FOOTGRADE: A SYNERGISTIC MULTI-MODEL ENSEMBLE FOR THE AUTOMATED EARLY DETECTION AND FINE-GRAINED STRATIFICATION OF DIABETIC FOOT ULCERS

Abstract

The rising incidence of Diabetes Mellitus worldwide has made Diabetic Foot Ulcers (DFUs) a leading cause of non-traumatic amputation of lower limbs and this has necessitated the development of automated diagnostic accuracy. Although the current computer-aided systems have enhanced binary detection (ulcer vs. healthy) they often cannot provide the granular assessment of severity necessary to support successful clinical staging. In the proposed work, FootGrade, a new-fangled diagnostic framework based on the synergistic multi-model ensemble, is introduced, with its capability of realizing both early detection and fine-grained stratification of DFUs. FootGrade combats typical challenges to deep learning like overfitting through the local feature extractors of deep convolutional backbones (EfficientNet-V2) and the global context of Vision Transformers (ViT), which affects sensitive clinical lighting. A hierarchical grading logic to scale ulcers into a multi-stage scale of severity, providing risk-stratification information is proposed. Experimental tests on the DFUC2020 dataset show that the proposed ensemble is much better than the current state-of-the-art performance with an Accuracy of 96.8%, F1-score of 0.96 and AUC of 0.988. These findings validate the idea that FootGrade is a high-quality, scalable system of remote patient monitoring and clinical decision support, which can potentially enable the early-stage intervention and the targeted, accurate intervention to lower the amputation rates.