APPLICATION OF REFINEMENTS ON FASTER-RCNN IN AUTOMATIC SCREENING OF DIABETIC FOOT WAGNER GRADES

Objective: To developed auto screening diabetic foot Wagner grade systems as a means of assisted diagnosis and assessment to alleviate part of the workload for podiatrists. Methods: we propose to use the Faster-RCNN algorithm based on the ResNet-50 backbone network to achieve automatic detection and localization of the Wagner grades of diabetic foot. To build a robust deep learning model, we collected 2,688 images of the diabetic foot as datasets for model training. We combines the Kmeans++ algorithm to improve the generation method of anchor boxes and obtains the Wagner grades automatic screening model for the diabetic foot with good robust performance. Results: By improving the generation method of anchor boxes, refinements on Faster-RCNN models reach a mean average precision (mAP) of 91 36% in the diabetic foot datasets. Conclusion: This work has the potential to lead to nursing methods shift in the clinical treatment of diabetic feet in the future, to provide a better self-management solution for patients with diabetic feet.