Assessing glioma grading with self-attention: comparative analysis of the diagnostic potential of different MRI sequences

Gliomas are a type of primary brain tumour which exhibit high variability in terms of histological characteristics and clinical behaviour. Accurate classification of gliomas plays a crucial role in prognosis, treatment planning, and patient management. We propose a novel approach for glioma classification using a self-attention mechanism, demonstrating promising results in capturing intra-slice dependencies within 3D volumetric MRI data due to its ability to capture dependencies in sequential data. Additionally, our study offered a unique comparison of the performances of different MRI sequences on glioma grading. We trained and evaluated our model on 365 MRI scans consisting of four MRI contrasts (T1, T2, T1ce, Flair) of the BraTS-19 dataset. Our findings reveal that self-attention performs well in classifying gliomas into HGG and LGG with T1-weighted sequences comparatively performing best in classifying gliomas with an overall accuracy of 95.52% and an AUC ROC score of 0.92. This study contributes to advancing glioma classification methodologies, highlighting the potential of self-attention mechanisms in capturing complex dependencies within MRI data and demonstrating the importance of selecting appropriate MRI sequences for accurate glioma grading, with implications for improving clinical decision-making in glioma management.