Deep whole brain segmentation of 7T structural MRI

Area of Science:

• Neuroimaging
• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medicine

Background:

• 7 Tesla (7T) MRI offers enhanced resolution and contrast for human brain imaging, crucial for detailed analysis.
• Whole brain segmentation is vital for region-specific analysis and provides spatial/volumetric data for neuroimaging pipelines.
• Spatially Localized Atlas Network Tiles (SLANT), a 3D convolutional neural network (CNN), excels at whole brain segmentation from 3T MRI but struggles with 7T MRI due to image inhomogeneity.

Purpose of the Study:

• To adapt the SLANT tool for accurate whole brain segmentation of structural 7T MRI scans.
• To overcome challenges posed by inhomogeneous image contrast in 7T MRI data.
• To improve the reliability and accuracy of brain segmentation at ultra-high field MRI.

Main Methods:

• Whole brain segmentation from 3T MRI was registered to 7T MRI.
• This registered 3T segmentation data was used to finetune the SLANT CNN model.
• The finetuned SLANT pipeline was evaluated on structural 7T MRI data.

Main Results:

• The mean relative difference in label volumes decreased to ~8.43% with the finetuned SLANT on 7T MRI data.
• This represents a significant improvement over the ~15.13% difference observed with standard SLANT between 3T and 7T scans.
• The Dice similarity coefficient between 3T and finetuned 7T segmentations increased from 0.79 to 0.83 (p<0.01).

Conclusions:

• Finetuning SLANT using registered 3T MRI data is a viable strategy for improving whole brain segmentation on 7T MRI.
• This approach enhances the accuracy and consistency of segmentation for high-resolution 7T structural neuroimaging.
• The finetuned SLANT pipeline facilitates more reliable brain region analysis using ultra-high field MRI data.