Effect of Different MRI Modalities on Brain Tumor Segmentation Performance: A Study With BraTS2020 and 2D U-Net Convolutional Neural Network

Gliomas are the most common form of primary brain malignancy in adults, with an incidence of approximately 5 cases per 100,000 inhabitants annually, and are known for their high morbidity and mortality. Glioblastomas, the most prevalent and aggressive form of gliomas, exhibit significant heterogeneity, including subregions such as peritumoral edema, necrotic foci, and active tumor. Effective diagnosis and treatment depend on accurate segmentation of these subregions. This study compares various combinations of imaging modalities for glioma segmentation using the 2D U-Net convolutional neural network (CNN) and the BraTS2020 dataset. Among the combinations tested, the T1ce_FLAIR model, which integrates T1-contrast-enhanced (T1CE) and Fluid-Attenuated Inversion Recovery (FLAIR) sequences, achieved the highest Dice coefficients: 84.81% for peritumoral edema, 76.99% for enhancing tumors, and 72.25% for necrotic regions. This combination provided the most comprehensive tumor representation and demonstrated superior performance compared to other combinations of imaging modalities.