves Dice scores of 92.15%, 89.57%, and 87.25% for the left ventricle, right ventricle, and myocardium, respectively, with consistent gains demonstrated on the LA dataset.