A dual-branch multi-scale network for thyroid nodule ultrasound image segmentation

Thyroid ultrasound images present intrinsic complexity due to the coexistence of low grayscale contrast, speckle noise, and ambiguous nodule boundaries, which jointly hinder reliable and stable segmentation. These challenges arise from the heterogeneous and multi-scale nature of ultrasound imaging. To address this issue, we propose a dual-branch multi-scale fusion network for thyroid nodule segmentation, designed to balance local detail preservation and global dependency modeling. The encoder integrates a residual network branch to capture fine-grained boundary and texture features, together with a visual state space branch that models long-range dependencies and global contextual relationships. Multi-scale feature fusion between the two branches enables complementary representation across different semantic levels. In addition, a convolutional refinement unit is embedded after each visual state space module to compensate for the loss of boundary details during network modeling. Meanwhile, an adaptive channel weighting mechanism is introduced to dynamically enhance effective nodule features and suppress noise interference, further improving the robustness of the model. Furthermore, a window-free lightweight token attention module is embedded in high-level residual network stages to efficiently capture long-range semantic dependencies without window partitioning. Extensive experiments conducted on two public thyroid ultrasound datasets demonstrate that the proposed method consistently outperforms representative advanced approaches in segmentation accuracy and boundary delineation. The proposed method shows consistent performance across the two evaluated thyroid ultrasound datasets.