Breast density, or the amount of fibroglandular tissue (FGT) present relative overall breast volume, has been shown in multiple studies to increase the risk of developing breast cancer. Breast density is typically assessed qualitatively by a radiologist or semi-automatically using quantitative tools. However, both methods are inaccurate and suffer from inter-user variability. Although studies have been done to utilize deep learning methods to assess breast density, the lack of a publicly available dataset hinders the development of better assessment tools. The objective of this study was to both develop and evaluate an automated segmentation method for breast, FGT, and blood vessels using convolutional neural networks (CNN) and to publicly share full magnetic resonance imaging (MRI) volume annotation masks for breast, FGT, and blood vessels.
We used the Duke Breast Cancer MRI dataset to randomly select 100 fat-saturated gradient echo T1-weighted pre-contrast MRI studies. For each MRI volume, manual segmentation of breast, FGT, and blood vessels was completed and reviewed by a board-certified radiologist. We split the 100 patients into a training set (n=70), validation set (n=15), and testing set (n=15). We used the training and validation set for model development and the testing set for final model evaluation. The data was preprocessed with intensity normalization. Two CNNs were developed. The first model used MRI data to output a breast mask. The second model utilized the first model’s breast mask prediction in addition to the MRI data to segment FGT and blood vessels. We tested various methods of data input into the models, including 2D inputs, 3D random sub-volume sampling, and 3D fixed-stride sub-volume sampling. Model performance was evaluated using the Dice similarity coefficient (DSC).
All current results are using the validation set as the testing set will not be used until model development is complete. For the 2D data inputs, the DSCs were 0.90 for the breast, 0.75 for the FGT, and 0.35 for the blood vessels. For the 3D random sub-volume inputs, the DSCs were 0.87 for the breast, 0.84 for the FGT, and 0.58 for the blood vessels. For the 3D fixed-stride sub-volume inputs, the DSCs were 0.84 for the breast, 0.85 for the FGT, and 0.59 for the blood vessels.
A dual 3D CNN processing pipeline is able to accurately segment breast, FGT, and blood vessels using pre-contrast breast MRI data. All data and code used was made publicly available for others to use and iterate on.