Comparative Evaluation of Structural, Diffusion Tensor, and Perfusion MRI for MGMT Methylation Prediction in Glioblastoma Using Deep Learning Approach

Glioblastoma Multiforme (GBM) brain cancer poses a significant clinical challenge due to its aggressive behavior and poor prognosis. The survival rate of these types of cancer patients is also minimal. O(6)-Methylguanine-DNA Methyltransferase (MGMT) enzyme which is encoded by the MGMT gene, plays an important role in restoring the cancer cell. When the MGMT gene promoter is methylated, the expression of the MGMT enzyme is reduced. This makes the cancer cells more sensitive to alkylating chemotherapy for damage. Therefore, MGMT gene methylation status is considered a prognostic biomarker in clinical practices. However, the molecular diagnostics of MGMT promoter methylation status are invasive and time-consuming. Previous studies have shown the significance of different MR imaging with radiomics and deep features for the identification of MGMT gene methylation using machine and deep learning approaches. Nevertheless, the selection of the specific type of MR imaging modality for better prediction outcomes is still challenging. This study attempts to analyze different 3D MRI modalities to determine the suitable modality for MGMT methylation identification. This study analyzes derivatives from 836 images for each of the following modalities: Structural MRI, Diffusion Tensor Imaging (DTI), and Dynamic Susceptibility Contrast (DSC) perfusion MRI. These images are preprocessed and fed into the deep 3D CNN convolutional neural network and evaluated with a five-fold stratified cross-validation. The prediction performance between the two classes is measured by accuracy, precision, recall, f1 score, and confusion matrices. The model performance with sMRI achieves an accuracy of 94.62% and an f1 score of 93.25%. Whereas using the same model with DSC imaging shows an accuracy of 87.74% and an f1 score of 84.15%. For DTI imaging, the accuracy is 76.18% with an f1 score of 75.72%. The findings suggest that the deep 3D CNN method could noninvasively predict the MGMT methylation status with superior performance for sMRI among the three modalities of MRI. It also offers clear guidance in selecting the most effective MR imaging modality for assessing MGMT methylation status.