Tumour heterogeneity revealed by unsupervised decomposition of dynamic contrast-enhanced magnetic resonance imaging is associated with underlying gene expression patterns and poor survival in breast cancer patients

Ming Fan, Pingping Xia, Bin Liu, Lin Zhang, Yue Wang, Xin Gao, Lihua LiBreast 2019 · DOI

Abstract

Tumour heterogeneity is a hallmark of cancer that complicates prognosis and treatment. This study employs an unsupervised decomposition method, Convex Analysis of Mixtures (CAM), to "unmix" Dynamic Contrast-Enhanced MRI (DCE-MRI) signals into distinct kinetic subregions. We demonstrate that imaging features from these subregions, particularly the "maximum probability" texture feature, are significantly associated with recurrence-free survival and underlying gene expression patterns, specifically the Ras signaling pathway.

Introduction

Breast tumours are not monolithic; they exhibit significant internal variation in genetics and vascularity. Standard analysis of Dynamic Contrast-Enhanced MRI (DCE-MRI) often averages signals across the entire tumour volume, which can obscure aggressive "hotspots" and dilute prognostic signals—a phenomenon known as the partial volume effect. To address this, researchers investigated whether mathematically decomposing mixed MRI signals into distinct kinetic subregions could reveal more precise biomarkers for patient outcomes.

Methodology: The Radiogenomic Framework

The study's core innovation is the application of Convex Analysis of Mixtures (CAM) to DCE-MRI data. This unsupervised algorithm treats each pixel as a weighted combination of three "pure" kinetic patterns: plasma input, fast-flow kinetics, and slow-flow kinetics.

Figure 1: Overall study framework, illustrating the identification of prognostic imaging signatures, the radiogenomic association analysis, and the assessment of derived gene signatures.

Following decomposition, 14 features (texture, shape, and histogram-based) were extracted from these subregions. To link these imaging phenotypes to biological drivers, the team utilized Weighted Gene Co-expression Network Analysis (WGCNA) to identify correlations between MRI features and gene expression modules.

Data and Cohorts

The robustness of the findings was tested across four distinct datasets: a Training Cohort (n=61) for initial identification, a Reproducibility Cohort (n=173) for independent validation of imaging features, a Radiogenomic Cohort (n=87) to establish biological links, and a Large-scale Validation Cohort (n=1010) of TCGA gene expression data to confirm the prognostic value of derived multigene signatures.

Key Results: Survival and the Ras Pathway

Analysis revealed that imaging features derived from the "fast-flow" subregion were superior predictors of survival compared to whole-tumour averages. Specifically, a texture feature called "maximum probability" emerged as a potent biomarker. This feature measures the most frequent intensity relationship between neighboring pixels within the fast-flow region.

Figure 5: Kaplan-Meier curves showing recurrence-free survival stratified by (a) maximum probability and (b) tumour volume. The maximum probability feature from the fast-flow subregion provides significant risk stratification.

In the reproducibility cohort, high "maximum probability" in the fast-flow region was significantly associated with poorer Recurrence-Free Survival (RFS). Furthermore, a 43-gene signature associated with this imaging feature successfully stratified the 1,010-patient TCGA cohort into high- and low-risk groups ($p = 0.035$ for RFS). Pathway enrichment analysis linked these prognostic imaging features to the Ras signaling pathway (corrected $p = 0.0044$), a known driver of tumour progression.

Discussion and Conclusion

By "unmixing" MRI pixels, this method effectively performs a mathematical biopsy that reveals biological heterogeneity invisible to standard clinical assessment. While limitations such as data heterogeneity across institutions and the need for high temporal resolution in MRI sequences remain, this radiogenomic approach provides a bridge between macroscopic imaging and microscopic genetic drivers. These subregion-based biomarkers offer a promising path toward more personalized breast cancer management.

Summary updated

Abstract

Dynamic pattern deconvolution revealed that tumour heterogeneity was associated with poor survival and cancer-related pathways in breast cancer.
BibTeX
@inproceedings{fan2019tumour,
  author = {Ming Fan and Pingping Xia and Bin Liu and Lin Zhang and Yue Wang and Xin Gao and Lihua Li},
  title = {{Tumour heterogeneity revealed by unsupervised decomposition of dynamic contrast-enhanced magnetic resonance imaging is associated with underlying gene expression patterns and poor survival in breast cancer patients}},
  booktitle = {Breast},
  year = {2019},
  doi = {10.1186/s13058-019-1199-8},
}
Figure