Despina Kontos PhD, Presenter: Research Consultant, Hologic, Inc Support agreement, Hologic, Inc

Iosif Tournas BEng, MENG, Abstract Co-Author: Nothing to Disclose

Brad M. Keller PhD, Abstract Co-Author: Nothing to Disclose

Jinbo Chen, Abstract Co-Author: Nothing to Disclose

Emily F. Conant MD, Abstract Co-Author: Consultant, Hologic, Inc

To evaluate feasibility of developing a classification model based on computer-extracted imaging features of lesion and parenchymal pattern characteristics for predicting the probability of a false versus true positive biopsy after diagnostic evaluation from screening mammography.

A retrospective cohort study was performed in a total of 10,187 women who underwent screening mammography between 01/2009-06/2010. In this screening cohort, 255 women underwent biopsy after callback for diagnostic evaluation, yielding 66 malignancies (true-positives) and 189 benign or high risk lesions (false-positives). All digital mammograms (Selenia, Hologic Inc.) were analyzed for computer-aided lesion detection and breast density using FDA-cleared software (R2 ImageChecker CAD, Quantra, Hologic Inc.). A total of 8 features were extracted, specifically per case, bilateral, total number of lesion CAD marks, calcifications and masses, total CAD marks with low and high severity, average breast and glandular tissue volume, and volumetric percent density. Multivariable logistic regression analysis with stepwise feature selection was performed to predict false-positive biopsies. All extracted imaging features were used as model inputs as a surrogate of overall parenchymal tissue complexity. Age was considered as an additional predictor variable. Area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate prediction performance.

The logistic model has an AUC of 0.78 (p<0.001) in predicting false-positive biopsies. Significant predictor variables retained by the model after step-wise feature selection were total number of CAD marks (p<0.001), total number of calcifications (p=0.002), total number of high severity CAD marks (p=0.04), and age (p<0.001). Positive predictive value (PPV) was equal to 80% and negative predictive value (NPP) was equal to 67%.

Computer-extracted mammographic features may have value in predicting false versus true-positive biopsies. Larger studies are warranted in screening populations to validate the accuracy of false-positive biopsy risk prediction models, including additional clinical indicators and in conjunction to cancer detection rate.

Prediction models based on imaging and clinical patient factors may be feasible for estimating risk of false-positive biopsy after diagnostic evaluation for reducing the number of unnecessary biopsies

Kontos, D, Tournas, I, Keller, B, Chen, J, Conant, E, Computer-aided Mammographic Feature Analysis for Classification of False versus True Positive Biopsies.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12035140.html