Investigating Artificial Intelligence Performance on Mammographic Cases in High- and Low-Resourced Countries

Population-based screening in Australia for breast cancer using mammography has delivered key outcomes in reducing deaths, but such programs are usually not present in Vietnam. Breast cancer is common in both countries, with Vietnamese women having high breast density and Vietnam having low radiology expertise. This paper investigated the performance of two state-of-the-art Artificial Intelligence (AI) models for cancer detection on Datasets 1 and 2 (Vietnam and Australian mammographic cases) and determined whether the breast density in the two datasets with different vendors affected the AI performance. Both datasets consisted of the same number of mammographic cases (865 malignant; 865 normal). Mammographic images were enhanced using the contrast-limited adaptive histogram equalization algorithm. Transfer learning of the Globally-aware Multiple Instance Classifier (GMIC) and Global-Local Activation Maps (GLAM) AI models was conducted on enhanced images in the two national datasets. The breast density was classified as four levels (A, B, C, and D) by three radiologists. We combined Levels A and B into 0%-50%, and Levels C and D into 50%-100% breast density. The performance of these two AI models, both with and without transfer learning, was evaluated on the two datasets using McNemar's test. The GMIC in the pre-trained and transfer learning modes outperformed the GLAM in terms of specificity and sensitivity in the two datasets. The performance of the GMIC and GLAM with transfer learning in Dataset 2 was better, with 86.7% specificity and 88.2% sensitivity, outperforming their transfer learning model with 80.4% specificity and 81.2% sensitivity in Dataset 1. The specificity and sensitivity of the two AI models (p-values <0.05) were significantly improved using transfer learning in the two datasets. The specificity and sensitivity of GMIC with transfer learning (p-values <0.05) on the 0%-50% and 50%-100% breast density in Dataset 2 were significantly higher than those in Dataset 1, respectively.