Advancements in Developing an Automated Breast Density Detection Technique for Breast Cancer Risk Prediction: Synthesizing a Signal-dependent Noise Stochastic Process

Abstract

Breast density is an important breast cancer risk factor estimated from mammograms and useful for breast cancer risk prediction. We describe a new formulation built on the backbone of an established automated percentage of breast density detection method. This framework relies on signal-dependent noise (SDN), characterized by a statistical dependence between the variance and mean signal. Variations in this dependency due to different image data representations cause degradation in the algorithms performance; the current work addresses this problem by synthesizing a stochastic process conditioned on a given mammogram instead of analyzing the mammogram directly. Image data used in the analysis was derived from three breast cancer case-control studies employing different mammographic technologies including full field digital mammography (both raw and clinical images) and digital breast tomosynthesis. The new formulation produced significant odds ratios across all image data representations due to these methodological advancements: (1) synthesis of SDN to an optimal quadratic structure given an arbitrary image; and (2) ensemble averaging over a given image, boosting the signal. We also demonstrate methods to standardize and combine measurements from different technologies using a probability density transformation technique. This automated technique can be applied to images from different technologies with minimal adjustment, thereby making it suitable for both research and clinical applications.