Thermography-based Breast Abnormality Detection using Pre-Trained RadImageNet Models

With breast cancer accounting for a large number of cancer-related deaths worldwide, there exists a need to develop tools and imaging modalities that aid in the early detection of breast abnormalities. While mammography is considered the standard imaging modality for breast cancer diagnosis, it cannot be used on pregnant women and is not as effective with denser breasts. Thermography, which is a non-invasive and radiation-free technique, is an adjunct tool to help detect breast abnormalities. Various pre-trained models have been applied to breast thermograms in the literature to detect breast abnormalities. This is a preliminary study that explores the use of pre-trained models trained on a medically sourced database like RadImageNet, in the analysis of breast thermograms from the Database for Mastology Research (DMR). Models trained from medically-sourced databases like RadImageNet have been known to extract features and patterns more relevant for medical imaging analysis and classification. In this study, the performances of RadImageNet models were compared with models trained on non-medically sourced databases, like ImageNet. Preliminary findings suggest that ImageNet models outperform their equivalent RadImageNet models. This preliminary study emphasizes the importance of using relevant and targeted datasets for the development of models that aid in breast abnormality detection.