Contactless skin temperature management for preterm infants : a multi-modal approach

Maintaining stable skin temperature is critical for the survival and healthy development of premature infants. Traditional methods that rely on wired skin sensors carry the risk of skin damage, infection and interference with parent-infant interaction. This thesis presents a novel, non-contact approach using Infrared Thermography to address these challenges, enabling continuous and accurate temperature monitoring without physical contact. For this purpose, an automated body part segmentation algorithm has been developed as a critical first step for accurate temperature extraction from infrared images. A dataset of multimodal RGB and long-wave infrared data was created and annotated. Deep learning models, specifically U-Net architectures, were trained and evaluated, demonstrating that multi-modal fusion approaches combining information from both RGB and LWIR images significantly outperform single-modality methods in segmenting body regions such as the head, torso, arms and legs. Accurate IRT-based temperature measurement within an incubator requires careful consideration of environmental influences such as IR window transmission and reflections. This work introduces a novel temperature correction algorithm that addresses these challenges and has been calibrated using a black body. This algorithm significantly reduced measurement errors, achieving a mean absolute error of 0.17 °C compared to the black body. A clinical study of premature infants validated the performance of the combined segmentation and temperature correction pipeline, demonstrating its accuracy and robustness in a real-world environment. In the clinical setting, the system achieved a mean absolute error of 0.41 °C compared to adhesive skin sensors. Furthermore, the torso and arms were found to be the most reliable regions for contactless temperature measurement due to their lower fluctuations and higher detectability. A novel neonatal thermal phantom was created to enable the development and validation of non-contact skin temperature control algorithms. This phantom, which incorporates a 3D-printed anatomical structure with independently controllable heating zones and embedded temperature sensors, accurately simulates the thermal behavior of a premature infant, allowing realistic testing of various physiological and pathological scenarios. Finally, a hardware-in-the-loop (HIL) test bed was designed, integrating the IRT system, segmentation models and the thermal phantom. PI-based control strategies were implemented and evaluated, demonstrating successful closed-loop control of infant skin temperature by regulating the incubator air temperature, while minimizing fluctuations and responding effectively to simulated perturbations and impaired thermoregulation. This research provides a foundation for the future development of fully contactless closed-loop skin temperature control systems, promising improved safety and comfort for premature infants.