Quantitative Glandular Tissue Component Assessment on Breast Ultrasound Using an Anatomically Constrained Deep Learning Framework

Background/Objectives: Glandular tissue component (GTC) has been incorporated into BI-RADS v2025 as part of breast tissue composition assessment, but visual grading remains subjective and does not provide continuous quantification. We developed and evaluated an ROI-guided, anatomically constrained deep learning framework for quantitative GTC assessment on breast ultrasound. Reader performance was compared between unaided and AI-assisted sessions. Methods: This retrospective single-center study included 421 ultrasound images from 114 women for model development and 150 women for clinical evaluation. The framework segmented fibroglandular tissue (FGT) and glandular tissue to calculate continuous GTC percentage and assign the corresponding image-derived P1–P4 category. Four model configurations were compared during model development using patient-level five-fold cross-validation. Standalone AI performance was evaluated against expert consensus. Two senior and two junior radiologists assessed the clinical cohort without and with AI assistance, separated by a 4-week washout period. Results: During model development, the anatomical constraint markedly improved anatomical consistency while quantitative error and segmentation performance remained comparable. In the clinical cohort, continuous GTC percentage showed an MAE of 3.88 percentage points and strong correlation with the reference percentage (Spearman ρ = 0.957, p < 0.001). Exact P1–P4 agreement was 78.0%, adjacent agreement was 99.3%, and weighted κ was 0.888. However, 26 of 36 reference P4 cases (72.2%) were underestimated. Exact agreement was significantly higher for all four readers in the second, AI-assisted session than in the first, unaided session (all p < 0.001). Conclusions: The framework enabled anatomically consistent continuous GTC quantification and categorical assessment on individual breast ultrasound images.