Improving Non-Invasive Brain Tumor Categorization using Transformers on MRI Data

Recent years have seen a surge in the number of studies utilizing Artificial Intelligence (AI) on Magnetic Resonance Imaging (MRI) to analyze and categorize brain tumors. Despite the advances, most of the existing computer-aided brain tumor classification models are severely limited to smaller datasets of only 4 MRI contrasts: T2, T2/FLAIR, and Tl pre and post-contrast which leads to unsatisfactory performance since the imaging protocols significantly depend on magnetic field strength and acquisition parameters. As a result, this research aims to address the issue by incorporating the most up-to-date Glioma MRI dataset, UCSF-PDGM that includes standardized 3- T three-dimensional preoperative MRI protocol, diffusion MRI, and perfusion MRI. In order to acquire a better computational efficiency while extracting image features both locally and globally, we have presented two Transformer based approach: Swin Transformer and MaxViT-Tiny, to categorize three types of tumors: Astrocytoma, Glioblastoma, and Oligodendroglioma. Considering, Tl and T2 weighted MR images are more eligible to classify brain tumors, we have trained the two models on these imaging protocols. After training and evaluating both the models on performance metrics, we have found out that MaxViTTiny slightly outperforms Swin Transformer in classifying brain tumors with an accuracy of 94.84% on T1-dataset and 98% on T2-dataset; whereas, Swin Transformer achieved 91.05% and 96.97% respectively.