Abstract 1494: Non-invasive prediction of molecular subtype in glioblastoma using multi-parametric magnetic resonance imaging pattern analysis and machine learning

Abstract

Abstract Introduction: Imaging studies better capture the spatial heterogeneity of tumors compared to histopathological analysis. Glioblastoma (GB), the most common primary malignant brain tumor, is genetically diverse and may be classified into four subtypes based on gene expression. In this study, we use novel image analysis methods and machine learning to examine the entirety of multi-parametric imaging data and develop an integrative image-based model to predict GB molecular subtype. Methods: We performed a retrospective cohort study of patients with de novo GB at a single academic institution. Tissue samples underwent molecular subtyping using a novel RNA isoform-based classifier. Subtypes included classical, mesenchymal, proneural, and neural. Using initial magnetic resonance (MR) images at time of diagnosis, image pattern analysis techniques identified multiple imaging features from sequences that included T1, T2, T2-FLAIR, and derivatives of diffusion-weighted and perfusion imaging. A machine learning algorithm was then used to analyze multiple features simultaneously to determine which features were most predictive of subtype. Ten-fold cross validation was performed. Results: Molecular subtype was determined for 99 tissue samples. The number of classical, mesenchymal, proneural, and neural subtypes was 29, 22, 20, and 28, respectively. After image feature extraction, a machine learning algorithm identified the following features as most predictive of molecular subtype: T2-FLAIR intensity of enhancing tumor, size of enhancing tumor, and peak height of perfusion signal in edema for the classical subtype, mean T1 intensity in enhancing tumor and T2-FLAIR intensity in edema for the mesenchymal subtype, T2 intensity in edema for the neural subtype, and T2-FLAIR intensity in enhancing tumor and mean T1 intensity in enhancing tumor for the proneural subtype. The balanced accuracy of our image-based model in predicting molecular subtype was 0.71 for proneural, 0.79 for neural, 0.77 for mesenchymal, and 0.68 for classical. Area under the curve was 0.87 for proneural, 0.92 for neural, 0.89 for mesenchymal, and 0.75 for classical subtype. Conclusions: We non-invasively predicted GB molecular subtype with high accuracy using pre-operative MR imaging alone. Imaging features predictive of subtype corresponded with underlying tumor physiology and gene expression. Only through advanced quantitative image analysis are predictive patterns revealed, which would otherwise not be appreciated by examination of individual features. Our image-based model uses standard imaging sequences in clinical practice and is therefore easily translatable to the clinic. Such informatics-derived imaging biomarkers of molecular composition may be applied in future studies to evaluate treatment response over time and in response to targeted agents. Citation Format: Jared Pisapia, Lukasz Macyszyn, Hamed Akbari, Xiao Da, Mark Attiah, Yingtao Bi, Sharmistha Pal, Ramana Davaluri, Laura Roccograndi, Nadia Dahmane, Ronald Wolf, Donald M. O'Rourke, Christos Davatzikos. Non-invasive prediction of molecular subtype in glioblastoma using multi-parametric magnetic resonance imaging pattern analysis and machine learning. [abstract]. In: Proceedings of the 106th Annual Meeting of the American Association for Cancer Research; 2015 Apr 18-22; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2015;75(15 Suppl):Abstract nr 1494. doi:10.1158/1538-7445.AM2015-1494