Morphological characterization of median nerve and transverse carpal ligament from ultrasound images using convolutional neural networks

Objectives: The purpose of this study was to automatically segment and quantify the median nerve and carpal arch from ultrasound images using convolutional neural network (CNN). Methods: A U-Net method based on CNN was implemented for median nerve and transverse carpal ligament segmentation from cross-sectional ultrasound images of the distal carpal tunnel. Median nerve and ligament were measured using the manual segmentations and model predictions. Model performance was evaluated using Dice score coefficient (DSC), recall, and precision. Model performance parameters and morphological parameters were compared between the healthy and carpal tunnel syndrome patients using Wilcoxon signed-rank test. The reliability of the morphological measurements from the predictions was assessed by calculating mean average error and the intra-class coefficient (ICC). Results: The DSC, recall, and precision were 0.89 ± 0.81, 0.94 ± 0.04, and 0.86 ± 0.08 for healthy subjects, respectively, for median nerve segmentation; the corresponding values for patients were 0.81 ± 0.08, 0.86 ± 0.10, and 0.77 ± 0.11, respectively. For ligament segmentation, the DSC, recall, and precision were 0.87 ± 0.03, 0.88 ± 0.04, and 0.87 ± 0.05, respectively, for healthy subjects; the corresponding values for patients were 0.77 ± 0.10, 0.77 ± 0.12, and 0.77 ± 0.09, respectively. Acceptable to excellent agreement was found between morphological measurements calculated using manual segmentations and model predictions. The carpal tunnel syndrome patients had larger median nerve cross-sectional area and carpal arch height than the healthy subjects when measured from the model predictions (p < 0.05). Conclusions: CNNs were used to automatically segment the median nerve and TCL with high accuracy. The model predictions provided reliable quantification of the carpal tunnel anatomy, demonstrating the potential diagnostic value using CNNs.