Quantitative MRI Phenotyping of Breast Cancer across Molecular Classification Subtypes

Abstract

The goal of our study was to investigate the potential usefulness of quantitative MRI analysis (i.e., phenotyping) in characterizing and data mining the molecular subtypes of breast cancer in order to better understand the difference among HER2, ER, and PR expression, triple negative, and other molecular classifications. Analyses were performed on 168 biopsy-proven breast cancer MRI studies acquired between November 2008 and August 2011, on which molecular classification was known. MRI-based phenotyping analysis included: 3D lesion segmentation based on a fuzzy c-means clustering algorithm, computerized feature extraction, leave-one-out linear stepwise feature selection, and discriminant score estimation using Linear Discriminant Analysis (LDA). The classification performance between the molecular subtypes of breast cancer was evaluated using ROC analysis with area under the ROC curve (AUC) as the figure of merit. AUC values obtained for 26 HER2+ vs. 142 HER2-, 118 ER+ vs. 50 ER-, 93 PR+ vs. 75 PR-, 40 Triple Negative (ER-, PR-, and HER2-) vs. 128 all others are 0.65, 0.70, 0.57, and 0.68, respectively for the combined datasets that included images from both 1.5T and 3T scanners. Contributions to the classifiers come from the shape, texture, and kinetics of the lesion, triple negative cases exhibiting increased margin variability, distinct kinetics, and increased surface area. Analyzing the datasets within magnet strength substantially improved performances, e.g., the AUC for triple negative vs. all other cancer subtypes increased from 0.69 (SE=0.05) to 0.88 (SE=0.05). The results from this study indicate that quantitative MRI analysis shows promise as a means for high-throughput image-based phenotyping in the discrimination of breast cancer subtypes.