Multi-site Segmentation of Breast and Fibroglandular Tissue in MRI with a Focus on Clinical Practicality

Breast cancer is the most prevalent cancer among women globally, and MRI is increasingly being proposed as a screening tool in addition to its role as a diagnostic imaging modality. This creates a need for automated methods that analyse breast MR images. Two important basic tasks in this area are segmentation of the breast region and the fibroglandular tissue. We train 3D U-Nets and nnU-Nets for these tasks and assess the segmentation quality as well as the inference time on a diverse multi-centric dataset. In particular we analyse how well models generalize to data from other sites if they are only trained on a subset of sites. For breast segmentation, our models achieve a mean (SD) Dice score of 0.93 (0.02) with a mean (SD) inference time of 18 (13) seconds on a CPU, and for fibroglandular tissue segmentation a mean (SD) Dice score of 0.76 (0.11) with a mean (SD) inference time of 6 (4) seconds on a CPU. These inference times would permit integration of the algorithms into clinical workflows that allow only little time to process the images. Regarding the influence of data from different sites we find that data from one site can be sufficient to create models that generalize well across various sites, if selected appropriately. We suggest from our experiments that the decisive characteristic is the inclusion of both fat-suppressed and non fat-suppressed images.