Abstract
Pressure injuries remain a critical concern in clinical care, with early detection essential for preventing progression and reducing morbidity. While thermal imaging has demonstrated promise for early pressure injury detection, the impact of imaging protocol variations and patient skin tone on detection accuracy remains underexplored. In this study, we systematically assess how variations in lighting, camera distance, patient positioning, and camera type influence deep learning model performance for early pressure injury detection using both optical and thermal images. A total of 1680 images were collected from 35 healthy adults across diverse skin tones using a factorial design of 12 imaging protocols in a controlled environment where localized cooling was induced to simulate temperature changes. Three deep learning model architectures (MobileNetV2, InceptionNetV3, ResNet50) were evaluated to assess protocol robustness. Thermal imaging significantly outperformed optical imaging, achieving >90% accuracy across models with minimal sensitivity to protocol variations. In contrast, optical performance varied substantially across protocols (32-55% accuracy), demonstrating significant protocol dependency that could impact clinical implementation. Cross-subject error analysis revealed that models focused on both the cool and warm regions in the images, suggesting that current static labeling approaches may be inadequate for dynamic thermal imaging applications. These findings establish the robustness of deep learning models trained on thermal imaging data across diverse skin tones and imaging conditions, providing a critical foundation for future clinical validation in pressure injury detection applications.