Automating Brachial Plexus Scan: Wireless Handheld Ultrasound with Deep Learning over Ten Locations

Topographical ultrasound is gaining traction for brachial plexus visualization due to its value in regional anesthesia. However, existing artificial intelligence models for nerve localization are trained on high-resolution stationary ultrasound images, limiting their applicability to more convenient, low-resolution handheld devices. This study addresses this challenge by proposing a novel image segmentation model suitable for low-resolution images.

Thirty adult patients provided informed consent for participation in this study. A high-frequency, portable ultrasound probe was used to acquire B-mode images and video clips at 20 predefined positions. A training dataset of 60,000 images was constructed with expert annotations for landmark localization. A two-stage convolutional neural network architecture was implemented: Stage 1 for image classification and Stage 2 for segmentation with centroid refinement. Four novice physicians underwent testing for comparison. Model performance was evaluated using classification accuracy and segmentation precision metrics.

Our model achieved high accuracy in classifying brachial plexus ultrasound image positions (99.2% sensitivity and 84.5% specificity) compared to novice physicians (81.1% sensitivity and 59.8% specificity). In addition, the model demonstrated significantly superior performance in landmark segmentation, with lower median distance error (0.19 mm vs. 4.9 mm) and superior shape similarity metrics (average symmetric surface distance, Hausdorff distance and intersection over union) compared to novice physicians.

We present a novel image segmentation model for brachial plexus ultrasound images. The model achieved high classification accuracy and significantly surpassed novice physicians in landmark segmentation. This performance suggests the potential for the model as an educational tool to aid novice physicians in learning brachial plexus anatomy.