Martin Lillholm PhD, Presenter: Employee, Biomediq A/S Shareholder, Biomediq A/S

Joselene Marques PhD, Abstract Co-Author: Employee, Biomediq A/S

Dan Richter Jorgensen PhD, Abstract Co-Author: Employee, Biomediq A/S

Kersten Petersen, Abstract Co-Author: Employee, Biomediq A/S

Nico Karssemeijer PhD, Abstract Co-Author: Shareholder, Matakina International Limited Scientific Board, Matakina International Limited Shareholder, QView Medical, Inc Research Grant, Riverain Technologies, LLC

Mads Nielsen PhD, Abstract Co-Author: Stockholder, Biomediq A/S Research Grant, Nordic Bioscience A/S Research Grant, SYNARC Inc Research Grant, AstraZeneca PLC

Mammographic Density (MD) is scientifically well established as an independent breast cancer risk factor and is increasingly used in clinical practice for personalized screening. Mammographic parenchymal patterns/textures beyond MD is known to provide further risk segregation. We investigate the risk segregation potential of an ensemble of mammographic density and texture measures from FFDM.

A case/control study was selected from the Dutch Breast Cancer Screening program. Mammograms were acquired on a Hologic Selina FFDM system with a 70μm pixel size. 250 screen detected cancer cases were chosen randomly and likewise for 750 cancer-free controls. Cases were represented by the latest cancer-free prior contralateral MLO view (if available) and otherwise by the contralateral MLO view at time of diagnosis. Controls were represented by the latest available MLO view (laterality matched individually per case). Volumetric density was calculated using Volpara 1.45. Based on a large literature review (e.g. Giger, Manduca, Häberle, Heine), a total of 56 measures of mammographic texture were implemented and measured for each mammogram. Furthermore, a novel machine learning based texture measure was trained on an independent training set and measured on each mammogram. A multivariate logistic regression model for all 58 measured (and linearly age-corrected) markers was 5-fold cross-validated and evaluated for association to cancer outcome through AUC.

In total, 28 of the age-corrected markers resulted in AUCs significantly better than chance. Specifically, the novel machine learning based measure resulted in an AUC of 0.65 (95% CIs 0.61-0.69) whereas volumetric density was non-significant (AUC 0.51). The multivariate logistic regression yielded a pooled AUC across the five folds of 0.75 (0.71-0.79).

By combining a representative range of published mammographic texture measures with a novel machine learning based approach, it was possible to separate future cancer cases from healthy controls to a degree that a) clearly improves on what is attainable through density alone and b) could facilitate personalized screening of, e.g., high risk women.

Clinical practice includes screening based on, e.g., age and family history through risk models as Gail and Tyrer-Cuzick. Such models could be enhanced by texture measures from routine mammograms.

Lillholm, M, Marques, J, Jorgensen, D, Petersen, K, Karssemeijer, N, Nielsen, M, Enhanced Personalized Breast Cancer Screening Using an Ensemble Model of Mammographic Texture and Density.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14006527.html