Imaging signatures of genetic mutations in glioblastoma using machine learning

点击翻译按钮获取中文摘要

Glioblastoma (GB) is the most common and aggressive malignant adult brain tumor, with grim prognosis and heterogeneous molecular and imaging profiles. Although the currently applicable treatment options (i.e., surgery, radiotherapy, chemotherapy) have expanded during the last 20 years, there is no substantial improvement in the OS rates. The major obstacle in treating GBM patients is the heterogeneity of their molecular landscape. Determination of molecular targets requires ex vivo postoperative tissue analyses, which are limited in assessing the tumor's spatial heterogeneity (sampling error due to single sample histopathological and molecular analysis) and temporal heterogeneity (not possible to continuously assess the molecular transformation of the tumor during treatment). Herein we propose to develop quantitative imaging phenomic (QIP) markers of a range of mutations of interest in GB. We will build on prior work on EGFR, IDH1 mutations and MGMT methylation QIP signatures, and develop an extensive panel of imaging signatures of 10 gene mutations, as well as MGMT promoter methylation, using machine learning methods applied to relatively routine clinical mpMRI (standard plus diffusion tensor and perfusion protocols). Availability of such biomarkers can contribute to non-invasive i) patient stratification into appropriate treatments, ii) measurement of individual molecular characteristics. In particular, we propose to carry out the following specific aims: Specific Aim 1 (SA1): To develop the enabling methodologies for constructing Quantitative Imaging Phenomic signatures of GB mutations Specific Aim 2 (SA2): Establish QIP signatures of 10 mutations of interest in gliomas, plus MGMT promoter status, using next generation sequencing (NGS). We will use 709 datasets. Specific Aim 3 (SA3): Characterize the molecular heterogeneity of GB using QIP signatures, leveraging the NGS samples of SA1, as well as a new sample that we will genotype, adding to a total of 600 tumor samples obtained from 4 different locations per patient from 150 patients. The first 150 tissue samples are already analyzed as part of ongoing work. Specific Aim 4 (SA4): Integrate our methods into the Cancer Imaging Phenomics Toolkit (CaPTk), in order to allow easy access to them by users