Low-rank distribution embedding of dynamic thermographic data for breast cancer detection.

Dynamic infrared thermography has emerged as a robust adjunctive modality for early-stage breast cancer screening, complementing clinical breast examination (CBE) through its capacity to capture physiologic alterations in breast tissue. By enabling the extraction of thermal biomarkers indicative of vasodilation, thermography provides a noninvasive means of detecting vascular and metabolic abnormalities. Heterogeneous thermal distributions often reflect angiogenesis-the neovascularization process widely recognized as a hallmark of malignant transformation. This study systematically evaluates computational frameworks for the detection and characterization of such heterogeneous thermal signatures. While low-rank approximation has shown utility in extracting vasodilatory features from temporal thermographic sequences, limitations remain in optimizing the integration of basis vectors. To address this, we examine cosine similarity, euclidean distance, and Canberra distance as advanced spatiotemporal metrics. Furthermore, triangular, Poisson, Weibull, and Chi-squared embedding techniques are introduced as innovative strategies to enhance discrimination of heterogeneous patterns. A deep neural network, augmented with histogram equalization (HE), is employed to refine these embeddings by accurately delineating regions of interest (ROI) and maximizing dynamic range. The proposed framework demonstrates strong diagnostic performance (accuracy: 83.6±5.5%, kappa = 0.31), underscoring its potential to advance computational thermography for early breast cancer detection and to provide clinically meaningful insights into vascular heterogeneity.