Diagnostic performance of deep learning for brachial plexus ultrasound: A systematic review

Ultrasound-guided nerve block is a safe and effective regional anesthesia technique; however, accurate identification of the brachial plexus remains challenging due to its small size and low contrast in ultrasound images. Recent advances in deep learning offer promising solutions to enhance brachial plexus segmentation and improve perioperative regional anesthesia precision and safety. This review systematically summarizes current deep learning approaches applied to ultrasound-based brachial plexus segmentation. We highlight key models, including Convolutional Neural Networks, the U-shaped Convolutional Neural Networks and their variants, Mask RegionBased Convolutional Neural Networks, and Generative Adversarial Network-based architectures, and compare their reported performances, with Dice Similarity Coefficients ranging from 0.5865 to 0.882 and Intersection over Union values up to 0.6957. Among them, U-Net remains the most frequently employed due to its balance of accuracy and computational efficiency. Moreover, novel models such as multi-objective brachial plexus segmentation network and BPMSegNet have demonstrated superior segmentation performance by incorporating attention mechanisms and spatial contrast features. Notwithstanding these advancements, challenges persist, particularly limited dataset availability and insufficient model generalization. This review provides a comprehensive overview of recent progress, evaluates comparative performance metrics, and outlines future directions to improve model robustness and clinical applicability and clinical applicability in the perioperative setting.