Improving performance robustness of subject-based brain segmentation software.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Commercial AI tools for brain image analysis face limitations due to insufficient training data and scanner specificity.
• High-quality brain segmentation is crucial for accurate quantification and diagnosis.

Purpose of the Study:

• To improve the performance of personalized brain segmentation software using AI models trained on single-institution data when applied to multicenter datasets.
• To investigate the impact of preprocessing techniques on the generalizability and stability of AI-based brain segmentation.

Main Methods:

• Utilized preindicators of brain white matter (WM) information for preprocessing.
• Trained AI models on data from cognitively normal (CN) individuals at a single center.
• Tested models on multicenter data including CN individuals and Alzheimer disease (AD) patients.
• Employed data augmentation and preindicator-based preprocessing techniques.

Main Results:

• Preindicator-based preprocessing significantly improved segmentation performance (Dice Similarity Coefficient [DSC] 0.8567 vs. 0.7921).
• Standard deviation of WM intensity had a greater impact on performance than average intensity.
• Preprocessing increased the correlation of mean cortical thickness between Atroscan and FreeSurfer (0.7584 to 0.8165).

Conclusions:

• Data augmentation and preindicator-based preprocessing enhance the performance, generalizability, and stability of AI-based brain segmentation software.
• These methods address limitations posed by multicenter data and scanner variability.