Current Methodology Towards a Standardized and Automated Background Parenchymal Enhancement Categorization: A Literature Review

Abstract

Background parenchymal enhancement (BPE) is defined as the amount of enhancement observed in normal fibroglandular tissue (FGT) after contrast administration in breast MR protocols. This phenomenon has been associated to sensitivity reductions, as a consequence of lesion occlusion, and most recently described as a potential imaging biomarker for breast cancer prediction. Thus, a standardized automated methodology for its categorization may be of utmost importance. Nonetheless, the lexicon utilized as a gold-standard by radiologists is associated to large variability and susceptibility in BPE assessment between readers. This study aims to perform a review that recollects the most recent methodologies in the literature for BPE categorization, which entails three different frameworks: BPE quantification, radiomics and deep learning models. Findings indicated that, while the former is widely utilized, its applicability for four-way categorization by discretization is associated with lower correlations to the reference than that of radiologists, thus limiting its applicability to the task under study. Furthermore, machine learning (ML) approaches provided substantially better results towards standardized automated categorization. The potential superiority of ML techniques is shown to be associated with higher correlation coefficients than those achieved by quantification techniques. Within this group, radiomic architectures, which rely on the manual selection of features for proper representation of the tissue, presented an advantage against deep learning architectures. However, their natural dependence on segmentation techniques must be taken into account, as it may be significant of error propagation and a reduction in BPE estimation and categorization. Therefore, although the goal of standardized BPE categorization is still far, this review presents some important observations that may guide future efforts. Additionally, the relevance of balanced data in the development of ML techniques, and the limitation attributed to the wide variability of breast MR protocols among institutions, is highlighted. Future outlooks may focus on differentiations between symmetrical or asymmetrical BPE breasts, and two-way classifications towards more clinically relevant BPE categorization methodologies.