Ultra-High Resolution 9.4T Brain MRI Segmentation via a Newly Engineered Multi-Scale Residual Nested U-Net with Gated...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• 9.4T MRI offers unprecedented resolution for brain imaging, crucial for detecting subtle neurological changes.
• Existing segmentation models struggle with the high resolution of 9.4T MRI data.
• There is a need for specialized models to accurately segment 9.4T brain MRIs.

Purpose of the Study:

• To introduce the GA-MS-UNet++, the first deep learning model designed for 9.4T brain MRI segmentation.
• To evaluate the performance of GA-MS-UNet++ against existing leading segmentation models.
• To demonstrate the potential of GA-MS-UNet++ for clinical applications in automated neuroimaging.

Main Methods:

• Developed GA-MS-UNet++, integrating multi-scale residual blocks, gated skip connections, and spatial channel attention.
• Trained and evaluated the model on the UltraCortex 9.4T dataset (12 patients).
• Benchmarked against Attention U-Net, Nested U-Net, VDSR, and R2UNet, using both manual and synthetic ground truth masks.

Main Results:

• GA-MS-UNet++ achieved state-of-the-art performance, with Dice scores of 0.93 (manual) and 0.89 (synthetic).
• Achieved 97.29% accuracy, 90.02% precision, and 94.00% recall across evaluations.
• Demonstrated high correlation (R² = 0.90) in volumetric validation and near-exact alignment with ground truth masks.

Conclusions:

• GA-MS-UNet++ is the first specialized deep learning model for 9.4T brain MRI segmentation.
• The model shows significant potential for clinical use and advancing automated neuroimaging analysis.
• This algorithm provides a powerful tool for high-resolution brain segmentation, overcoming limitations of current models.