Multi-class Glioma Classification from MRI Images Using 3D Convolutional Neural Networks

Abstract

AbstractGlioma is a prevalent and deadly form of brain tumor. Most of the existing models have used 2D MRI slices for general tumor classification. In this paper, a 3D convolutional neural network has been proposed to automatically classify gliomas namely, astrocytoma, oligodendroglioma, and glioblastoma from MRI images. MRI modalities like T1, T1ce, FLAIR, and T2 are used for glioma classification. This step is essential for diagnosis and treatment planning for a patient. Manual classification is costly, time-consuming, and human-error prone. So, there is a need for accurate, robust, and automatic classification of glioma. Convolution neural networks have achieved state-of-the-art accuracy in many image processing classification tasks. In this work, 3D and 2D CNN models have been studied with multiple (T1, T1ce, FLAIR, and T2) and single (T1ce) MRI image modalities as input. The effect of segmentation of glioma on the classification accuracy has also been studied. The CNN models have been validated on the publicly available CPM-RadPath2019 dataset. It has been observed that the 3D CNN with segmented glioma along with multiple MRI modalities (T1, T1ce, FLAIR, and T2) has achieved an overall accuracy of 75.45%.KeywordsGlioma/Brain tumor classification3D convolutional neural networkSegmentationMRI images