Synergizing SAM2 and U-Mamba as Mutual Teachers in a Dual-Branch Dynamic Knowledge Transfer Network for Orbital Segmentation

Accurate segmentation of orbital structures in thyroid eye disease (TED) is critical for clinical staging and surgical planning. However, as an under-explored task, orbital multi-organ segmentation suffers from inherent data scarcity, resulting in substantial domain gaps when directly applying general-purpose segmentation models. These foundation models often lack sufficient exposure to orbital imaging data, producing anatomically implausible yet overconfident predictions. To address this challenge, we propose a Dual-Branch Dynamic Knowledge Transfer Network (DDKTN), which integrates the global semantic generalization capability of SAM2 with the fine-grained structural modeling strength of U-Mamba. Specifically, a Dynamic Feature Interaction (DFI) mechanism enables adaptively controlled inter-branch knowledge transfer, facilitating a progressive transition from SAM2-guided global semantics to U-Mamba-driven structural refinement. Moreover, a Bidirectional Consensus Confidence (BCC) regularization mechanism identifies consensus and discrepancy regions between the foundation and domain-specific branches, enforcing anatomy-aware consistency via mutual supervision. In addition, a Prompt Adaptation Module (PAM) leverages the anatomical sensitivity of the domain-specific branch to automatically generate structured prompts, thereby injecting reliable anatomical priors into SAM2. Experiments conducted on a private orbital CT dataset and a public coronal MRI benchmark demonstrate that DDKTN achieves superior performance, with average Dice scores of 93.68% and 94.65%, respectively. Ablation studies further validate the effectiveness of each component in improving generalization and anatomical fidelity. The implementation of our method is available at https://github.com/zltshadow/DDKTN.