Presentations

Abstract

Introduction: Depending on the severity, site, or cause of peripheral nerve injuries, inaccurate or late detection of the injured nerve poses a challenge for patient prognosis and the ability to offer surgical intervention. Peripheral nerve injuries are diagnosed by initial cross-sectional imaging i.e. CT, MRI, or US, followed by physical exam maneuvers and EMG testing. Inaccurate radiological detection during the initial visit can lead to patients having permanent disability and lower quality of life. Modern advances in artificial intelligence and machine learning algorithms have enabled automated localization and segmentation of peripheral nerves. By using cross-sectional imaging that is already obtained during initial patient encounters, applying machine learning models can assist clinicians in early detection of peripheral nerve injuries. This review investigates the use of machine learning algorithms to detect Carpal Tunnel Syndrome on ultrasound images. Carpal Tunnel Syndrome is caused by compression of the median nerve which can be difficult to visualize on ultrasound images due to operator dependence and morphological variations.

Methods: Independent reviewers conducted a literature search using PubMed, Web of Science, and MEDLINE (through July 1, 2024) to identify studies using machine learning algorithms for detecting carpal tunnel syndrome on ultrasound images. The database searches used were ("median nerve" OR "carpal tunnel syndrome") AND ("ultrasonography" OR "ultrasound" OR "sonography") AND ("artificial intelligence" OR "deep learning" OR "machine learning" OR "convolutional neural network"). Articles were screened for sonographic detection of the median nerve using machine learning algorithms by either localization, segmentation, or both.

Results: Eleven articles were used after screening 74 articles, comprising 593 participants. The machine learning algorithms used include U-Net, SegNet, MATLAB, Mask R-CNN, Resnet, MNT-DeepSL, DeepLabV3+, and FPN. Machine learning algorithm performance metrics obtained include precision, accuracy, recall, F-score, IoU, and dice coefficient.

Conclusion: Machine learning algorithms show capability for automating localization and segmentation of the median nerve on ultrasound images by statistically significant measures. Carpal tunnel syndrome can be difficult to visualize by ultrasound due to operator dependence and morphological variations of the median nerve. Further research should involve more participants varying in age, sex, and comorbidities to account for morphological variations of the median nerve in order for machine learning algorithms to be a reliable tool for sonographers in clinical practice.