Interobserver and intraobserver variability in determining breast density according to the fifth edition of the BI-RADS® Atlas

To determine the level of agreement within and between observers in the categorization of breast density on mammograms in a group of professionals using the fifth edition of the American College of Radiology's BI-RADS® Atlas and to analyze the concordance between experts' categorization and automatic categorization by commercial software on digital mammograms. Six radiologists categorized breast density on 451 mammograms on two occasions one month apart. We calculated the linear weighted kappa coefficients for inter- and intra-observer agreement for the group of radiologists and between the commercial software and the majority report. We analyzed the results for the four categories of breast density and for dichotomous classification as dense versus not dense. The interobserver agreement among radiologists and the majority report was between moderate and nearly perfect for the analysis by category (κ=0.64 to 0.84) and for the dichotomous classification (κ=0.63 to 0.84). The intraobserver agreement was between substantial and nearly perfect (κ=0.68 to 0.85 for 4 categories and k=0.70 to 0.87 for the dichotomous classification). The agreement between the majority report and the commercial software was moderate both for the four categories (κ=0.43) and for the dichotomous classification (κ=0.51). Agreement on breast density within and between radiologists using the criteria established in the fifth edition of the BI-RADS® Atlas was between moderate and nearly perfect. The level of agreement between the specialists and the commercial software was moderate.