A Data-Driven Normalization Framework for Subject-Specific Cerebrovascular Reactivity Assessment in Cerebrovascular...

Area of Science:

• Neuroimaging
• Medical Physics
• Machine Learning

Background:

• Acetazolamide-augmented BOLD-MRI is used for cerebrovascular reactivity (CVR) assessment.
• High intersubject variability in baseline CVR complicates interpretation.

Purpose of the Study:

• To develop an integrated normalization framework for BOLD-MRI CVR assessment.
• Combine machine learning for healthy voxel identification with atlas-based rescaling.
• Improve subject-specific baseline rescaling for enhanced interpretation.

Main Methods:

• Retrospective study of 35 patients with unilateral or bilateral steno-occlusive disease.
• Utilized T1-weighted MPRAGE and gradient-echo EPI at 3T.
• Trained machine learning models to identify healthy voxels based on resting-state BOLD features.
• Constructed a healthy reference CVR atlas and rescaled individual CVR maps.
• Compared integrated normalization with atlas-based normalization using various statistical metrics.

Main Results:

• Integrated normalization significantly decreased coefficient of variation in normal hemispheres, improving inter-subject stability.
• Lesion contrast-to-noise ratio and ROC AUC showed improvement across impairment thresholds.
• Asymmetry index increased significantly at higher risk thresholds, indicating enhanced discriminability.
• Chi-square analysis revealed significantly increased separation between normal and abnormal CVR distributions.

Conclusions:

• The proposed integrated normalization framework enhances the stability and discriminability of acetazolamide-augmented BOLD-CVR.
• This method improves the interpretability of CVR maps for detecting cerebrovascular impairment.
• The approach offers a more robust assessment of cerebrovascular reactivity in clinical settings.