Improving across-dataset brain tissue segmentation for MRI imaging using transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automated brain tissue segmentation is crucial for MRI data analysis but faces challenges with generalization.
• Existing automated methods, including convolutional neural networks (CNNs), often struggle with variations in MRI acquisition.
• There is a significant need for robust and generalizable brain segmentation tools.

Purpose of the Study:

• To introduce a novel CNN-Transformer hybrid architecture for enhanced brain tissue segmentation in 3D medical imaging.
• To improve the performance and generality of automated segmentation for T1w MRI data.
• To address the limitations of current automated segmentation techniques in handling diverse MRI acquisition properties.

Main Methods:

• Developed a novel CNN-Transformer hybrid architecture (TABS network) for 3D medical image segmentation.
• Evaluated the model's performance on various T1w MRI datasets.
• Validated the model's generality across multi-site datasets with diverse vendors, field strengths, scan parameters, and neuropsychiatric conditions.
• Assessed model reliability using test-retest scans.

Main Results:

• The proposed CNN-Transformer hybrid model demonstrated superior performance on multiple T1w MRI datasets.
• Rigorous validation confirmed the model's excellent generality across diverse multi-site datasets.
• The model exhibited high reliability on test-retest scans, indicating robustness.
• Outperformed benchmark methods in terms of generality and reliability.

Conclusions:

• The novel CNN-Transformer hybrid network offers a robust and generalizable solution for brain tissue segmentation in T1w MRI.
• This method significantly improves upon existing automated segmentation tools, addressing limitations in MRI data variability.
• The TABS network is a valuable tool for quantitative analysis in brain-related MRI studies.