Unmasking Malignancy in Dense Breast Tissue (American College of Radiology Types C and D): A Biopsy-Proven Analysis of Digital Breast Tomosynthesis Performance, Clinical Decision Dynamics, and Correlation with Histological Grade and Molecular Subtype

Background/Objectives: This study aimed to evaluate the diagnostic performance and clinical decision-making impact of adding digital breast tomosynthesis (DBT) to two-dimensional digital mammography (2D DM) in a 100% biopsy-proven cohort of women with dense breast parenchyma (ACR types C and D), focusing on the relationship between tumor biology and radiologic conspicuity. Methods: This retrospective study analyzed 123 women (mean age 46.7 ± 9.9 years) with ACR type C (n = 67) or D (n = 56) breast density who underwent sequential 2D DM and DBT imaging, all with definitive histopathology (malignant, n = 78; high-risk, n = 8; benign, n = 37). A single expert radiologist performed independent readings of 2D DM alone and then the combined 2D DM + DBT dataset after a 4-week washout. Outcomes included sensitivity, specificity, PPV/NPV, lesion visibility (0–5 scale), and BI-RADS upgrade rate. Results: 2D DM + DBT significantly outperformed 2D DM alone in sensitivity (97.4% vs. 69.2%; p < 0.000001) and NPV (93.1% vs. 57.1%), with AUC improving from 0.727 to 0.922 (p < 0.001). Mean lesion visibility increased from 2.94 ± 1.19 to 4.52 ± 0.78 (p < 0.0001), with the largest gains for architectural distortion (100% sensitivity) and, notably, for Grade 1 tumors (delta +1.90), which reached the highest DBT visibility of any grade. Major BI-RADS upgrades occurred in 27 cases (22.0%; upgrade PPV 88.9%). Conclusions: In this biopsy-enriched ACR C/D cohort, adding DBT to 2D DM was diagnostically superior across all metrics. The disproportionate visibility gain for Grade 1 tumors is a hypothesis-generating finding relevant to the overdiagnosis debate, not proof that DBT reduces overdiagnosis. These findings support DBT as a valuable adjunct to 2D DM in dense breasts, pending prospective, multi-reader validation in unselected populations.