Development and validation of deep learning-based automatic brain segmentation for East Asians: A comparison with...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Accurate brain segmentation is crucial for neurological research and clinical diagnosis.
• Existing automated segmentation tools like Freesurfer may have limitations in diverse populations.
• Deep learning approaches show promise for improving the accuracy and reliability of brain segmentation.

Purpose of the Study:

• To develop and validate a deep learning-based automatic brain segmentation tool (Neuro I) for East Asian populations.
• To compare the performance of Neuro I against Freesurfer using a ground truth.
• To assess the utility of Neuro I for brain volume assessment.

Main Methods:

• Developed Neuro I using a 3D convolutional neural networks (CNNs)-based deep learning algorithm.
• Trained the algorithm on T1-weighted MRI data from 776 healthy Koreans.
• Validated the tool on 30 healthy participants, comparing segmentation accuracy (Dice coefficient) with Freesurfer (ver6.0).

Main Results:

• Neuro I achieved significantly higher Dice coefficient values than Freesurfer.
• Inter-method reliability analysis showed poor to moderate correlations between Freesurfer and Neuro I (ICC: 0.498–0.688).
• Neuro I demonstrated consistent brain volume measurements across different age groups.

Conclusions:

• Neuro I exhibits superior performance in automatic brain segmentation compared to Freesurfer when validated against a ground truth.
• Neuro I is a reliable and potentially more accurate tool for brain volume assessment in East Asian populations.
• The findings suggest Neuro I as a valuable alternative for neuroimaging research and clinical applications.