Artificial Intelligence in Mammography: A Study of Diagnostic Accuracy and Efficiency

Breast cancer continues to be a considerable global health problem, highlighting the need for early and accurate diagnosis to improve patient outcomes. Although mammography is widely considered the gold standard for screening, its interpretation is not straightforward and varies among readers. Our study aimed to compare the performance and computational efficiency of three leading Convolutional Neural Network (CNN) architectures for classifying breast cancer automatically from mammogram images. We used a publicly available dataset consisting of 3,383 mammogram images, which were labeled as either Benign or Malignant, and we trained and evaluated three models: EfficientNetB7, EfficientNetv2-Small, and RexNet-200. We found the RexNet-200 architecture had the best performance across the performance metrics we measured, achieving the best accuracy (76.47%), precision (75.18%), and F1-score (77.44%). Even though EfficientNetB7 had a slightly better recall than the RexNet-200 model; the RexNet-200 model showed a more compelling accuracy-board balance in diagnosis. Furthermore, RexNet-200 had the best performance and lowest computational cost with a very low parameters count (13.81M) and lowest GFLOPS (3.0529) of the three models. Our study demonstrated that RexNet-200 had the best prospects for achieving the ideal balance of high diagnostic accuracy and economical use of resources. Therefore, RexNet-200 is a very promising candidate for incorporation into clinical decision support systems designed to assist radiologists in the early detection of breast cancer.