Deep learning for segmentation of brain tumors from MRI scans

Abstract

Accurate segmentation of brain tumors from MRI scans is crucial for effective treatment planning. However, manual segmentation remains time-consuming and subjective. Deep learning offers the potential to automate this process, improving diagnostic accuracy and efficiency. This study compares convolutional neural networks (CNNs), transformer-based models, hybrid architectures, and self-supervised approaches for brain tumor segmentation on the BraTS2018 dataset.SegResNet (CNN), SwinUNET-3D (transformer model), and SwinUNeTR (hybrid model) were trained using Dice loss. Self-supervised learning was also explored using the DINO approach with SwinUNet3D. Our research findings underscore the continued effectiveness of CNNs for brain tumor segmentation. Additionally, we understood that while transformers and hybrid models show promising results, they need further optimization. The use of larger datasets, more computational power, and refined self-supervised strategies may lead to significant improvements in model accuracy and robustness.