Advancements in an Automated Breast Density Detection Technique for Breast Cancer Risk Prediction: a Synthetic Signal-dependent Noise Construct

Breast density is a significant breast cancer risk factor estimated from mammograms and useful for breast cancer risk prediction. We describe enhancements to a previously developed automated percentage of breast density detection method. The approach relies on signal-dependent noise (SDN), where the expected variance and mean (i.e., the signal) are functionally related. SDN variations in different image data representations caused degradation in the algorithm’s performance; this work addressed this problem. Image data used in the analysis were from three breast cancer case-control studies employing different mammographic technologies i.e., full field digital mammography (including raw and clinical images) and digital breast tomosynthesis. Advancements demonstrated that the algorithm produced significant odds ratios across all image data representations by adding these capabilities: (1) the ability to transform noise to a specific optimized quadric SDN form given an arbitrary image data representation; and (2) the ability to perform ensemble averaging over a given image, boosting the signal. We also demonstrated methods to combine measurements from different technologies using a probability density transformation approach. The technique can be applied to images from different technologies with minimal adjustments thereby making it suitable for both research and clinical applications.