P03.16.A MGMT PROMOTER METHYLATION PREDICTION IN HIGH GRADE GLIOMAS USING CONVENTIONAL MRI AND MACHINE LEARNING SEGMENTATIONS

Abstract BACKGROUND The clinical efficacy of temozolomide varies significantly in gliomas, partly due to methylation of the O6-methylguanine-DNA methyltransferase (MGMT) promoter. Testing for MGMT promoter methylation (MGMTpm) requires tumor tissue. Magnetic resonance imaging (MRI) features have shown potential in predicting MGMTpm, possibly improving pre-operative management. The aim of this study was to evaluate volumetric data from manual and machine learning (ML) segmentation methods to distinguish MGMTpm status in high grade gliomas (HGG). MATERIAL AND METHODS Data from 40 patients (mean age 60 years, 14 females) with HGG, 32 glioblastomas (GBM) and 8 astrocytomas grade 3-4, was retrospectively included from a larger cohort study. MGMTpm was routinely assessed using pyrosequencing, with a cut-off ≥10 % indicating methylation. Conventional MRI sequences were collected pre-operatively, or during follow-up when a recurrent tumor was present prior to re-operation. The images were manually segmented to extract the volume of contrast enhancing tumor, necrosis, edema, total tumor and tumor/edema ratio. The same segmentations were generated using two ML-models, Deep Brain Tumor Image Analysis (DeepBraTumIA) and Raidionics, and compared with the manual volumes to calculate Dice coefficients. All volumes were compared based on MGMT status using Mann Whitney U tests. RESULTS Using Raidionics, MGMTpm tumors displayed significantly decreased mean tumor/edema ratio compared to unmethylated tumors (0.24 vs 0.44, p = 0.04). A trend was seen using DeepBraTumIA (0.39 vs 0.51, p = 0.051) but not with manual segmentation (0.24 vs 0.36, p = 0.17). No significant results were found in the other segmented volumes. Compared to manual segmentation, both DeepBraTumIA and Raidonics generated similar Dice coefficients regarding tumor (0.80 and 0.77) and edema (0.79 and 0.76) but performed worse when segmenting necrosis (0.56 and 0.59). CONCLUSION Our results indicate that ML models such as Raidionics may distinguish MGMTpm from unmethylated tumors using the segmented tumor/edema ratio on conventional MRI sequences. This agrees with some previous studies that have also been able to separate MGMT status based on imaging features. The results could not be replicated using manual segmentation, possibly due to variability particularly in defining necrosis. Both ML models used in this study were trained on pre-operative GBM, emphasizing the importance of model selection and external validation based on the clinical context. Future studies should investigate other ML models in a larger sample size, incorporating different MRI sequences to further distinguish MGMT status based on a multiparametric study design.