Transformers for Neuroimage Segmentation: Scoping Review

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Automated neuroimaging segmentation is crucial for diagnosing and treating neurological diseases.
• Manual segmentation is labor-intensive and prone to errors.
• Transformers offer a powerful deep learning solution for automated segmentation.

Purpose of the Study:

• To systematically review and assess transformer models used in neuroimaging segmentation.
• To synthesize current literature on transformer applications in the field.

Main Methods:

• Systematic literature search across major databases (Scopus, IEEE Xplore, PubMed, ACM Digital Library) from 2019-2023.
• Inclusion of peer-reviewed journal and conference papers on transformer-based segmentation of human brain imaging data.
• Exclusion of non-neuroimaging data, raw brain signals, and electroencephalogram data.
• Narrative synthesis of extracted data on image modalities, datasets, conditions, models, and metrics.

Main Results:

• 67 out of 1246 publications met inclusion criteria, with a surge in 2022.
• Over two-thirds of studies focused on brain tumor segmentation using Magnetic Resonance Imaging (MRI).
• Hybrid Convolutional Neural Network-Transformer architectures, especially Vision Transformers, were prevalent and showed high performance.
• Dice score was the most common evaluation metric, with studies reporting improved accuracy and ability to capture local/global features.

Conclusions:

• Transformers, particularly hybrid CNN-Transformer models, are increasingly adopted for neuroimaging segmentation, especially for brain tumors.
• Current models demonstrate state-of-the-art performance but face limitations in computational cost and overfitting.
• Diversifying datasets beyond brain tumors is essential for broader clinical applicability.
• Further research is needed to optimize transformer architectures and training for clinical use, potentially revolutionizing brain MRI segmentation.