Interobserver agreement in breast radiological density attribution according to BI-RADS quantitative classification

The authors sought to assess interobserver agreement in classifying mammography density according to quantitative Breast Imaging Reporting and Data System (BI-RADS) criteria.

Six expert mammography readers were tested on a set of 100 mammograms. Interobserver agreement was determined according to the kappa statistic, adjusting for chance agreement, on a four-category (D1 vs. D2 vs. D3 vs. D4) or two-category (D1-2 vs. D3-4) basis. Agreement with a panel of 12 readers who had been tested on the same set in a previous study was also assessed.

The six readers showed good agreement when compared in pairs [agreement on a four-category basis was substantial (kappa=0.60-0.80) for 13 pairs and almost perfect (kappa>0.80) for two pairs); agreement on a two-category basis was substantial for 12 pairs and almost perfect for three pairs) or compared with the panel (on a four-category basis, agreement was substantial for five of six readers and almost perfect for one; on a two-category basis, agreement was substantial for all readers).

In agreement with previous studies, visual classification of mammography density according to BI-RADS quantitative criteria was highly reproducible among readers; nevertheless, attribution to the "dense breast" (BI-RADS D3-4) category, which might be adopted as a determinant of different screening protocols (such as adjunct ultrasonography or yearly interval) varied among readers (range 6-15%). Controlled studies should be performed comparing visual with computer-density category attribution, the latter possibly being a better alternative due to its absolute reproducibility.