A deep learning-based performance-oriented approach for diabetic foot ulcer identification and localization using Faster R-CNN

The major health problem linked to diabetes mellitus is the diabetic foot ulcer (DFU). It must be noticed periodically in order to stop foot ulcers from spreading. In this research work, five architectures (VGG16, VGG19, ResNet50, NASNetMobile, and DenseNet121) were evaluated for ulcer classification. These models were used 2830 photos of both healthy and diabetic foot ulcer to detect DFU images. The accuracy of NASNetMobile was 98.58% where ResNet50 got 92.92%, VGG16 achieved 97.64%, VGG19 gained 96.23% and lastly DenseNet121 achieved 98.29%. So the researcher selected NASNetMobile as our base model for classification because the model achieved the highest accuracy 98.58%. The proposed model used NASNetMobile for classification (with high confidence: sigmoid output > 0.3) and used bounding box regressor network of Faster R-CNN (ResNet-50-FPN) to locate ulcers position, filtering region proposals using an IoU threshold of 0.3 for Non-Maximum Suppression (NMS). The size limitations (area > 1,000 pixels) were applied to ignore small false positives. Through decision fusion, an image is classified as abnormal if either NASNetMobile predicts abnormality (confidence > 0.3) or Faster R-CNN detects valid ulcer regions. The real-time localization of DFU is demonstrated by this work using deep learning, which can be further enhanced with a larger data set.