Unifying Genetics and Imaging: MRI-Based Classification of MGMT Genetic Subtypes Using Visual Transformers
I. Babar, Asjid Tahir, Arslan Haider, Hafeez Ur Rehman, Moutaz Alazab · 2023
Brain cancer can take many forms, but glioblastoma (GBM) is one of the most aggressive. To treat it effectively, doctors need to know the genetic subtype of a specific part of the tumor called the O-6-methylguanine-DNA-methyltransferase (MGMT) promoter. This information is critical for making treatment decisions and predicting outcomes. However, traditional methods for predicting MGMT status have significant limitations. One of the main reasons for this poor performance is that these methods fail to capture small variations in MRI signatures that are essential for accurately predicting MGMT promoter methylation status. In this work, we propose a two-part approach to predict the MGMT methylation status. First, we use autoencoders to extract features from the MRI images, capturing intricate spatial patterns and representations. Then, we feed these encoded features into a visual transformer architecture to learn long-range dependencies and contextually significant relationships among image regions. This innovative fusion of autoencoders and visual transformers allows us to harness both local and global information present in the MRI data. We utilized a dataset consisting of 1010 3D MRI images from patients with GBM. Half of the patients had MGMT promoter methylation, while the other half did not. The model achieved impressive results, including an overall accuracy score of 0.904, recall score of 0.902, precision score of 0.914, F1 score of 0.906, and specificity score of 0.906. These metrics far surpass the accuracy of conventional imaging techniques.