Preventing Diabetic Foot Disability Through Dual-backbone Siamese Learning for Ulcer Risk Classification

Diabetic foot ulcers (DFUs) represent one of the most disabling complications of diabetes, frequently culminating in limb amputation, permanent mobility loss, and long-term physical disability when infection or ischaemia is not recognized early. Addressing this disability burden, the present study proposes a disability-focused deep learning framework for four-class DFU image classification, separating wounds into infection, ischaemia, both, and none. The method combines EfficientNetV2-S for local wound texture, DINOv2-Base for global visual context, and a Siamese metric-learning network trained with large margin cotangent loss and class-balanced focal loss. A dynamic sigmoid attention module fuses the two feature streams, and the final normalized embeddings are classified using eXtreme Gradient Boosting (XGBoost). Experiments were performed on the licensed Diabetic Foot Ulcer Challenge 2021 (DFUC2021) dataset, where targeted class-conditional augmentation was used to reduce severe imbalance. Compared with the k-nearest neighbour (KNN)-based baseline, the optimized XGBoost model improved overall accuracy from 64.16% to 65.57% and increased macro area under the receiver operating characteristic curve (AUC-ROC) from 0.7387 to 0.8257. The strongest gain was observed for ischaemia, reaching an AUC of 0.9324. These findings demonstrate that dual-backbone metric learning can support earlier DFU risk recognition, with the potential to support earlier identification of high-risk ulcers that may ultimately contribute to disability prevention and improved patient outcomes following prospective clinical validation.