Style Randomization Improves the Robustness of Breast Density Estimation in MR Images

Goksenin Yuksel, Koen Eppenhof, Jaap Kroes, Marcel Worring

Keywords: Breast Density Estimation, MRI, Style Randomization, Deep Learning, FGT Segmentation, Robustness, Representation Learning, Dixon Images

Abstract:

Breast density, a crucial risk factor for future breast cancer development, is defined by
the ratio of fat to fibro-glandular tissue (FGT) in the breast. Accurate breast and FGT
segmentation is essential for robust density estimation. Previous research on FGT segmen-
tation in MRI has highlighted the significance of training on both images with and without
fat suppression to enhance network’s robustness. In this study, we propose a novel data
augmentation technique to further exploit the multi-modal training setup motivated by the
research in style randomization. We demonstrate that the network trained with the pro-
posed augmentation is resilient to variations in fat content, showcasing improved robustness
compared to solely training with multi-modal data. Our method effectively improves FGT
segmentation, thereby enhancing the overall reliability of breast density estimation

Submission Number: 95