Spatiotemporal Deep Learning on Dynamic Infrared Thermography for Classification of Post-COVID-19 and Post-Myocardial Infarction Patients

Abstract

Dynamic infrared thermography is emerging as a noninvasive technique for monitoring microvascular health, yet its interpretation remains largely qualitative and labor-intensive. This work systematically benchmarks four deep learning architectures: 2D CNN, 3D CNN, CNN–LSTM, and CNN–Transformer, evaluated for automated DIRT sequence classification in a clinically relevant cohort of post-COVID-19 and post-myocardial infarction patients. The study introduces a rigorous pipeline encompassing thermal image acquisition, standardized preprocessing, tailored data augmentation, and stratified cross-validation to ensure reliable evaluation. Purely spatial models such as the 2D CNN underperform, achieving a macro F1 score of 73.5% and accuracy of 80.1%, while temporally aware models yield substantial gains: CNN–LSTM reaches a macro F1 score of 91.4% and accuracy of 92.7%, and the CNN–Transformer achieves 88.8% and 90.6% prior to hyperparameter optimization. After automated hyperparameter optimization, both models converge to a macro F1 score of 93.8% and accuracy of 94.8%, with the Transformer requiring less than half the parameters. Functional ANOVA analysis highlights that learning rate is the most influential factor for LSTM tuning, while dropout dominates for the Transformer. These findings establish a foundation for robust, sequence-aware DIRT analysis, demonstrating that modern deep learning models, when rigorously validated, can transform DIRT into a quantitative biomarker for longitudinal vascular assessment.