Open-Source Software for Thermographic Analysis of the Human Hand Using Yolo Instance Segmentation

Introduction: This note presents the concept and implementation of dedicated software for the automated analysis of thermograms of the skin surface of the human hand. Standard tools supplied by thermal camera manufacturers often restrict access to raw radiometric data, hamper batch data processing, and force reliance on closed ecosystems. Method: In response to these limitations, a custom data processing chain was developed. The process comprises the extraction of raw radiometric matrices and calibration parameters from the EXIF metadata of JPEG files, followed by the conversion of the radiometric signal into temperatures using a formula that constitutes an algebraic inversion of the simplified Planck model. The key component of the system is a YOLOv8n-seg instance segmentation model, initialized with weights pretrained on the COCO dataset and fine-tuned on a set of 52 precisely labeled thermal images (prepared in Label Studio) covering hands in three thermal states: after warming, at room temperature, and after cooling. The use of a model based on deep convolutional neural networks (CNNs) made it possible to isolate the hand region from the background automatically and to extract precise temperature statistics exclusively from pixels within the segmentation mask. Results: The trained model achieved very high segmentation precision (mAP50 = 0.995 and mAP50-95 = 0.944). A demonstration comparison of two measurement sessions of the same person revealed a difference in mean hand surface temperature of −7.4°C between the hand at rest and the cooled hand, while the model maintained high detection confidence (0.974 and 0.964) in both cases. Conclusions: The demonstration results indicate the effectiveness of the method in detecting skin-temperature differences between measurement sessions, providing a basis for further validation of the software in physiological studies.