Unpaired RGB-to-thermal learning for explainable diabetic-foot risk screening

Researchers have developed a novel approach to screen for diabetic foot ulcers using a combination of RGB and thermal imaging. Traditional infrared cameras are expensive and not commonly accessible at home, but smartphones can capture regular photographs that can be converted into thermal images. The study, published in Scientific Reports, introduces an unpaired cross-modal framework that translates RGB DFU photographs into thermal data using a CycleGAN backbone.

This framework incorporates structural-consistency constraints and anatomical priors to maintain the accuracy of toe, metatarsal, arch, and heel geometry. A thermal-domain anomaly detector, EfficientNet-B0 with a feature pyramid, identifies thermal abnormalities and projects its findings back onto the original RGB images through a cross-domain attention back-projection module.

This process generates interpretable thermal risk maps, which are supervised by weak labels derived from thermal asymmetry and adaptive thresholding. The model's performance is robust, achieving high cross-domain fidelity (SSIM = 0.82, LPIPS = 0.14) and clinically useful detection of thermal abnormalities (AUROC = 0.91, AUPRC = 0.88).

The localized RGB-domain predictions show high accuracy (IoU = 0.76; Top-1 hit rate = 87%), focusing on interdigital and metatarsophalangeal regions where pressure-related ulcers often develop. The thermal risk maps demonstrate physiological consistency, with predicted temperature rise closely matching clinical improvements in foot temperature.

Furthermore, a decline in the model-predicted risk area correlates significantly with vascular recovery (ABI/TBI; r = − 0.68, p 0.01). The researchers conducted ablations to assess the importance of each component, finding that removing structural loss or anatomical priors negatively impacts geometry and clinical alignment. Conversely, eliminating cross-domain attention causes the most significant drop in performance (AUROC = 0.80, IoU = 0.64, correlation = 0.32), highlighting its necessity for semantic alignment and interpretability.

The study is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, allowing non-commercial use and sharing of the research.