Breast Density Estimation in Mammograms Using Unsupervised Image Segmentation

Abstract

Breast cancer, the most prevalent cancer in women as of 2020, poses significant health risks. Early detection is crucial for effective management, and mammography serves as a key screening method. Mammograms, produced through mammography, are x-ray images which allow radiologists to assess mammographic density—a measure of non-fatty tissue in the breast. The Breast Imaging-Reporting and Data System (BI-RADS) is the current standard for density measurement, and it categorizes density into four classes. The importance of studying breast density is twofold. First, high breast density can be a factor in lowering the mammogram sensitivity as dense tissue, which looks bright on a mammogram, can mask cancers in mammograms. Second, higher breast density is associated with an increased risk of breast cancer, making accurate assessments vital. However, challenges arise in interpreting mammograms, particularly regarding inter-observer variability, especially with dense breasts. This work addresses the need for accurate and objective breast density assessments. It proposes a label-noise-tolerant unsupervised-learning-based method for quantitative breast density estimation. The study begins with a comprehensive review of existing literature and state-of-the-art techniques. A framework for breast density estimation is then introduced, involving mammogram preprocessing, unsupervised segmentation, and percentage density estimation. A convolutional neural network (CNN) with a loss function combining similarity and continuity is adapted for segmentation. The framework is tested on two public datasets (DDSM and INbreast), and its segmentation quality, classification capability, and unsupervised labeling ability are evaluated. Silhouette scores exceed 0.92 for subsets of both datasets, demonstrating strong segmentation performance. Per-patient agreements of 71.43% and 79.28% are achieved for DDSM and INbreast datasets, respectively, comparable to state-of-the-art techniques. The clustering quality assessment confirms reasonable unsupervised labeling, with Silhouette scores averaging around 0.57 for DDSM and 0.50 for INbreast. The proposed framework provides a non-subjective model for quantitative breast density estimation. Its potential benefits extend to clinical settings, where it can aid radiologists in assessing breast density.