November-December 2017
Review Article

Radiomics in Brain Tumor: Image Assessment, Quantitative Feature Descriptors, and Machine-Learning Approaches

Gevaert Pic

Olivier Gevaert

Radiomics in brain tumors fits within a larger effort to investigate the links between quantitative imaging and cellular and molecular properties of complex diseases. Radiomics is an essential component in representing medical images (eg, CT, MRI, and PET) as data and linking them in mathematic models to molecular data (eg, DNA or RNA sequencing). This field has been coined radiogenomics, or imaging genomics, and has been the central focus of our work, as we are working toward an eventual multiscale model of complex diseases. Radiogenomics as a novel discipline brings together data from multiple scales for studying complex diseases. We have been working on this topic since 2010, studying several cancer types such as lung cancer,1-3 glioblastoma,4,5 and hepatocellular carcinoma.6,7

Radiogenomics/imaging genomics can result in noninvasive biomarkers that can be used to guide treatment. Radiomics has shown its potential through its ability to predict clinical outcomes (ie, prognosis) and actionable molecular properties of tumors (eg, the activity of epidermal growth factor receptor [EGFR], a major drug target in many cancers). In addition, radiomics maps—multivariate links between image features and drug targets’ activity—provide networks of associations between image phenotypes and potential treatment. For example, we have shown how quantitative image features can be mapped to metagenes for patients with lung cancer,1 we have defined 3 imaging subtypes of adult brain tumors with implications for targeted treatment,5 and we have shown how EGFR mutation status can be predicted from lung CT images.2

Our current research focus is directed toward radiogenomics/imaging genomics, which expands the radiomics field by integrating the molecular makeup of a patient into the model. We are applying this research to several additional tumor types, such as medulloblastoma and head and neck squamous cell carcinoma. For medulloblastomas, we have new results showing that we can predict molecular subtypes using MRI features; we show that we can predict the sonic hedgehog group with AUC = 0.79. For head and neck squamous cell carcinomas, we have discovered 5 novel molecular subtypes using epigenomics analysis, and now we have preliminary results to show that, using CT image features, we can predict these subtypes using noninvasive biomarkers. This means that our work can be more rapidly translated to clinical environments and be used to determine treatment.

In the longer term, we are expanding this research direction by including other types of information and developing a more complete, multiscale model of complex diseases such as cancer. Especially in cancer research, multiscale data are now available to integrate molecular, cellular, and tissue scales and establish a more comprehensive view of lesions. Many previous approaches have focused on mechanistic multiscale modeling by modeling the biochemical reactions and mechanistic relationships between multiscale entities; however, we propose a computational approach by using high-dimensional molecular data together with cellular and tissue scale image data to develop a statistical multiscale modeling approach to cancer.

We are planning to present 2 projects at RSNA 2017 in Chicago.

Read this article at AJNR.org …

References

  1. Gevaert O, Xu J, Hoang CD, et al. Non-small cell lung cancer: identifying prognostic imaging biomarkers by leveraging public gene expression microarray data--methods and preliminary resultsRadiology 2012;264:387–96, 10.1148/radiol.12111607
  2. Gevaert O, Echegaray S, Khuong A, et al. Predictive radiogenomics modeling of EGFR mutation status in lung cancerSci Rep 2017;7:41674, 10.1038/srep41674
  3. Zhou M, Leung A, Echegaray S, et al. Non-small cell lung cancer radiogenomics map identifies relationships between molecular and imaging phenotypes with prognostic implicationsRadiology 2017 [Epub ahead of print], 10.1148/radiol.2017161845
  4. Gevaert O, Mitchell LA, Achrol AS, et al. Glioblastoma multiforme: exploratory radiogenomic analysis by using quantitative image features. Radiology 2014;273:168–74, 10.1148/radiol.14131731
  5. Itakura H, Achrol AS, Mitchell LA, et al. Magnetic resonance image features identify glioblastoma phenotypic subtypes with distinct molecular pathway activities. Sci Transl Med 2015;7:303ra138, 10.1126/scitranslmed.aaa7582
  6. Echegaray S, Gevaert O, Shah R, et al. Core samples for radiomics features that are insensitive to tumor segmentation: method and pilot study using CT images of hepatocellular carcinomaJ Med Imaging 2015;2:041011, 10.1117/1.JMI.2.4.041011
  7. Bakr SH, Echegaray S, Shah R, et al. Noninvasive radiomics signature based on quantitative analysis of computed tomography images as a surrogate for microvascular invasion in hepatocellular carcinoma: a pilot studyJ Med Imaging 2017;4:041303, 10.1117/1.JMI.4.4.041303