Radiogenomics of breast cancer using dynamic contrast enhanced MRI and gene expression profiling

Recent advances in image acquisition, computational power, and algorithmic development have increased the awareness of computational image analysis methods, allowing for computer-extracted features, i.e. phenotypes, from computer-aided diagnosis and quantitative imaging algorithms to yield “radiomics”, i.e., the high throughput conversion of complex imaging data sets into a multi-dimensional feature space with extractable characteristics [1‐3]. Imaging techniques can provide extractable information about the tumor non-invasively and have been shown to provide information about the underlying histopathology [4‐9] as well as genetic makeup of the tumor and tumor microenvironment [10‐13]. Adoption of this technology in an effort to phenotypically characterize solid tumors have been performed primarily in patients with glioblastomas, head and neck, lung, and breast cancers [10, 12, 14‐26].

Correlating specific image-based phenotypes (radiomics) with large-scale genomic analyses (genomics) is an emerging area of research commonly referred to as “radiogenomics” or more specifically “imaging-genomics”. This emerging field addresses novel high-throughput methods of associating information-rich radiographic images with genomic data as well as other clinically relevant information [2, 12, 13, 21‐25, 27]. Radiogenomics has the potential to impact diagnostic and therapeutic strategies by creating more individualized prognostic signatures and real time measurements in response to therapy.

The application of radiogenomics to breast cancer has recently been gaining more attention. For example, studies using mammographic images showed that radiographic texture analysis could identify patients who are more likely to carry BRCA1/2 mutation [28], and parenchymal pattern and breast density were associated with UGT2B gene variation [29]. Higher-dimensional imaging data such as magnetic resonance imaging (MRI) have also been used to associate radiologic features with underlying clinical or histological characteristics [12, 13, 16, 17, 30‐34]. However, most of these studies have focused on a few clinical, histopathologic, or genetic features. For example, Bhooshan et al. classified features that were correlated with the ability to distinguish invasive vs. non-invasive lesions as well as with tumor grade [12, 13]. Mazurowski et al. demonstrated a relationship between MRI enhancement dynamics with the luminal subtypes of breast cancer [33]. Agrawal et al. was able to show a correlation between certain MRI phenotypes and HER-2 receptor subtype [16]. Gene expression profiling was examined to a limited degree in another study that correlated MRI phenotypes with Oncotype DX output that predicted for survival [31].

Recent studies conducted by the TCGA Breast Phenotype Group examined breast MRI radiomics with clinical stage, molecular classification, risk of recurrence, gene expression profiling, and other genomic analyses [21‐25]. In the most comprehensive breast radiogenomics study to date, they analyzed 91 breast cancer cases that were extracted from The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA) datasets. In that study, MRI tumor phenotypes derived from 1.5 Tesla scanners were examined in relation to genomic features including DNA mutation, copy number variation, miRNA expression, protein expression, and gene expression profiles. Their data showed that gene expression profiling provided the most robust information in terms of correlation with radiomic phenotypes [25].

Here, we sought to further elucidate and validate relationships between imaging phenotypes and gene expression pathways by analyzing clinical MRI data of breast cancers taken from higher resolution 3 T scanners at the University of Chicago Medical Center using RNA sequencing transcriptomic data extracted from corresponding frozen tissue samples.

We successfully obtained radiomic and gene expression data from 47 cases (Fig. 1). Analysis of clinical DCE-MRI images of these cases along with corresponding gene expression data derived from RNA sequencing revealed 119 gene expression pathways that had at least one significant association with a radiomic phenotype (Additional file 1: Table S1). Several of these pathways were disease-specific pathways for particular organ systems, such as cardiovascular or neurological disease associations. As we were interested in basic cellular and physiologic pathways, we proceeded to analyze only non disease-specific pathways, which resulted in a total of 91 expression pathways representing 26 major categories, as denoted in the KEGG database (Additional file 1: Table S1). To analyze which pathway categories had the largest number of associations with radiomic phenotypes, we examined the average number of significant associations per pathway according to each of the 26 major categories (Fig. 2). The most heavily represented category involved immune regulation pathways. Other pathways that were also highly represented included those involved in cytokine-cytokine and cell adhesion receptor signaling, cellular proliferation and growth, lipid and glycan metabolism, and canonical signaling cascades including the JAK/STAT and VEGF pathways. We examine several of these major categories in further detail.

