Abstract
PURPOSE: To develop and validate a fully automated deep-learning pipeline that quantifies mammographic dense rate (MDR) on processed mammograms, and to characterize age-related MDR trajectories in large-scale screening cohort. MATERIALS AND METHODS: Segmentation AI was built with a U-Net segmentation network. The AI model was trained with 300 mediolateral-oblique images (240 for training, 60 for testing) in which pectoral muscle, glandular and fatty tissue were manually annotated. Segmentation accuracy was evaluated with the DICE coefficient whether AI could extract the segment of gland, pectoral muscle and fatty tissue. MDR was calculated, using the AI model. MDR was defined as the proportion of glandular pixels whose intensity exceeded the mean intensity of the segmented pectoral muscle. For each 240,465 processed MLO images from 134,411 women aged 40–79 years acquired in the Yokohama municipal cancer-screening program between January 2019 and April 2022, MDR was calculated. RESULTS: The segmentation model achieved DICE coefficient of 0.967. Mean age was 56.6 ± 11.2 years. Mean MDR was 31.7 ± 25.9% and the median MDR was 26.9%. MDR declined steeply from the early 40s to the late 50s, then plateaued. Longitudinal analysis revealed two distinct patterns; a rapidly decreasing group and a persistently high-density group. Increased MDR may serve as an imaging biomarker preceding cancer diagnosis. CONCLUSION: The proposed vendor-agnostic U-Net pipeline enables accurate, standardized MDR measurement on processed images of mammograms, eliminating observer variability. Large-scale deployment elucidated granular age-specific density dynamics and supports quantitative MDR as a practical tool for risk-adapted breast-cancer screening and personalized selection of adjunct imaging modalities.