Diabetic Foot Ulcers Detection Model Using a Hybrid Convolutional Neural Networks-Vision Transformers

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Diabetology

Background:

• Diabetic foot ulcers (DFUs) are a serious diabetes complication requiring precise classification for effective management.
• Current DFU classification methods face challenges in performance, generalization, and interpretability, limiting clinical utility.

Purpose of the Study:

• To develop an innovative, robust, and interpretable model for classifying diabetic foot ulcer severity.
• To address the limitations of existing DFU classification techniques in clinical settings.

Main Methods:

• Integration of deep learning architectures (MobileNet V3-SWIN, LeViT-Peformer) with tensor-based feature fusion.
• Application of ensemble splines-based Kolmogorov-Arnold Networks (KANs) combined with Shapley Additive exPlanations (SHAP) for interpretability.
• Training and validation using the Diabetic Foot Ulcer Challenge (DFUC) 2021 and 2020 datasets.

Main Results:

• Achieved state-of-the-art performance on DFUC 2021 with 98.7% accuracy, 97.3% precision, 97.4% recall, and 97.3% F1-score.
• Demonstrated robust generalization on DFUC 2020 with 96.9% accuracy, outperforming baseline models.
• The model successfully classified DFUs into ischemia and infection categories with high performance and interpretability.

Conclusions:

• The proposed model offers a significant advancement in automated, explainable DFU classification.
• Findings support improved clinical decision-making, patient outcomes, and scalable management of diabetic foot complications.