Brad M. Keller PhD, Presenter: Nothing to Disclose

Jinbo Chen PhD, Abstract Co-Author: Nothing to Disclose

Nigel Sloan Bristol, Abstract Co-Author: Nothing to Disclose

Meng-Kang Hsieh, Abstract Co-Author: Nothing to Disclose

Shonket Ray PhD, Abstract Co-Author: Nothing to Disclose

Marie Synnestvedt, Abstract Co-Author: Nothing to Disclose

Emily F. Conant MD, Abstract Co-Author: Scientific Advisory Board, Hologic, Inc

Despina Kontos PhD, Abstract Co-Author: Nothing to Disclose

Breast density assessment is known to be subject to substantial intra- and inter-reader variability. Given the increased legislation mandating routine reporting of breast density, we evaluate the feasibility of fully-automated breast density assessment in a large screening cohort.

We report data from 10,751 screening mammography exams from an entire one-year cohort at our institution (2010-11). All digital mammograms were acquired with either a GE Essential (N=1,511) or Hologic Selenia (N=9,240) system. All “For Presentation” images were analyzed using a previously-validated algorithm developed at our institution that provides estimates of dense area, percent density (PD%) and BI-RADS density categories from either “For Processing” or “For Presentation” digital mammograms. Agreement between left and right breast density estimates were assessed via Pearson correlation (r) as a measure of the algorithm’s consistency. Cohen’s weighted-kappa (k) was used to evaluate agreement between the automated estimates and BI-RADS density categories assigned by the interpreting radiologists. Logistic regression was performed to determine if automated density measures are significant predictors in identifying women recalled for additional imaging (N=1,116), after adjusting for age, race and the availability of prior mammograms.

Both the absolute area and PD% automated measures demonstrate high reproducibility with a strong bilateral per-woman correlation (r>0.93, p<0.001). Substantial agreement (k=0.63; p<0.001; CI: 0.62-0.60) is observed between the algorithm-estimated and radiologists’ BI-RADS density scores, which is in range of previously reported inter-radiologist agreement in the literature. The automated BI-RADS density estimate is also a significant predictor of recall (OR: 1.17 per increasing density category; test-for-trend p=0.002), as were age (p<0.001) and the availability of prior mammograms (p<0.001).

Fully-automated analysis of “For Presentation” digital mammograms can be used to obtain reproducible measures of both continuous and categorical breast density estimates. This could be of particular use when “For Processing” images are not routinely available for analysis.

Accurate and reproducible breast density estimation using fully-automated software may be feasible for large-volume breast screening centers for the purpose of standardized density reporting.

Keller, B, Chen, J, Bristol, N, Hsieh, M, Ray, S, Synnestvedt, M, Conant, E, Kontos, D, Feasibility of Fully-Automated Breast Density Reporting in a Large Academic Center: Prospective Data from a One-Year Screening Cohort.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14045668.html