Unpaired RGB-to-thermal learning for explainable diabetic-foot risk screening
- 1School of Clinical Medicine, Shandong University
- 2Shengli Oilfield Central Hospital
Abstract
Diabetic foot ulcer (DFU) screening benefits from thermography, but infrared cameras are costly and rarely available at home. We present an unpaired cross-modal framework that virtualizes thermal imaging from ordinary smartphone photos. Using only public Kaggle data (RGB DFU photographs and a thermography foot dataset), Stage A performs RGB↔thermal translation with a CycleGAN backbone enhanced by structural-consistency constraints and anatomical priors to preserve toe, metatarsal, arch, and heel geometry. Stage B trains a thermal-domain anomaly detector (EfficientNet-B0 with a feature pyramid) and projects its responses back to RGB via a cross-domain attention back-projection module, yielding interpretable thermal risk maps. Weak labels derived from thermal asymmetry and adaptive thresholding supervise a lightweight segmentation head; uncertainty gating filters noisy pseudo-masks. In experiments (PyTorch 2.1; dual RTX 3090s), the model achieves strong cross-domain fidelity (SSIM = 0.82, LPIPS = 0.14) and clinically usable detection of thermal abnormalities (AUROC = 0.91, AUPRC = 0.88). RGB-domain localization is accurate (IoU = 0.76; Top-1 hit rate = 87%), highlighting interdigital and metatarsophalangeal regions commonly implicated in pressure-related ulceration. Physiological consistency is demonstrated by close agreement between predicted temperature rise and clinical improvement (+ 4.3 °C vs. +4.5 °C dorsal foot temperature) and by a significant association between declining model-predicted risk area and vascular recovery (ABI/TBI; r = − 0.68, p < 0.01). Ablations confirm the role of each component: removing structural loss or anatomical priors degrades geometry and clinical alignment, while removing cross-domain attention causes the largest drop (AUROC = 0.80, IoU = 0.64, correlation = 0.32), underscoring its necessity for semantic alignment and interpretability.
Related
- Biomechanical and movement-quality predictors of lower-limb load-related injury in collegiate athletes: a prospective cohort study using functional movement screening, wearable inertial sensing, and explainable machine learning
- A Robust and Trustable Approach to Incorporate Medical Domain Knowledge in Machine Learning Models for Diabetic Retinopathy Screening Using Routine Lab Results
- Development and temporal validation of an explainable machine learning model for predicting in-hospital diabetic foot ulcer in patients with type 2 diabetes using inflammatory–nutritional composite indices
- Identification of key clinical features and development of a machine learning model for screening-based detection of diabetic retinopathy
- An explainable artificial intelligence system for prognostic assessment and decision support in diabetic foot ulcer care: a multicentre derivation and validation study with prospective randomised evaluation
BibTeX
@article{sui2026unpaired,
title = {{Unpaired RGB-to-thermal learning for explainable diabetic-foot risk screening}},
author = {Ping Sui and Ganchun Wu and Juan Xu and Shouguang Sui},
journal = {Scientific Reports},
year = {2026},
month = aug,
doi = {10.1038/s41598-026-61426-5},
url = {https://doi.org/10.1038/s41598-026-61426-5},
note = {CC BY},
}
RIS
TY - JOUR
TI - Unpaired RGB-to-thermal learning for explainable diabetic-foot risk screening
AU - Ping Sui
AU - Ganchun Wu
AU - Juan Xu
AU - Shouguang Sui
PY - 2026
DA - 2026/08/01
JO - Scientific Reports
DO - 10.1038/s41598-026-61426-5
UR - https://doi.org/10.1038/s41598-026-61426-5
AB - Diabetic foot ulcer (DFU) screening benefits from thermography, but infrared cameras are costly and rarely available at home. We present an unpaired cross-modal framework that virtualizes thermal imaging from ordinary smartphone photos. Using only public Kaggle data (RGB DFU photographs and a thermography foot dataset), Stage A performs RGB↔thermal translation with a CycleGAN backbone enhanced by structural-consistency constraints and anatomical priors to preserve toe, metatarsal, arch, and heel geometry. Stage B trains a thermal-domain anomaly detector (EfficientNet-B0 with a feature pyramid) and projects its responses back to RGB via a cross-domain attention back-projection module, yielding interpretable thermal risk maps. Weak labels derived from thermal asymmetry and adaptive thresholding supervise a lightweight segmentation head; uncertainty gating filters noisy pseudo-masks. In experiments (PyTorch 2.1; dual RTX 3090s), the model achieves strong cross-domain fidelity (SSIM = 0.82, LPIPS = 0.14) and clinically usable detection of thermal abnormalities (AUROC = 0.91, AUPRC = 0.88). RGB-domain localization is accurate (IoU = 0.76; Top-1 hit rate = 87%), highlighting interdigital and metatarsophalangeal regions commonly implicated in pressure-related ulceration. Physiological consistency is demonstrated by close agreement between predicted temperature rise and clinical improvement (+ 4.3 °C vs. +4.5 °C dorsal foot temperature) and by a significant association between declining model-predicted risk area and vascular recovery (ABI/TBI; r = − 0.68, p < 0.01). Ablations confirm the role of each component: removing structural loss or anatomical priors degrades geometry and clinical alignment, while removing cross-domain attention causes the largest drop (AUROC = 0.80, IoU = 0.64, correlation = 0.32), underscoring its necessity for semantic alignment and interpretability.
SN - 2045-2322
ER -