Automated Grading of Diabetic Foot Ulcer Severity Using Deep Learning on Medical Images

Diabetes mellitus is associated with a range of complications, among which Diabetic Foot Ulcers (DFUs) represent one of the most critical, often resulting in lower-limb amputation. Early and accurate identification of DFU severity can contribute to informed clinical decision-making and the development of effective treatment strategies. Given the need for objective and scalable DFU assessment, researchers have advanced toward automated, rapid, and reliable computational methods. In this study, we conduct an initial comparative analysis of state-of-the-art deep learning architectures to determine suitable models for DFU evaluation. Subsequently, we employ a two-stage pipeline to achieve comprehensive wound analysis. In the first stage, wound type identification is performed to classify images into ten categories, thereby isolating diabetic ulcers from non-diabetic cases. In the second stage, confirmed diabetic ulcers are further categorized into four clinically relevant severity levels. To ensure robust feature extraction, ConvNeXt V2 and GC-ViT architectures are utilized. Experimental results indicate that, in Stage 1, ConvNeXt V2 and GC-ViT attain classification accuracies of 96.68% and 96.33%, respectively, with ROC-AUC values approaching 1.0. In Stage 2, ConvNeXt V2 achieves an accuracy of 76.29%, while GC-ViT attains 77.66%, both demonstrating strong discriminative capability. The outcomes of this investigation show that the proposed system, incorporating advanced deep learning models across a multi-stage framework, offers reliable performance and exhibits potential for integration into clinical workflows aimed at early DFU diagnosis and improved patient management.