Lund University Publications

Noninvasive MGMT-promotor methylation prediction in high grade gliomas using conventional MRI and deep learning-based segmentations

• Mark

Zahirovic, Edin ^LU ; Salomonsson, Tim ^LU ; Knutsson, Malte ^LU ; Sarda, Xavier Saenz ^LU ; Lätt, Jimmy ^LU ; Kinhult, Sara ^LU ; Belting, Mattias ^LU ; Rydelius, Anna ^LU ; Bengzon, Johan ^LU and Knutsson, Linda ^LU , et al.

Abstract

BACKGROUND/OBJECTIVES: High grade gliomas (HGG) are aggressive brain tumors, most frequently glioblastoma and astrocytoma grade 4. Methylation of O6-methylguanine-DNA methyltransferase (MGMT) promoter in HGG is crucial for temozolomide efficacy. As MGMT promoter methylation (MGMTpm) assessment requires tumor tissue, magnetic resonance imaging (MRI) is of interest for non-invasive prediction. We aimed to analyze volumetric data from edema, contrast-enhancing tumor, necrosis, total-tumor and total-tumor/edema ratio for MGMTpm prediction in HGG. Further we assessed overall survival (OS) and progression free survival (PFS) between groups and volumes.

METHODS: Segmentation was performed using deep learning models (DL-models),... (More)

BACKGROUND/OBJECTIVES: High grade gliomas (HGG) are aggressive brain tumors, most frequently glioblastoma and astrocytoma grade 4. Methylation of O6-methylguanine-DNA methyltransferase (MGMT) promoter in HGG is crucial for temozolomide efficacy. As MGMT promoter methylation (MGMTpm) assessment requires tumor tissue, magnetic resonance imaging (MRI) is of interest for non-invasive prediction. We aimed to analyze volumetric data from edema, contrast-enhancing tumor, necrosis, total-tumor and total-tumor/edema ratio for MGMTpm prediction in HGG. Further we assessed overall survival (OS) and progression free survival (PFS) between groups and volumes.

METHODS: Segmentation was performed using deep learning models (DL-models), DeepBraTumIA and Raidionics, on 70 HGG patients (45 males, 32 MGMTpm). Manual segmentation was conducted in 37 for validation of DL-models. Group differences were evaluated using Man-Whitney U tests and receiver operation characteristic (ROC) curves. Multivariate analysis was conducted using logistic regression and bootstrapping. Dice coefficient, intraclass correlation coefficient (ICC) and Kruskal-Wallis test evaluated DL-model performance.

RESULTS: MGMTpm tumors displayed significantly larger edema, segmented by DeepBraTumIA (
p = 0.03), and lower total-tumor/edema ratio segmented by both DL-models (
p < 0.01). Raidionics segmented total-tumor/edema ratio showed highest univariate predictive ability with area under curve 0.687 (sensitivity 46.2%, specificity 87.5%). Multivariate analysis confirmed this, showing that the ratios from both DL-models were the only ROIs to remain independent, significant predictors (
p < 0.05) after controlling for clinical covariates. The overall multivariate models were significant (
p = 0.01) and improved prediction over baseline. ICC showed interclass correlation of 0.96 (contrast-enhancing tumor), 0.50 (tumor necrosis) and 0.90 (peritumoral edema). Segmentation methods demonstrated 83-91% median overlap in contrast-enhancing tumor, 67-80% in necrosis and 80-84% in edema regions. Significant OS and PFS differences were observed, notably being longer in MGMTpm tumors and lower tumor volumes.

CONCLUSION: This study suggests that significant radiological differences in MGMTpm can be found using deep learning models, primarily in tumor edema volume. MGMTpm status and region of interest volumes impact OS and PFS. Future studies should incorporate other molecular imaging sequences for methylation prediction.

(Less)