Detection of Peripheral Neuropathy from Walking Videos Using Pose Estimation and Anatomical Segmentation
Résumé
Peripheral neuropathy, a common complication of diabetes and kidney disease, impairs gait and often goes undetected in early stages. Current diagnostic methods require specialized equipment, limiting their accessibility in low-resource and home environments.
This research aims to develop a contactless, video-based deep learning pipeline for early detection of peripheral neuropathy. The method integrates YOLOv11-based pose estimation and SAM-guided anatomical segmentation to extract lower limb features, followed by gait classification using a pretrained ResNet34 model.
The approach was evaluated on 113 sagittal-view walking videos from 57 clinical participants, with strict participant-level separation between training and testing. Our pipeline achieved an accuracy of 80% when using the left ankle keypoint to guide segmentation, outperforming both YOLO-only and SAMonly baselines. The system also localized clinically relevant gait features without requiring wearable sensors.
This work demonstrates the feasibility of a low-cost, scalable, and fully contactless screening tool for early peripheral neuropathy detection, with strong potential for deployment in remote or home settings.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
