Machine Learning Approaches for Breast Cancer Risk Prediction Using MRI Radiomics Features

Abstract

There is increasing recognition of the potential value of medical imaging to support personalized breast cancer risk assessment. Techniques such as mammography and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) offer non-invasive, patient-friendly methods for evaluating tissue composition, enhancement patterns, and potential abnormalities. Elevated breast density on mammography and elevated background parenchymal enhancement (BPE) on DCE-MRI have each shown association with increased risk of developing breast cancer. In our study, we investigated radiomics features extracted from DCE-MRI to further enhance the precision of breast cancer risk prediction. In the first part of this study, we compared a fully automated deep learning-based segmentation approach with a semi-automated fuzzy c-means clustering method for fibroglandular tissue (FGT) segmentation, which is critical for extracting meaningful imaging biomarkers. Through a radiologist reader study and quantitative evaluation, we found that the deep learning model produced clinically acceptable FGT delineations and reproduced quantitative background parenchymal enhancement (BPE) values, which have been associated with increased breast cancer risk. These findings support the use of deep learning for more standardized and reproducible quantification of FGT-based imaging features. In the second part, we applied this segmentation model to extract radiomics features from post-contrast DCE-MRI and percent enhancement (PE) maps on whole breast and FGT regions. We developed multiple machine learning classifiers, incorporating principal component analysis (PCA) for dimensionality reduction, and evaluated various combinations of radiomics features, including models that integrated age and conventional imaging factors. Among the approaches tested, the support vector machine (SVM) combined with PCA achieved the highest classification performance (AUC=0.697), outperforming conventional density (AUC=0.572) and qualitative BPE (AUC=0.622) metrics, when incorporating all radiomics features from both image types and both masks in distinguishing women who developed breast cancer within five years (cases) from those who remained cancer-free (controls). In summary, this study demonstrates the utility of fully automated FGT segmentation and MRI-based radiomics for breast cancer risk assessment. Our findings highlight the potential for more standardized, efficient, and personalized screening strategies by integrating imaging biomarkers with machine learning-based prediction models.