Abstract
Breast cancer remains the most prevalent cancer among women worldwide, emphasizing the demand for accessible and accurate screening technologies. Infrared thermography offers a noninvasive and radiation-free alternative; however, automated segmentation remains challenging due to low contrast, noise, and high intersubject variability. Existing approaches—from classical computer-vision pipelines to advanced deep networks such as UNet, SegNet, YOLOv8-Seg, TransUNet, and Dense Multiscale UNet—often depend on large annotated datasets and lack anatomical constraints, leading to unstable boundaries and inconsistent thermal quantification. We propose a hybrid UNet that integrates thermographic anatomical landmarks as spatial priors, combined with targeted geometric and spectral augmentation to enhance robustness to variations in anatomy, sensor calibration, and acquisition protocols. This design enforces anatomically plausible breast contours and minimizes dependence on extensive manual labeling. Validated on an independent held-out test set with bootstrap-based confidence estimation, the proposed model achieved DSC = 0.988 [Formula: see text] 0.004, IoU = 0.958 [Formula: see text] 0.006, AUC = 0.995 [Formula: see text] 0.003, SEN = 0.990 [Formula: see text] 0.004, SPC = 0.995 [Formula: see text] 0.002, and NHD95 = 0.012 [Formula: see text] 0.002. These results surpass all compared methods, including Dense Multiscale UNet (DSC = 0.977 [Formula: see text] 0.009) and TransUNet (DSC = 0.979 [Formula: see text] 0.010), demonstrating superior boundary precision and stability under bootstrap resampling ([Formula: see text]0.002–[Formula: see text]0.017). By coupling data-driven learning with explicit anatomical priors and controlled augmentation, this framework advances thermography-based breast segmentation toward clinically reliable, anatomy-aware, and statistically reproducible workflows for early breast cancer screening.