Deep Learning for Brain Tumor Radiogenomic Classification: A Multi-Parametric MRI Images

Abstract

In this research, we will evaluate the ability of advanced deep learning models to predict methylation of O6-methylguanine-DNA methyltransferase (MGMT) gene promoter methylation status using the data from mpMRI in a more performative manner than traditional methods. The present study involves usage of a variety of architectures such as 3D Vision Transformers (ViT3D), Xception, ResNet50, EfficientNet-B3, and other models, where different sequences of MRI including T1, T1-C, T2, and FLAIR were used for training. The present study showed that the Xception model has the highest Area Under Curve (AUC) at the test stage was 0.617, indicating that this model's degree of discrimination of positive and negative samples is relatively high. Following closely was the 512 ViT3D model which achieved a AUC of 0.6006, whereas 0.58078 and 0.55817 were recorded respectively for fairly low performing architectures in this setting, ResNet50 and EfficientNet-B3. All these results emphasize the effectiveness of the best architectures, especially Xception, in determining the status of MGMT gene promoter from mpMRI images.