3D hemisphere-based convolutional neural network for whole-brain MRI segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Whole-brain segmentation is vital for neuroimaging analysis.
• Current methods using 2D slices or 3D patches with CNNs face limitations in capturing global context and efficiency due to GPU memory constraints.
• Sliding window approaches in CNNs for whole-brain segmentation lack comprehensive spatial information.

Purpose of the Study:

• To introduce a novel 3D hemisphere-based CNN for automatic whole-brain segmentation of T1-weighted MRI scans.
• To overcome the limitations of existing slice-based or patch-based CNN methods.
• To enhance both the accuracy and efficiency of brain structure segmentation.

Main Methods:

• A 3D hemisphere-based CNN approach was developed.
• A localization network was trained to identify hemispheric bounding boxes.
• A segmentation network processed one hemisphere, with the other mirrored across the mid-sagittal plane.

Main Results:

• The proposed method achieved high accuracy, with an overall Dice similarity of 0.84 and a Hausdorff distance of 6.1 mm for 102 brain structures.
• Competitive performance was observed in subcortical segmentation tasks across independent test datasets.
• High consistency in volumetric measurements for intra-session scans was demonstrated.

Conclusions:

• The 3D hemisphere-based CNN offers an efficient and accurate solution for whole-brain segmentation.
• This method effectively addresses the limitations of previous CNN-based approaches.
• The approach shows promise for clinical applications requiring precise brain structure analysis and volumetric measurements.