Unsupervised Clustering of Quantitative Image Phenotypes Reveals Breast Cancer Subtypes with Distinct Prognoses and Molecular Pathways

Abstract

Purpose: To identify novel breast cancer subtypes by extracting quantitative imaging phenotypes of the tumor and surrounding parenchyma and to elucidate the underlying biologic underpinnings and evaluate the prognostic capacity for predicting recurrence-free survival (RFS).Experimental Design: We retrospectively analyzed dynamic contrast-enhanced MRI data of patients from a single-center discovery cohort (n = 60) and an independent multicenter validation cohort (n = 96). Quantitative image features were extracted to characterize tumor morphology, intratumor heterogeneity of contrast agent wash-in/wash-out patterns, and tumor-surrounding parenchyma enhancement. On the basis of these image features, we used unsupervised consensus clustering to identify robust imaging subtypes and evaluated their clinical and biologic relevance. We built a gene expression-based classifier of imaging subtypes and tested their prognostic significance in five additional cohorts with publically available gene expression data but without imaging data (n = 1,160).Results: Three distinct imaging subtypes, that is, homogeneous intratumoral enhancing, minimal parenchymal enhancing, and prominent parenchymal enhancing, were identified and validated. In the discovery cohort, imaging subtypes stratified patients with significantly different 5-year RFS rates of 79.6%, 65.2%, 52.5% (log-rank P = 0.025) and remained as an independent predictor after adjusting for clinicopathologic factors (HR, 2.79; P = 0.016). The prognostic value of imaging subtypes was further validated in five independent gene expression cohorts, with average 5-year RFS rates of 88.1%, 74.0%, 59.5% (log-rank P from <0.0001 to 0.008). Each imaging subtype was associated with specific dysregulated molecular pathways that can be therapeutically targeted.Conclusions: Imaging subtypes provide complimentary value to established histopathologic or molecular subtypes and may help stratify patients with breast cancer. Clin Cancer Res; 23(13); 3334-42. ©2017 AACR.

Figure 1

Flowchart of the overall design of the study. This study contains three phases. Abbreviations: DCE = Dynamic contrast-enhanced, TCIA = The Cancer Imaging Archive, TCGA = The Cancer Genome Atlas, GEO = Gene Expression Omnibus.

Figure 2

Unsupervised consensus clustering of quantitative imaging phenotypes. A) and C) The consensus matrices represented as heat maps for the chosen optimal cluster number (k = 3) for discovery and validation cohorts respectively. Patient samples are both rows and columns, and consensus values range from 0 (never grouped together) to 1 (always clustered together). The dendrogram above the heat map illustrates the ordering of patient samples in three clusters. B) and D) The corresponding relative change in area under the cumulative distribution function (CDF) curves when cluster number changing from k to k+1. The range of k changed from 2 to 10, and the optimal k = 3.

Figure 3

A) Selected four quantitative imaging features significantly associated with three imaging subtypes, including tumor volume, tumor sphericity, tumor homogeneity measured at early enhancement phase, and background parenchymal enhancement (BPE) fraction with percentage enhancement > 0.6. B) Details of analyzing the tumor and BPE for a representative patient from each imaging subtype. The tumor active function was measured and color-coded with signal enhancement ratio (SER). The BPE was measured and color coded with percentage enhancement at early enhancement phase.

Figure 4

Kaplan-Meier curves of recurrence-free survival stratified by the imaging subtypes. The plots are for A) the discovery cohort, and B–F) five independent validation cohorts, with predicted imaging subtypes via gene expression-based imaging subtype classifiers built in TCGA cohort.

Figure 5

A) Stacked Venn plots of the significantly associated (FDR < 25%) KEGG pathways for three imaging subtypes with Gene Set Enrichment Analysis. B) Pathway activity scores for three imaging subtypes. The pathways are from NCI Pathway Interaction Database Pathways, which are significantly (FDR<25%) associated with three imaging subtypes with PARADIGM analysis. The bar length indicates the magnitude of activity score. From two independent pathway analyses, we observed consistent pathway dysregulation patterns across imaging subtypes, which might explain the differential prognoses associated with the three imaging subtypes.