Smart Diagnosis of Diabetic Foot Ulcers Using Convolutional Neural Networks on a Real-World Dataset

Abstract

Diabetic foot ulcers (DFUs) are a significant occurrence among diabetic sufferers that augment the danger of potential amputation and substantially advance hopefulness as a result of secondary infections. The paper presents a reliable DFU detection mechanism based on a sophisticated deep learning approach that incorporates a Convolutional Neural Network (CNN), which is expected to spot DFUs at their initial stages so that immediate intervention can be provided and reduce the problems associated with poor, timely detection. The proposed model minimizes feature extraction and enhances the levels of accuracy in predicting the classification of the images using a multi-branch CNN architecture with numerous convolutional layers. Using the base networks such as ResNet and DenseNet to transfer learning to obtain more accuracy and efficiency in our learning, we used transfer learning in particular. Moreover, we incorporated the model with a space attention block, which enables it to focus on targeted zones of the feet in infrared thermal images, which document alterations in temperature that are indicative of ulcer risk. The model was trained on a dataset containing more than 6000 labels of the annotated images of DFU, with an F1 greater than 97.12%. To enhance resilience and prevent overfitting, complex data augmentation techniques were applied, which ensure that the model would perform decently on a variety of patient demographics. To save hundreds of thousands of healthcare dollars spent on DFU complications and to better patient outcomes, this study tries to provide medical practitioners with a powerful, non-invasive means of early DFU detection.

Key words: Diabetic Foot Ulcer Detection, Deep Learning for Medical Imaging, Multi-branch CNN Architecture, Healthcare Cost Reduction through AI