Validation of Deep Learning based Tissue Segmentation for Efficient and Robust Quantitation of Background Parenchymal Enhancement on Breast MRI

Motivation: Quantitative measurements, such as background parenchymal enhancement (BPE) on breast dynamic contrast-enhanced (DCE) MRI, show potential as markers for breast cancer risk. Goal(s): This study compares the efficiency and reproducibility of automated deep learning (DL) vs. semi-automated methods for fibroglandular tissue (FGT) segmentation and quantitation of BPE on DCE-MRI. Approach: Quality of FGT segmentations generated by fuzzy c-means (semi-automated) and DL (fully automated) algorithms were scored by a radiologist, and the agreement of quantitative BPE measurements across segmentation approaches was assessed. Results: Deep learning-based FGT segmentation offers more accurate and robust FGT segmentation, enhancing quality and maintaining BPE reproducibility. Impact: Application of deep learning for segmentation of fibroglandular tissue on breast MRI can improve the robustness and reliability of quantitative imaging biomarkers, with the potential to improve risk stratification and clinical decision-making for high-risk breast cancer screening.