Noninvasive Assessment of the Ulnar Nerve in the Upper Extremity Using Machine Learning.
Authors
Affiliations (1)
Affiliations (1)
- Department of Orthopaedic Surgery, University of Rochester Medical Center, Rochester, NY.
Abstract
The purpose of this study was to develop a machine learning algorithm trained on ultrasound images of the cubital tunnel that can be used to automatically identify, segment, and measure the cross-sectional area (CSA) of the ulnar nerve. Control subjects and patients were scanned using a Fujifilm-SonoSite ultrasound system, equipped with a high-frequency linear array probe by a trained technician or physician. The ulnar nerve was identified and segmented in individual frames using a custom graphical user interface application. A convolutional neural network (YOLOv8) was then trained on these images to automatically detect the ulnar nerve, resulting in a binary map. The CSA and Dice score were then computed using the binary map of the nerve outline from the prediction and ground truth to assess the prediction accuracy. In total, 34 subjects (11 patients and 23 controls) were imaged with ultrasound. A total of 2,011 ultrasound grayscale images were extracted, and the ulnar nerve was segmented from these scans. In total, 1,286 images from 23 subjects (seven patients and 16 controls) were used to train the model, 425 images from five subjects (two patients and three controls) were used for validation, and 300 images from six subjects (two patients and four controls) for testing. The machine learning model (YOLOv8) resulted in an average Dice score of 0.90 with 296 images (99%) achieving a Dice score above 0.75. When comparing the CSA predictions from the machine learning model to the area derived from the ground truth, the mean difference and mean absolute difference were 1.78 (13.80%) mm<sup>2</sup> and 2.08 mm<sup>2</sup> (16.05%), respectively. Ultrasound is an emerging noninvasive modality for the diagnosis of cubital tunnel syndrome. However, its implementation remains challenging because of operator dependence and technical variability. This work establishes a foundational step toward fully automated, operator-independent ultrasound assessment of the ulnar nerve. Diagnostic III.