Thermal diffusion–based subsurface Depth Estimation in Active Infrared Thermography Using Thermal Signal Reconstruction and Deep Learning

Breast cancer remains one of the most prevalent malignancies worldwide, with early detection being critical for improved patient outcomes. Infrared thermography offers a non-invasive, radiation-free technique to monitor physiological changes associated with malignancies, capturing subtle heat variations on the skin surface. In this study, we present an integrated approach combining Thermographic Signal Reconstruction (TSR), automatic region-of-interest (ROI) detection, and deep learning for precise depth estimation in tissue-mimicking phantoms embedded with resistive heat sources. Variable-frame thermal sequences are acquired using a long-wave infrared camera, and TSR is applied with shape-safe reconstruction to model surface temperature decay. Automatic ROI selection identifies regions exhibiting maximum thermal contrast, and second-derivative analysis with sub-frame quadratic interpolation extracts peak-time signals. Peak-time versus depth is analyzed using log–log regression weighted by signal-to-noise ratio (SNR). To further enhance prediction accuracy, a lightweight 3D convolutional neural network (3D-CNN) is trained on TSR-processed sequences, and Bayesian fusion of TSR and CNN outputs yields final depth estimates. The method achieves over 97% accuracy across multiple phantom depths, with Bland–Altman analysis confirming high agreement between predicted and true depths. Multi-depth decay curves, log–log peak-time plots, and confidence intervals demonstrate robust performance across variable thermal sequences. This pipeline provides a reliable framework for early breast cancer detection in a controlled phantom setting, and the methodology can be extended to clinical thermography applications, offering a rapid, non-contact, and interpretable diagnostic tool.