AI-Powered Early Detection of Diabetic Foot Ulcers: Integrating Deep Learning and Clinical Insights

Abstract

Diabetic Foot Ulcers (DFUs) are among the most severe complications of diabetes and can lead to infections, amputations, and increased mortality if not detected early. This study presents an AI-based multiclass DFU detection system using a convolutional neural network built upon the EfficientNetB5 architecture to classify foot images into four clinically relevant categories: no ulcer, immediately treatable, treatable within four weeks, and complex wounds. A curated dataset of 6,247 foot images was constructed from public and clinical sources through expert-guided manual filtering and annotation. Comprehensive preprocessing, including noise removal, normalization, resizing, and targeted data augmentation, was applied to address image quality issues and class imbalance. The proposed model achieved an overall accuracy of 80.35%, with a macro-averaged precision of 81.2%, recall of 78.6%, and F1-score of 79.7%, demonstrating balanced performance across all classes. The system is further integrated into a mobile health application to support early DFU screening, offering a scalable and low-cost solution for resource-constrained healthcare settings.