An anomaly detection model for breast cancer on MRI scans achieved high performance, surpassing binary classification in high- and low-cancer-prevalence scenarios and producing explanation heat maps not significantly different from radiologist annotations.