DIFNet: A Dual-Branch Interactive Fusion Network for Femoral Nerve Segmentation in Ultrasound Images

Accurate femoral nerve (FN) segmentation in ultrasound images is crucial for nerve blocks but remains challenging due to its tortuous structure, large-scale variations, and low-contrast boundaries with the closely adjacent femoral artery. Existing methods often yield fragmented segmentations or high false-positive rates due to inadequate long-range dependency modeling and insufficient discriminative feature learning. To address these FN-specific challenges, we propose a Dual-branch Interactive Fusion Network (DIFNet). It employs parallel CNN and Transformer encoders to capture both local details and global context. Specifically, the Cross-Branch Interaction Module (CBIM) bridges semantic gaps and enhances structural continuity; the Multi-Scale Dilated Fusion module (MSDF) handles large-scale variations; and the Region-Guided Enhancement Module (RGEM) refines boundaries and suppresses false positives by explicitly modeling foreground, background, and boundary regions. Experiments demonstrate that DIFNet outperforms state-of-the-art methods on both public (mDice: 90.72%, mIoU: 83.31%) and private datasets (mDice: 91.69%, mIoU: 84.97%), providing more continuous, complete, and accurate segmentations. These results highlight DIFNet's robustness and effectiveness for femoral nerve segmentation in ultrasound imaging.