Methodology for the diagnosis of chronic skin ulcer infections through thermal imaging and clinical trial

Chronic skin ulcers affect about 2% of the global population and represent a major clinical and economic challenge for healthcare systems worldwide. Their management entails prolonged treatments, frequent hospital visits, and high costs, while patients often experience pain, reduced mobility, and psychological distress. Infection further worsens these outcomes, leading to serious complications such as localized tissue necrosis, osteomyelitis, necrotizing fasciitis, sepsis, and septic shock. When untreated, infections can transform a chronic wound into a life-threatening systemic condition, highlighting the urgent need for early detection and timely intervention. Although clinical, microbiological, and imaging-based methods allow pathogen identification and functional assessment, they often suffer from limitations in invasiveness, accessibility, cost, or processing time. Imaging approaches, including X-ray, MRI, ultrasound, and fluorescence techniques, provide structural and functional insights but are often costly, operator-dependent, or limited to specialized centers. In routine practice, empirical evaluation frequently guides diagnosis. Thermography, in contrast, is a non-invasive, portable, and cost-effective technique capable of detecting perfusion alterations and early inflammatory signs, providing real-time insights into wound status. Its ease of use makes it particularly suitable for primary care, outpatient, and home-monitoring contexts, where rapid, contactless, and operator-independent assessments are essential. This thesis combines clinical and computational components developed in collaboration with Niguarda Hospital in Milan. The clinical component involved the collection of RGB and thermal images of chronic wounds under standardized conditions. Ground truth labeling relied on tissue biopsies and, in some cases, the C&H clinical scale to assess the presence of infection. Building upon this dataset, the research focused on data processing, feature extraction, and the development of classification models for infection assessment. Gradient-based, RGB-derived, and manually annotated parameters, including the Wound Bed Preparation (WBP) scale, were integrated to train and compare machine learning and deep learning models, such as CatBoost, XGBoost, Random Forest, SVM, k-NN, logistic regression, and YOLOv11. Model performance was evaluated using balanced accuracy, precision, recall, and F1 score, providing robust validation of the predictive framework. Results indicate that thermal imaging combined with computational models enables accurate, rapid, and non-invasive assessment of infection in chronic wounds. The analyses produced good and promising results, with a level of confidence regarding the actual presence of infection that is clinically satisfactory. These findings support the potential to establish an effective methodology for infection diagnosis capable of aiding clinical decision-making. This work represents a step forward toward AI-driven wound management, facilitating timely interventions, reducing reliance on invasive procedures, and improving patient outcomes. Furthermore, it lays a solid foundation for the integration of non-invasive diagnostic technologies into routine wound care and telemedicine, particularly where conventional approaches are limited or delayed.