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

Area of Science:

• Sports Science
• Medical Imaging
• Artificial Intelligence

Background:

• Infrared thermography (IRT) shows potential in sports science for monitoring skin surface radiation temperature (Tsr).
• Manual analysis of thermograms is subjective, lacks reproducibility, and is impractical during exercise.
• Existing methods require expert knowledge and are not suitable for dynamic movement analysis.

Purpose of the Study:

• To develop deep neural networks (DNNs) for automated, objective segmentation of body parts in thermal images.
• To enable continuous and precise measurement of Tsr during exercise, including recognition of blood vessel patterns.
• To establish a foundation for advanced exercise radiomics and non-invasive physiological diagnostics.

Main Methods:

• Development of two DNNs: a body part network and a vessel network for semantic segmentation of thermal images.
• Training and validation of DNNs using a dataset of 1,107,855 thermal images from 38 cardiopulmonary exercise tests.
• Comparison of DNN-based analysis with traditional manual thermogram analysis.

Main Results:

• High performance metrics for DNNs: mean IoU of 0.8 for body part segmentation and 0.6 for vessel segmentation.
• Excellent agreement between manual and DNN-based Tsr measurements (r = 0.999, p < 0.001), with minimal mean difference (0.01 °C).
• Bland Altman analysis confirmed high agreement within 95% limits (-0.086 °C to 0.228 °C).

Conclusions:

• The developed DNNs provide an automated, objective, and continuous method for Tsr measurement during exercise.
• These DNNs eliminate the need for manual region selection, surpassing previous analytical approaches.
• This technology supports extensive investigation of Tsr distributions for non-invasive diagnostics in exercise radiomics.