Breast density assessment: image feature extraction and density classification with machine intelligence

Breast composition density has been identified to be a risk factor of developing breast cancer and an indicator of lesion diagnostic obstruction due to masking effect in x-ray mammography images. Volumetric density measurement evaluates fibro-glandular volume, breast volume, and breast volume density measures that have potential advantages over area density measurement in risk assessment. Compared to traditional x-ray absorption computing based areal and volumetric tissue density assessments, image feature detection based density classification approaches emulate the clinical density evaluation process by radiologists instead of using indirect information (e.g., percentage density values). We have modeled breast density assessment as a machine intelligence task which automatically extract the image features and dynamically improves density classification performance in clinical environment: (1) a bank of deep learning networks are explored to automatically extract the image features that emulate the radiologists' image review process; (2) the pretrained networks are retrained with clinical 2D digital mammography images (for processing and for presentation DICOM images) using transfer learning; (3) a deep reinforcement network is incorporated through human-machine gaming process. The data preprocessing, trained models / processes have been described, and the classification inference have been evaluated with the predicted breast density category values of the clinical validation 2D digital mammographic images in terms of statistic measures. The experimental results have shown that the method is promising for breast density assessment.