Bio-Inspired DFU-ThermoNet: A Spiking Transformer-Driven Framework for Early Diagnosis and Risk Prediction of Diabetic Foot Ulcers Using Thermal Imaging

Diabetic Foot Ulcers (DFUs) are among the most serious problems for people with diabetes, often leading to infection and amputation. Early identification of risk levels can prevent severe complications. DFU risk evaluation using foot thermal imaging has gained attention with recent deep learning (DL) advances. Current practices focus on subjective clinical evaluation, which relies on subjective clinical assessment and manual visual inspection, often causing delays in diagnosis. We propose DFU-ThermoNet, a bio-inspired DL framework for early DFU diagnosis and risk prediction using plantar thermal images. The model combines a Spiking Transformer Block (STB) for temporal pattern recognition, a Multi-Scale Capsule Network (MCN) for preserving spatial hierarchies, and a Dynamic Attention Fusion (DAF) mechanism to effectively merge features from different modules. A self-attention convolutional layer further enhances the discrimination of local features. Before training, the thermal images underwent a structured preprocessing pipeline including Thermal Change Index (TCI) calculation, risk-level classification into five categories— normal, mild, moderate, high, and critical risk—and class-balanced augmentation with perceptual hashing to prevent duplication. Training and testing were conducted using a subject-wise 80-20% split. This integrated strategy ensures that DFU-ThermoNet captures fine-grained spatial details and temporal changes critical for accurate DFU risk classification. DFU-ThermoNet was trained and tested on the Plantar Thermogram Database (PTD), which consists of 6000 thermal foot images, and the following measures were attained. Classification accuracy: 98.67%; Precision: 0.99; Recall: 0.99; F1 score: 0.99; and ROC-AUC score: 0.99 across the five risk categories, subject-wise 5-fold cross-validation confirmed consistent performance (97.84 % ± 0.53 %) with 95% confidence interval (CI), demonstrating the dependability of the proposed framework. DFU-ThermoNet provides an accurate, non-invasive assessment for early DFU risk detection compared with the state-of-the-art models.