Abstract 5879: The impact of radiomic imaging features in predicting short term risk of breast cancer after a negative mammogram using machine learning

Abstract

Abstract Introduction & Purpose: While mammography reduces breast cancer mortality, up to 20% of breast cancers are diagnosed within the screening interval of a prior negative mammogram. Advances in breast imaging and machine learning have enabled evaluation of radiomic markers from mammography imaging with respect to breast cancer risk. Identification of radiomic markers associated with short term risk of breast cancer would be useful to direct supplemental screening. The aim of this study was to evaluate the potential of mammographic imaging features beyond breast density to predict screening failure using machine learning. Methods: We utilized full-field digital mammography (FFDM) studies from a cohort of women aged 40-85 who had a negative screening mammogram at the University of Pennsylvania Health System from 2010 - 2015. BRCA1/2 carriers, women with prior breast cancer, and women with breast implants were excluded. We identified 125 breast cancer cases diagnosed within 2 years of a negative mammogram whom we matched to controls not diagnosed with breast cancer within 2 years in a 1:5 ratio on BI-RADS breast density category, age, race, and screening date, resulting in a total of 750 FFDM studies. For each FFDM image, a total of 344 radiomic features were extracted using a previously validated lattice-based computational approach for parenchymal pattern characterization. Radiomic features were summarized across both breasts and across mammographic image views (CC and MLO) using means and standard deviations of radiomic features. Principal component analysis was applied to z-score normalized radiomic features and four principal components (PCs) were retained. Widely used machine learning models, such as Support Vector Machines (SVMs) and Random Forests (RFs), were applied. The performance of the machine learning models in predicting screening failure on the basis of the radiomic PCs was evaluated via cross-validated calculations of the area under the ROC curve (AUC). Results: Upon fine-tuning, SVM models using the mean summaries of radiomic features demonstrated a cross-validated case-control discriminatory capacity of AUC = 0.61 (95% CI 0.52, 0.68) while RFs did not achieve significant performance (AUC = 0.53, 95% CI 0.45, 0.60). Neither SVM models, nor RFs achieved significant predictive performance as measured by the cross-validated AUC when the standard deviation summaries of radiomic features was used. Conclusions: Given that BI-RADS density is the standard measure of breast density used clinically to direct supplemental screening, measuring the performance of radiomic features beyond BI-RADS provides the most clinically useful findings. In this small sample, our results suggest that mammographic radiomic features provided only moderate discriminatory accuracy, when BI-RADS density, age, race, and screening date were controlled for via matching. Citation Format: Aimilia Gastounioti, Sarah Ehsan, Walter Mankowski, Emily F. Conant, Despina Kontos, Anne Marie McCarthy. The impact of radiomic imaging features in predicting short term risk of breast cancer after a negative mammogram using machine learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5879.