Anomaly Detection Model Surpasses Binary Classification Model in Cancer Detection Performance

Background: MRI is highly sensitive for detecting breast cancer, especially in high-risk groups, but MRI screening can often result in false positives. It is hoped that recent advances in artificial intelligence (AI) applications in breast MRI could reduce radiologist workload and false positives. Deep learning models have been proposed for MRI cancer detection and triage but have performance limitations (binary classification) due to small number of positive cases (even in high-risk groups) and due to the heterogeneity of malignant lesions. Prior breast MRI AI models have been evaluated using balanced datasets of cancer and noncancer patients leading to overestimation of predictive power, which can limit clinical utility. An additional limitation is that heat maps generated by AI programs are often difficult to interpret due to lack of spatial annotation making exact localization of malignant lesions challenging. Objective: To develop and evaluate an AI model for cancer detection us more...