NIMG-48. ASSOCIATIONS BETWEEN HETEROGENEITY OBSERVED IN O-6-METHYLGUANINE-DNA METHYLTRANSFERASE (MGMT) ASSAY MEASUREMENTS AND MGMT-BASED IMAGING SIGNATURE BASED ON MRI AND DEEP LEARNING

Abstract

Abstract BACKGROUND O-6-methylguanine-DNA methyltransferase(MGMT), a gene promoter, corresponds with the efficacy of standard therapy for Glioblastoma (GBM) patients. If a single sample is sequenced, an appropriate diagnosis of MGMT methylation may be hampered by intratumoral heterogeneity. This work proposes a non-invasive in-vivo estimation of MGMT methylation using deep learning and multi-parametric MRI (mpMRI). Further, it quantifies spatial heterogeneity and establishes the associations between MGMT assay measurement and MGMT imaging signature heterogeneity. METHODS For this study, mpMRI structural scans (T1, T2, FLAIR and T1-Gd) of 498 newly diagnosed GBM patients were considered. These patients underwent surgical tumor resection, and MGMT methylation testing was retrospectively collected. By overlaying local and regional 2D patches of size 32 X 32 over the whole tumor on mpMRI scans, a 10-fold cross-validated deep convolutional neural network model was trained considering five layers on the discovery cohort (n=458). The trained model was applied to an independent multi-sample cohort (n=40) to compute the MGMT methylation at the pixel level for imaging-derived maps generation and heterogeneity quantification. The independent cohort included the patients for whom multiple tissue specimens were sequenced for MGMT methylation. RESULTS The patch-based deep model produced an area under the ROC curve of 0.76 (95% CI: 0.71–0.82) for MGMT status prediction. This model showed a strong association between homogeneously MGMT methylated and heterogeneous cases with MGMT Assay measurement of multiple samples with Pearson’s correlation coefficients of 0.64 (p < 0.05) and 0.78 (p < 0.01), respectively. Further, heterogeneity with a significantly higher variance was also observed in heterogeneous cases in developed MGMT imaging signature. CONCLUSION The proposed model efficiently determines MGMT promoter methylation and quantifies GBM spatial heterogeneity non-invasively. Also, it showed a significant association between the heterogeneity observed in MGMT-based imaging signature and MGMT assay measurements.