Classification of Healthy and Diabetic Patients Using Plantar Foot Infrared Images

This study explores multimodal imaging and machine learning for classifying diabetic foot disease (DFD) and assessing risk. A total of 415 co-registered RGB and thermal (IR) image pairs from the STANDUP dataset were analyzed, including healthy controls and diabetic patients across three clinical stages (R0, R1, R2). Deep learning models (ResNet, LSNet) were trained using RGB, IR, and fused inputs, and evaluated under binary and multi-class classification. Foot shape masks were also used to assess the diagnostic value of morphological features. Temporal thermal modeling using cold stress test images (T0–T10) captured vascular reactivity. IR-based models outperformed RGB-only ones, achieving up to 99.5% balanced accuracy in binary classification. Hybrid models combining statistical features and CNN outputs improved multi-class performance. Surprisingly, models using only foot shape masks performed comparably well. Temporal dynamics improved classification, highlighting their diagnostic relevance. However, the absence of longitudinal data limited validation of thermal asymmetry metrics.