High-resolution synthesis of high-density breast mammograms: Application to improved fairness in deep learning based mass detection

Abstract

Computer-aided detection systems based on deep learning have shown goodperformance in breast cancer detection. However, high-density breasts showpoorer detection performance since dense tissues can mask or even simulatemasses. Therefore, the sensitivity of mammography for breast cancer detectioncan be reduced by more than 20% in dense breasts. Additionally, extremelydense cases reported an increased risk of cancer compared to low-densitybreasts. This study aims to improve the mass detection performance in highdensitybreasts using synthetic high-density full-field digital mammograms(FFDM) as data augmentation during breast mass detection model training. Tothis end, a total of five cycle-consistent GAN (CycleGAN) models using threeFFDM datasets were trained for low-to-high-density image translation in highresolutionmammograms. The training images were split by breast density <em>BIRADS</em>categories, being <em>BI-RADS A </em>almost entirely fatty and <em>BI-RADS D</em>extremely dense breasts. Our results showed that the proposed dataaugmentation technique improved the sensitivity and precision of massdetection in models trained with small datasets and improved the domaingeneralization of the models trained with large databases. In addition, theclinical realism of the synthetic images was evaluated in a reader studyinvolving two expert radiologists and one surgical oncologist.