Comparative Analysis of Convolutional Neural Networks and Transformers for Brain Tumor Segmentation: A Study Using BRATS 2021 Dataset
Al-Musalami, Aisha Abduljabbar Ahmed
Al-Musalami, Aisha Abduljabbar Ahmed
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis focuses on a comparative analysis of CNN-based and Transformer-based segmentation models using the BraTS 2021 dataset. The primary hypothesis is that Transformer-based architectures, by virtue of their self-attention mechanisms, can capture long-range dependencies and complex spatial relationships more effectively than traditional CNNs, potentially leading to improved segmentation accuracy. Notably, our experimental results demonstrate that Transformer-based models not only achieve higher Dice Similarity Coefficient (DSC) scores but also converge faster during training. In contrast, while CNN-based models have traditionally been valued for their computational efficiency, our study shows that they may lag in convergence speed when dealing with complex brain tumor segmentation.
Citation
Aisha Abduljabbar Ahmed Al-Musalami, “Comparative Analysis of Convolutional Neural Networks and Transformers for Brain Tumor Segmentation: A Study Using BRATS 2021 Dataset,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Transformers, CNNs, Brain Tumors, BraTS2021, Segmentation