MRI-Based Deep-Learning Method for Determining Glioma MGMT Promoter Methylation Status

Area of Science:

• Neuroimaging
• Oncology
• Artificial Intelligence

Background:

• O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is a key prognostic and predictive biomarker in gliomas.
• Accurate determination of MGMT methylation status is crucial for guiding treatment decisions and predicting patient outcomes.

Purpose of the Study:

• To develop and validate a deep learning network (MGMT-net) for predicting MGMT promoter methylation status using only T2-weighted MRI (T2WI).
• To simultaneously perform tumor segmentation alongside methylation status prediction.

Main Methods:

• A deep learning network (MGMT-net) utilizing 3D-dense-UNets was developed.
• The network was trained and validated on brain MR imaging (T2WI) and genomic data from 247 glioma patients.
• Three-fold cross-validation was employed to assess generalization performance, with Dice scores used for segmentation accuracy.

Main Results:

• MGMT-net achieved a mean cross-validation accuracy of 94.73% for predicting MGMT methylation status.
• The network demonstrated high sensitivity (96.31%) and specificity (91.66%), with a mean AUC of 0.93.
• The mean Dice score for whole tumor segmentation was 0.82.

Conclusions:

• Deep learning analysis of T2WI alone can accurately predict MGMT promoter methylation status in gliomas.
• This AI-driven approach offers a non-invasive alternative that surpasses traditional histologic and molecular methods.
• This represents a significant advancement towards utilizing MR imaging for predicting glioma prognosis and treatment response.