Deep learning architectures for enhanced multimodal brain tumor segmentation

Abstract

Brain tumors represent some of the most aggressive malignancies, MRI imaging is the main tool used by clinicians for brain tumor diagnosis. Accurate segmentation is essential for clinical diagnosis. Manual methods are labor-intensive and prone to inter-observer variability, motivating automated alternatives. Transformer-based models excel at capturing global contextual dependencies and long-range dependencies across the full image domain through their self-attention mechanism. In contrast, U-Net achieves precise segmentation by combining global context from the encoder with fine local details from skip connections for accurate localization. In this work, we present a comparative analysis of transformer-based and U-Net architectures for multimodal brain tumor segmentation in MRI. We evaluate SegFormer and SwinUNETR models against the U-Net architecture on the Brain Tumor MRI Dataset. A preprocessing pipeline was implemented, the models were trained and evaluated, segmentation performance was quantified using the Dice Score. Experimental results indicate that SegFormer model outperformed SwinUNETR and the U-Net architecture with 0.89 Dice Score average.