CNN-based Whole Breast Segmentation in Longitudinal High-risk MRI Study

Segmentation of breast MRI in a longitudinal high-risk cohort is challenging due to considerable variability in image quality and artefacts, imaging protocols, scanner types and field strengths, and not least anatomy. Automated segmentation of the whole breast is a prerequisite for quantitative tissue characterization in early detection as well as image-based risk modeling and predictive machine learning. We investigated the behavior of a 2D U-net architecture when being faced with a large set of error-prone training input and compared the results to the training on a smaller but human-corrected high-quality annotation set, both when being used for fine-tuning and training from scratch. Our dataset consists of a total of 876 pre-contrast axial T1 weighted MRI volumes from 166 subjects acquired between 2006 and 2021 in a longitudinal high-risk breast cancer screening cohort study. All images were previously segmented using a fully automated heuristic algorithm, resulting in error-prone segmentation masks, which were used in an initial ‘human-free’ experiment. We randomly separated on a per-subject level 102 volumes from 23 subjects, for which an expert radiologist manually corrected these segmentation masks, providing the basis for the additional two experiments. Our results indicate a subclass of input errors can be compensated for by the regularization capacity of standard deep convolutional neural networks while other errors are learnt or even newly introduced. In particular, tissue boundaries in axial directions were missed in our experiment. In the inner region, the median Dice coefficient of our fine-tuning U-net exceeded 98% at a reasonable robustness and consistency, which is promising given the simplicity of our approach. Future work will address efficient learning schemes, aiming at boosting the segmentation quality with minimal human input, and the boundary issue.