Automated Segmentation of Median Nerve Cross-Sectional Area in Healthy Controls and Patients With Carpal Tunnel...

The cross-sectional area (CSA) of the median nerve (MN) is a key parameter for confirming carpal tunnel syndrome (CTS) with ultrasound. This study evaluates the performance of a convolutional neural network (CNN) with a 2D U-Net architecture for automated MN CSA segmentation in both healthy individuals and patients with CTS. Automated segmentation supports large-scale analysis, improves standardization across centers, and may serve as a foundation for future diagnostic tools.

Three hundred static ultrasound images from 50 healthy participants and 300 from 74 patients with CTS were used to train and validate a five-layer U-Net model. Each CSA measurement was performed in triplicate. Model performance was evaluated using the Dice similarity coefficient (DSC) and by comparing automated with manually annotated CSA values. Prospective evaluation was conducted on a small, unseen dataset and on data acquired using a different ultrasound system.

Manual CSA measurements showed excellent repeatability (ICC 0.982). The model achieved high DSCs (0.95 in controls, 0.96 in CTS) and showed no statistically significant difference from manual CSA measurements (p = 0.227). Segmentation inaccuracies in the test set were primarily attributable to minor contouring differences at the epineurial border. On unseen data, errors occurred in more proximally scanned images or scans showing intraneural hyperechogenicity.

In line with current literature, a CNN with a 2D U-Net architecture demonstrates strong potential for MN segmentation on static ultrasound images. However, reduced accuracy on unseen data suggests overfitting. Further validation is required to ensure robustness across anatomical and technical variation before clinical implementation.