Convex Factorization Embedding Thermography for Breast Cancer Diagnostic

Thermographic has proven to be effective for the early detection of breast cancer and with clinical breast examination (CBE). There are many matrix factorization methods developed for computational thermography that can be used to extract thermal variations across the acquisition time. These methods are often used to summarize thermographic sequences and simultaneously highlight predominant thermal patterns. Finding a single predominant infrared image capturing the prevalent patterns of changes remains a challenging task in the field. This study presents the applications of convex factor analysis combined with the bell-curve membership function embedding approach to tackle this task and generate one image to represent the entire sequence. This low-dimensional (LD) representation of a thermal sequence was then used to extract thermomics and train tuned hyperparameters random forest model for early breast cancer diagnosis. A comparative analysis of different embedding methods and factorization approaches is also provided. The results of the proposed method combining clinical information, and demographics yield 78.9% (75.7% and 85.9%), while the convex-nonnegative matrix factorization (NMF) alone gave 76.9% (73.7% and 86.1%). The result of the proposed method suggests that the embedding can help preserve important thermal patterns, which significantly aid CBE and early detection of breast cancer.