Improving breast cancer diagnostics with artificial intelligence for MRI

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has a very high sensitivity in detecting breast cancer, but it often leads to unnecessary biopsies and patient workup. In this paper, we used an artificial intelligence (AI) system to improve the overall accuracy of breast cancer diagnosis and personalize management of patients undergoing DCE-MRI. On the internal test set (N=3,936 exams), our system achieved an area under the receiver operating characteristic curve (AUROC) of 0.92 (95% CI: 0.92-0.93). In a retrospective reader study, there was no statistically significant difference between 5 board-certified breast radiologists and the AI system (mean {Delta}AUROC +0.04 in favor of the AI system). Radiologists' performance improved when their predictions were averaged with AI's predictions (mean {Delta}AUPRC [area under the precision-recall curve] +0.07). Those hybrid predictions also increase interreader agreement (Fleiss' kappa {Delta} +0.21 (0.16-0.26)). We demonstrated the generalizability of the AI system using multiple data sets from Poland and the US. In subgroup analysis, we observed consistent results across different cancer subtypes and patient demographics. Using the decision curve analysis, we showed that the AI system can reduce unnecessary biopsies in the range of clinically relevant risk thresholds. This would lead to avoiding benign biopsies in up to 20% of all BI-RADS category 4 patients. Finally, we performed an error analysis, investigating situations where AI predictions were mostly incorrect. This exploratory work creates a foundation for deployment and prospective analysis of AI-based models for breast MRI.