Towards Exercise Radiomics: Deep Neural Network-Based Automatic Analysis of Thermal Images Captured During Exercise

Infrared thermography is increasingly applied in sports science due to promising observations regarding changes in skin’s surface radiation temperature ( $T_{sr}$ ) before, during, and after exercise. The common manual thermogram analysis limits an objective and reproducible measurement of $T_{sr}$ . Previous analysis approaches depend on expert knowledge and have not been applied during movement. We aimed to develop a deep neural network (DNN) capable of automatically and objectively segmenting body parts, recognizing blood vessel-associated $T_{sr}$ distributions, and continuously measuring $T_{sr}$ during exercise. We conducted 38 cardiopulmonary exercise tests on a treadmill. We developed two DNNs: body part network and vessel network, to perform semantic segmentation of 1 107 855 thermal images. Both DNNs were trained with 263 training and 75 validation images. Additionally, we compare the results of a common manual thermogram analysis with these of the DNNs. Performance analysis identified a mean IoU of 0.8 for body part network and 0.6 for vessel network. There is a high agreement between manual and automatic analysis (r = 0.999; p $< $ 0.001; T-test: p = 0.116), with a mean difference of 0.01 $^\circ$ C (0.08). Non-parametric Bland Altman’s analysis showed that the 95% agreement ranges between - 0.086 $^\circ$ C and 0.228 $^\circ$ C. The developed DNNs enable automatic, objective, and continuous measurement of $T_{sr}$ and recognition of blood vessel-associated $T_{sr}$ distributions in resting and moving legs. Hence, the DNNs surpass previous algorithms by eliminating manual region of interest selection and form the currently needed foundation to extensively investigate $T_{sr}$ distributions related to non-invasive diagnostics of (patho-)physiological traits in means of exercise radiomics.