SwinMalle-UNet: A Hybrid U-Net Integrating Swin Transformer and Deformable Fusion Modules for Brachial Plexus Nerve Segmentation in Ultrasound Imaging

Precise segmentation of brachial plexus nerves in ultrasound imaging is essential for clinical anesthesia guidance and neurological evaluation. However, achieving precise nerve boundary delineation remains challenging due to the inherently low contrast, speckle noise, and the complex, variable morphology of nerves in ultrasound data. While deep learning models have shown promise, traditional Convolutional Neural Networks (CNNs) are constrained by fixed receptive fields, limiting their adaptability to non-rigid nerve structures. Conversely, Transformer-based models, while excelling at global context modeling, often face challenges with fine-grained detail recovery in data-limited medical scenarios. To address these limitations, we propose SwinMalle-UNet, a novel hybrid segmentation network that innovatively integrates the Swin Transformer with a deformable fusion mechanism within the classic U-Net architecture. The encoder employs a hierarchical Swin Transformer to efficiently capture multi-scale global contextual information. The decoder, our core contribution, features a deformable fusion module that dynamically adjusts feature sampling positions in a data-driven manner. This design facilitates precise pixel-level feature alignment and enhances adaptability to the irregular morphology of brachial plexus nerves. We comprehensively evaluated our model on a core multi-target brachial plexus dataset (UBPD) and a large-scale public benchmark (USNS). Experimental results demonstrate that our approach achieves state-of-the-art performance, outperforming a wide range of baseline and state-of-the-art models such as U-Net, Attention U-Net, DeepLabV3+, TransUNet, an EfficientFormerV2-based model (EFv2-UNet), and a strong Swin-UNet baseline.