Classification of mammograms: Comparing a graphical to a geometrical approach

Breast carcinoma is the second most common cause of cancer-related deaths. Radiologists often use mammography, a noninvasive and inexpensive imaging tool, for the detection and classification of breast cancer (BC) lesions. However, manual analysis is labor-intensive and prone to diagnostic errors. In this scenario, the large-scale deployment of computer-aided diagnosis using well-trained algorithms could significantly reduce the morbidity and mortality associated with this carcinoma. In this study, we used a similarity metric-based classification of mammograms using graphical (with two different image sizes) and geometrical approaches (with a single image size) for comparison to improve the specificity, sensitivity, and accuracy of BC prediction and triage of patients in the order of disease severity. Both classification techniques use two novel algorithms, hereafter referred to as the normal and hybrid methods, to select representative images from the training sets of healthy and unhealthy groups of mammograms. The normal method identifies a representative image by comparing images within a cohort, whereas the hybrid method adopts a comprehensive approach by comparing images from both cohorts. This study explored the effects of image size and cardinality of the training set. Finally, we explored the uncharted territory of mapping accuracy versus computational expense for the different approaches adopted in the current study.