Author: Pelletier ED et al.
A & A Practice 19(8):e02040, August 2025. doi:10.1213/XAA.0000000000002040
This study evaluated deep-learning architectures for automating nerve identification in ultrasound images used for regional anesthesia. A dataset of 3594 images spanning 9 nerve block regions (e.g., TAP, femoral, brachial plexus, sciatic) was expanded through augmentation to 25,000 training images per region. Ten convolutional neural network models were trained with 5-fold cross-validation, and performance was measured against expert-labeled test images using Dice scores.
The R2U-Net model achieved the best overall accuracy with a mean Dice score of 0.76, slightly outperforming other models (0.71–0.76). Statistically significant performance variation was seen only for the TAP region. Expert validation showed high confidence in model-generated predictions—particularly in the popliteal region, where all 100 predictions were deemed safe for needle insertion. Logistic modeling suggested a positive correlation between Dice overlap and expert acceptance, notably in the supraclavicular region.
These findings highlight the promise of deep-learning tools for augmenting ultrasound-guided nerve block practice, though further clinical testing is needed before widespread adoption.
What You Should Know
• R2U-Net performed best for nerve segmentation, with a mean Dice score of 0.76.
• Performance differences were statistically significant only in the TAP region.
• Experts validated the predictions, with unanimous acceptance in the popliteal region.
• Greater Dice overlap correlated with higher expert trust, especially in supraclavicular blocks.
• Deep learning could enhance accuracy and safety in ultrasound-guided regional anesthesia but requires real-world validation.
Thank you to A & A Practice for making this work available.