Micro-parenchymal patterns for breast cancer risk assessment

Abstract

We evaluated small radiological regions of the parenchymal tissue in mammograms -micro-parenchymal (MP) patterns—for breast cancer risk assessment. We adapted path-based analysis, a computer vision technique, in order to build a model of the distribution of MP patterns in mammograms from a training population sample. Subsequently, the model was utilized to infer the level of risk of individual women based on the distribution of MP patterns in test mammograms. We validated our method using a pilot case/control study with 114 women diagnosed with cancer and 114 healthy controls matched by age, screening year and mammographic system. Experiments with 5-fold cross validation showed a statistically significant positive association between the MP-based risk scores and breast cancer risk with an OPERA (odds per standard deviation of the risk score) value of 1.66 (p-value < 0.001) and an area under the receiver operating characteristic curve (AUC) of 0.653. Results retain their statistical significance after adjusting for visual and quantitative breast densities, widely known imaging biomarkers for breast cancer risk. This work provides experimental evidence that there are specific MP patterns identifiable as cues of breast cancer and prompt the validation of these results in larger datasets.