Deep learning-based automated segmentation of athlete thermograms for real-time lower-limb heat response monitoring during treadmill running

Infrared thermography measures skin-surface temperature and dry heat exchange without contact, but continuous analysis during locomotion requires outlining moving anatomical regions in every frame. This study built an automated measurement chain that converts lower-limb thermograms into continuously updated thermophysical descriptors. RT-ThermoSeg, a compact segmentation network with thermal-gradient attention, was tested on the subject-independent split of the public StereoThermoLegs dataset (16,259 thermograms from 14 participants during an incremental treadmill test at 4–12 km h −1 ) against five segmentation architectures trained under the same protocol; all segmentation, thermometric and physiological results refer to the three held-out test participants ( n = 3). RT-ThermoSeg reached a dataset-level mean intersection over union of 0.798 (pooled over all test frames; frame-wise mean 0.756) with 1.03 million parameters and 0.314 giga multiply–accumulate operations (GMAC) per frame, comparable to a retrained DeepLabv3+ (0.792) at about one fifth of its arithmetic cost. The full acquisition-to-descriptor pipeline ran at 42.6 frames s −1 , above the 15 frames s −1 acquisition rate. For the whole lower limb, temperatures from predicted masks agreed closely with those from reference masks: bias −0.056 °C, limits of agreement −0.278 to 0.166 °C, and an intraclass correlation coefficient of 0.989. Across the reconstructed exercise stages, mean limb-surface temperature fell by 0.83 °C and dry heat-flux density fell from 63.6 to 55.7 W m −2 under an assumed still-air environment at 22 °C. Post-exercise temperature recovery followed first-order kinetics with a whole-limb time constant of 139 s. The experimental outcomes indicate that automated thermogram segmentation can support real-time, region-resolved thermal monitoring during treadmill exercise, and that segmentation performance can be expressed in physically interpretable temperature and heat-transfer quantities.