Automated Ultrasound-Based Severity Classification of Carpal Tunnel Syndrome: A Deep Learning Model Validated Against Nerve Conduction Studies

Ultrasound evaluation of the median nerve in carpal tunnel syndrome (CTS) is inherently operator-dependent. Automating CTS severity grading from ultrasound using deep learning may improve objectivity, reproducibility, and scalability in clinical practice. Background/Objectives: To develop convolutional neural network (CNN) models for automated CTS severity classification using B-mode ultrasound and electrodiagnostic results as the reference standard, and to compare performance between ultrasound scans. Methods: Fifty participants with suspected CTS provided 94 wrists, each classified into four severity levels (normal, mild, moderate, severe) through standardized nerve conduction studies. Ultrasound videos in transverse and longitudinal planes were acquired under a uniform protocol. From 11,895 frames, expert quality control produced a dataset of 2,518 valid images. Two CNN classifiers (transverse and longitudinal) were trained on nerve-centred crops using a 75/25 training–validation split, Adam optimization, and categorical cross-entropy loss. Performance was assessed through accuracy, class-wise precision, and confusion matrices. Results: Both models showed stable convergence. The transverse classifier achieved 0.92 validation accuracy, with precision values of 0.88 (normal), 0.94 (mild), 0.93 (moderate), and 0.97 (severe). The longitudinal classifier achieved 0.94 accuracy, with precision values of 0.98 (normal), 0.94 (mild), 0.96 (moderate), and 0.85 (severe). Misclassifications occurred mainly between adjacent categories, and reduced performance for severe longitudinal cases reflected limited sample representation. Conclusions: CNN-based classifiers can automatically predict CTS severity from B-mode ultrasound with high agreement to electrodiagnostic labels. These operator-independent models may support standardized diagnostic pathways and ultrasound-based screening. Future work should expand and balance datasets, incorporate multi-center data, and conduct external validation.