A Glimpse into the Future: Disease Progression Simulation for Breast Cancer in Mammograms

Simulating the disease progression of a suspicious finding in mammography (MG) images could assist in the early detection of breast cancer, where the telltale signs of malignancy are subtle and hard to detect. It could also decrease both unnecessary biopsies and treatments of indolent or low-grade disease that might otherwise remain asymptomatic. We propose a novel approach to simulating disease progression, as a significant step towards those goals. Our architecture uses the powerful Wasserstein GAN in combination with a novel component that simulates the progression of the disease in deep feature space. This allows us to learn from unlabeled longitudinal MG pairs of current and prior studies, stabilize the learning procedure, and overcome misalignment between the MG pairs. Our output image replicates an actual MG image, maintains the prior’s shape and general appearance while also containing a finding with characteristics that resemble the current image’s suspicious finding. We simulate a progression of: (i) a full MG prior image in low-resolution, and (ii) a high-resolution patch in suspicious areas of the prior image. We demonstrate the effectiveness of our pipeline in achieving the above goals using quantitative and qualitative metrics and a reader study. Our results show the high quality of our simulation and the promise it holds for early risk stratification.