Multiple Multi-Modal Methods of Malignant Mammogram Classification

Breast cancer is a tragic disease, which affects approximately 1 out of 8 women. Mammograms are X-ray scans used to detect early breast cancers, which are often susceptible to human error. As such, a strong emphasis has been placed on creating novel techniques to detect early cancers in mammogram scans. Analyzing and detecting breast cancer can be considered as a multi-modal task, which combines patient health information with their scan data. In this paper, we present a healthcare informatics solution-which consists of multiple multi-modal classification methods-and apply these methods in our healthcare informatics solution detect cancers in the Radiological Society of North America (RSNA) Breast Cancer Dataset. In this work, we perform preprocessing and describe novel architectures to classify raw mammogram scan data. Our evaluation results indicate that our destructive patching and embedding concatenation methods lead to relatively high accuracy scores and significantly faster convergence. Additionally, we compare and validate multiple methods for aggregating predictions made on multiple scans to determine whether a patient has cancer. Ultimately, we achieve a balanced accuracy of 70.2% on this task of classifying mammograms into benign (i.e., not harmful) or malignant (i.e., very virulent).