Creating a Radiogenomics Map of Multi-omics and Quantitative Image Features in Multiforme Olivier Gevaert; Lex A Mitchell; Achal Singh Achrol; Jiajing Xu; Gary K. Steinberg MD PhD; Samuel Henry Cheshier MD, PhD; Sandy Napel; Greg Zaharchuk; Sylvia K Plevritis

Introduction Results Conclusions Glioblastoma (GBM) is the most frequent The majority (63%) of the quantitative Radiogenomic is rapidly gaining recognition primary malignant brain tumor in adults. image features was robust to intra-reader as a powerful new field that has several Despite decades of research and variation and had meaningful correlations promising applications, such as non- multimodality treatment with microsurgical with survival outcomes, VASARI image invasive molecular lesion assessment. If resection followed by chemotherapy and features, and molecular GBM subtypes. For successful image surrogates can be , mean survival time is example, the irregularity of the enhanced identified that predict relevant molecular only 12–14 months. The development of a tumor edge had the strongest correlation aberrations (e.g. mutation in EGFR), the radiogenomic map – a link between image with overall survival and all but one value added by radiogenomic can be easily features and underlying molecular data - VASARI feature was correlated with at translated as medical imaging is part of holds the potential to address the clinical least one quantitative image feature. The Visualization of the association between routine management in oncology. need for surrogate biomarkers that radiogenomic map showed intriguing module 20 and the LAIIR2-var feature, In this study, we have shown exploratory accurately predict underlying tumor correlations between image features and defined as the variance of the necrosis results of two extensions to radiogenomic biology and therapy response in GBM. molecular pathways. For example Module edge shape. analysis of GBM cases from TCGA. We 20's expression signature is correlated to demonstrated the use of quantitative image Results Continued necrosis edge shape. This module is features in GBM and reported meaningful Methods enriched with genes related to neuronal We used AMARETTO to build a module correlations with VASARI image features, We obtained multi-omics data from 251 development. Overall, this technique network based on 426 TCGA patients. survival and molecular data. We also show patients and MR image data from a subset allowed annotation of 82% of quantitative Next, we created radiogenomic maps the power of creating radiogenomic maps of 55 patients in the Genome Atlas image features with biological pathways. by correlating only prognostic modules using AMARETTO, and using the module (TCGA) and The Cancer Imaging Archive network to indirectly associate image (TCIA) GBM databases. A board certified with quantitative image features. This features with underlying biological neuroradiologist traced 2D regions of establishes radiogenomic maps processes. Our results demonstrate that interest (ROI) around necrotic and between the quantitative image building radiogenomic maps with enhanced parts of the largest lesion in a features and 35 prognostic modules quantitative image features is a promising selected slice from a T1 post-contrast MR, separately for all ROIs. Each map complementary strategy towards non- and around the region of hyperintensity shows several significant associations invasive management of GBM. obtained from the enhancement on the between modules and image features. matched T2 FLAIR slice. These ROIs were used to compute quantitative image References features from their shapes and pixel - Verhaak RG, et al. Cancer Cell. 2010;17(1):98- 110 values. We used a module network - Gutman DA, et al. Radiology. 2013 algorithm called AMARETTO to integrate May;267(2):560-9 copy number, DNA methylation and gene - Jain R, et al. Radiology 2013 Apr;267(1):212- expression data into 100 co-expressed 20. gene modules. We established a - Zinn PO, PLoS One. 2011;6(10):e25451 radiogenomics map by correlating these - Gevaert O, Plevritis S. Pac Symp Biocomput. modules with the quantitative image 2013:123-34. features, and used significant module- - Gevaert O, et al. Interface Focus. Example of necrosis and enhancement 2013;3:20130013. image feature correlation for survival analysis using Cox proportional hazards ROI of a poor prognosis patient. modeling.