Mammographic Breast Density Classification by Integration of Deep Dictionaries and Multi-Model Sparse Approximations

Chelsea Harris, Uchenna Okorie, Sokratis Makrogiannis

Abstract: Highly dense breasts have been associated with higher risk for breast cancer. Breast cancer assessment is a challenging task that routinely requires visual evaluation by trained radiologists. Dense tissue patterns complicate this evaluation by causing a masking effect on breast lesions. In this work, we investigate the synergy of deep features extracted from CNNs with sparse coding classifiers for breast density assessment. Sparse approximation methods were employed to class-specific dictionaries of deep features instead of a combined class dictionary. Additionally, we investigate the performance of other machine learning techniques, such as SVMs, decision trees and CNN classifiers. We performed experiments on regions of interest in mammograms that include benign or malignant lesions. We evaluated our class-specific sparse representation classification using deep features (DF-SRC) method, with deep features extracted from five pre-trained CNNs. Our findings suggest that our DF-SRC technique performance is comparable to CNN classification with transfer learning and offers the benefit of decision interpretability.