Fig. 3

Overview of GAN-based anomaly detection and future lesion emergence prediction. a Tissue appearance change in negative follow-up CE-MRI scans of the breast that do not develop into breast cancer (normal data) forms a distribution in the space of different images that can be learned by a GAN model from normal cases. Anomalies are observations deviating from this distribution. b During training, a Generator G, a discriminator D, and an encoder E are trained on normal data. c During anomaly scoring a new image patch is processed by E, G, and D, resulting in an anomaly score. An anomaly score map for an entire CE-MRI volume is composed of anomaly scores evaluated in a sliding window across the entire volume. CE-MRI Contrast-enhanced magnetic resonance imaging, GAN Generative adversarial network