Deep Learning-Based Segmentation of the Brachial Plexus in Ultrasound Images: A Cross-Device Generalization Assessment

Ultrasound-guided brachial plexus regional anesthesia is a commonly used clinical technique for upper limb surgery analgesia. It provides excellent efficacy and high safety. However, due to the presence of substantial noise in ultrasound image pixels and the complex anatomical structures, accurate identification of nerves heavily depends on the operator’s experience. Precise nerve identification is crucial for patient recovery. Deep learning-based image segmentation can automatically identify the location of the brachial plexus in ultrasound images, thereby assisting clinicians in performing brachial plexus nerve blocks.In this study, we systematically compared the performance of three neural network architectures for brachial plexus segmentation in ultrasound images, including CNN-based, Transformer-based, and Mamba-based models. The experimental data come from ultrasound images acquired by three different devices (eSaote, Sonosite, and Butterfly). All models were trained on data from the eSaote device and tested on images from both eSaote and the Sonosite and Butterfly devices.The results indicate that on the eSaote device dataset, ConvUNeXt and UNet achieved the highest mean Intersection over Union (mIoU), with scores of 0.9027 and 0.9043, respectively. However, in cross-device testing, TransUNet and VMUNet exhibited better generalization ability. On the low-quality Butterfly test set, TransUNet maintained strong segmentation performance.In addition, the models showed some limitations in data dependency and cross-domain adaptability, and possible directions for improvement are suggested. This study can serve as a reference for selecting and optimizing ultrasound nerve segmentation models.