Comparison of three algorithms to measure breast density on mammograms in a population-based screening cohort

Abstract Objectives To compare breast density assessments from three automated algorithms by evaluating their agreement and associations with breast cancer risk in a large screening population. Materials and methods We included 61,514 women from the Dutch population-based PRISMA screening cohort (2014–2019). Breast density was quantified on one screening mammogram per participant using Volpara Density Grade 4th and 5th edition, Quantra Density Assessment Software, and iCAD Density Assessment. Agreement between categorical outputs was assessed via weighted kappa statistics. Breast cancer cases were identified through linkage with the Netherlands Cancer Registry (follow-up until December 2022). Associations with five-year breast cancer risk were analyzed using Cox proportional hazards models. Discriminative ability was evaluated using the area under the receiver operating characteristic curve (AUC). Results Good agreement was observed between algorithms (kappa: 0.725–0.826); the strongest across-vendor agreement was between Volpara Density Grade 5th edition and Quantra Density Assessment Software (kappa = 0.795). However, the proportion of women classified with extremely dense breasts (category D) varied (3.1–8.4%). During a median follow-up of 4.2 years, 777 breast cancer cases were identified. Both breast cancer risk and interval cancer risk increased with higher density, with similar magnitudes across algorithms. Categorical measures yielded similar AUCs to distinguish between women with and without breast cancer (0.53–0.55) and between screen-detected and interval cancers (0.57–0.60). Conclusion Three automated density assessment algorithms show strong agreement and comparable associations with breast cancer risk. However, differences in categorization—particularly the proportions classified as entirely fatty or extremely dense—should be considered in personalized screening strategies, as these may influence the capacity needed for potential supplemental screening. Key Points Question How do three automated breast density measurement algorithms compare in terms of agreement and their association with breast cancer risk in a large screening cohort ? Findings The three algorithms showed strong agreement (kappa  =  0.725–0.826) and similar associations with breast cancer risk. Proportions of mammograms classified as density categories A or D varied . Clinical relevance The strong agreement and similar associations with breast cancer risk support algorithm interchangeability, but variations in the proportion of mammograms classified as entirely fatty (category A) or extremely dense (category D) may impact capacity requirements for personalized screening .