Back to 2026 ePosters
An AI-Driven Pipeline for Localization, Segmentation, and Classification of Carpal Tunnel Syndrome Using Ultrasound Images of the Median Nerve
Rachel Hyzny, BA
1, Charles Patterson, BS
1, Abhigyan Kishor, MS
1, Nickolas Littlefield, MS
1, Fritz Steuer, MD
1, Jacob Weinberg, BS
1, John R Fowler, MD
2; Ahmad Tafti, PhD, FAMIA
1(1)University of Pittsburgh, Pittsburgh, PA, (2)Department of Orthopaedic Surgery, University of Pittsburgh, Pittsburgh, PA
Introduction:
Carpal tunnel syndrome (CTS) is a common peripheral neuropathy caused by compression of the median nerve. Although tools such as the CTS-6 score and ultrasound-based cross-sectional area (CSA) measurements are available, diagnosis often relies heavily on clinical evaluation. These methods, however, have limitations in both accuracy and objectivity. This study introduces an AI-driven ultrasound pipeline that integrates YOLOv11 for localization, U-Net for segmentation, and ConvNeXt for classification to enhance the precision and efficiency of CTS diagnosis.
Materials and Methods:
Patients ?18 years who were assessed for CTS between June 2023 and September 2024 were included. All patients were evaluated by a board-certified hand surgeon, and ultrasound images of the median nerve were obtained at the carpal tunnel inlet. 121 wrists (73 with CTS, 48 without CTS) were utilized to train and test the AI system. The pipeline includes 3 steps: (1) YOLOv11 identifies the median nerve, (2) U-Net outlines its structure, and (3) ConvNeXt classifies the image as CTS-positive or CTS-negative. A conservative learning rate of 0.0005 ensured training stability and reduced the risk of overfitting. An overview of the AI learning pipeline is provided in Figure 1.
Results:
The ConvNeXt classification achieved 94.1% accuracy in distinguishing CTS-positive from CTS-negative cases, with a positive predictive value (PPV) of 0.86 and a sensitivity of 1.0. The YOLOv11 localization model achieved an average PPV of 0.95 and sensitivity of 0.98. The U-Net segmentation model achieved a validation Intersection over Union of 0.86, indicating strong overlap between predicted and expert-annotated nerve boundaries.
Discussion and Conclusion:
This AI-driven ultrasound pipeline demonstrates strong diagnostic performance in identifying CTS and may serve as a cost-effective, objective diagnostic modality. By automating median nerve evaluation, it has the potential to reduce reliance on electrodiagnostic testing and minimize inter-operator variability. Despite being limited by a relatively small dataset, this study supports the clinical utility of an AI-based approach for CTS identification and warrants future large-scale validation studies.
Back to 2026 ePosters
