Distribution-based Low-rank Embedding
The early detection of breast abnormalities is a matter of critical significance. Notably, infrared thermography has emerged as a valuable tool in breast cancer screening and clinical breast examination (CBE). Measuring heterogeneous thermal patterns is the key to incorporating computational dynamic thermography, which can be achieved by matrix factorization techniques. These approaches focus on extracting the predominant thermal patterns from the entire thermal sequence. Yet, the task of singling out the dominant image that effectively represents the prevailing temporal changes remains a challenging pursuit within the field of computational thermography. In this context, we propose applying James-Stein for eigenvector (JSE) and Weibull embedding approaches, as two novel strategies in response to this challenge. The primary objective is to create a low-dimensional (LD) representation of the thermal data stream. This LD approximation serves as the foundation for extracting thermomics and training a classification model with optimized hyperparameters, for early breast cancer detection. Furthermore, we conduct a comparative analysis of various embedding adjuncts to matrix factorization methods. The results of the proposed method indicate an enhancement in the projection of the predominant basis vector, yielding classification accuracy of 81.7% (+/-5.2%) using Weibull embedding, which outperformed other embedding approaches we proposed previously. In comparison analysis, Sparse PCT and Deep SemiNMF showed the highest accuracies having 80.9% and 78.6%, respectively. These findings suggest that JSE and Weibull embedding techniques substantially help preserve crucial thermal patterns as a biomarker leading to improved CBE and enabling the very early detection of breast cancer.