Many of our findings suggest biological plausibility and highlight interesting associations between the gene expression pathways and radiomic phenotypes that may be clinically relevant. For example, all size phenotypes were positively correlated with cell proliferation pathways and negatively correlated with the apoptosis pathway. The relationship between cell death and shape phenotypes suggests that increased expression of the apoptosis pathway was associated with more spherical and less irregular tumors. While this particular relationship is biologically predictable, it provides reassurance about the validity of using these imaging features to gain insight about others, perhaps less obvious gene expression pathways.

Perhaps most interestingly, we found that a diverse array of immune related pathways had the most robust relationship with imaging features. As a group, these pathways tended to associate with similar features in the same directionality: tumors with increased degree of overall immune activation appeared to be more confined on imaging as shown by negative correlation with size features and positive correlation with sphericity. They also tended to be more heterogeneous in texture upon contrast enhancement as seen by positive correlation with entropy and negative correlation with energy. Many of these imaging features were also similarly associated with cytokine-cytokine and CAM pathways as well as the JAK/STAT signaling cascades, which play an important role in immune regulation and in particular, lymphocyte signaling and activation. For example, Additional file 3: Figure S2C and 2D show two tumors of similar size, but the former represents a tumor with the highest level of T-cell receptor signaling activity while the latter represents a tumor with the lowest (Additional file 3: Figure S2). Whether this may in part represent an influx of tumor infiltrating lymphocytes deserves further characterization, as it may assist in the prediction of response to immune based therapies and neoadjuvant chemotherapy [45, 46] and potentially other treatment options as well.

The association between the VEGF pathway and several radiomic phenotypes was also demonstrated in our study and has a plausible mechanistic link. Dynamic contrast enhanced MRI can quantify changes in the microvascular physiology of tumors [47], and studies have shown the ability for VEGF to alter the trans-endothelial transport of contrast agent by diffusion [48]. In our study, increased expression of VEGF was positively correlated with variability of enhancement and negatively correlated with enhancement texture homogeneity, supporting the idea that VEGF activation may result in the disorganized formation of “leaky vessels” quantifiable on DCE-MRI.

This is the first report to our knowledge that integrates whole transcriptome analysis using RNASeq with 3 T DCE MRI data of patients with breast cancer. However, we acknowledge limitations of the study including the relatively small sample size leading to an inability to determine the causal relationship between the gene pathway and radiomic features. Nevertheless, the associations that we do see are biologically plausible and could lead to further advances in the analysis presented here. Several other challenges remain that must be overcome in order to better utilize the potential offered by this approach. For example, studies to date have only centered on data gathered from a single tissue sample of a given primary tumor. It is well known that the genetic makeup of tumors is spatially heterogeneous, and thus, a more accurate approach would be to correlate multiple tumor biopsies with their respective anatomic location on MRI. With techniques such as whole mount pathology being developed, it may become more feasible to pursue this idea. Also, standardization of imaging features poses a challenge. While our focus on obtaining images from one type of scanner and protocol improved the reliability of our radiomics data set, we acknowledge that image acquisition protocols also vary across institutions, and generalization of our findings will need further validation with larger sample sizes and examination of robustness studies as previously conducted for computer aided detection [39]. Lastly, as advanced MRI and other imaging techniques [49] become more widely employed, further harmonization and refinement of novel relationships may be discovered.

The recently published TCGA/TCIA Breast Phenotype Group study also combined DCE-MRI with RNA sequencing data in 91 patients, using the same quantitative radiomic approach developed by the Giger lab, although the MRIs performed in that study were acquired across several different institutions and used 1.5 T scanners. Our work further supports the feasibility of incorporating radiomics from clinical MRIs with gene expression data in support of the TCGA/TCIA Breast Phenotype Group approach using higher resolution 3 T scanners [25].

The incorporation of automated feature extraction algorithms (i.e. quantitative radiomics) into routinely performed, noninvasive imaging modalities, such as DCE-MRI has the ability to stimulate the development and use of imaging biomarkers that may provide clinicians inferred biological information otherwise attainable only through direct tissue biopsy. Potential applications include not only improvement of diagnostic ability, but also adaptation of neoadjuvant treatment plans of primary tumors in real time, such as that seen in the recent I-SPY trials [50].

We provide further evidence that MR imaging of breast tumors can infer underlying gene expression using RNA sequencing data. As expected, size and shape features were appropriately correlated with proliferative and apoptotic pathways. We also discovered a large number of radiomic feature associations with immune pathways, which raises the interesting possibility of using MR imaging to distinguish tumors that are more immunologically active, although further studies are necessary to confirm this observation.