Abstract WP265: Aggregate Multi-tiered Normalization to Enhance Detection of Impairment in Dynamic BOLD-CVR for Assessment of Hemodynamic Impairment

Introduction: Cerebrovascular reactivity (CVR) is a common measure of hemodynamic impairment. Conventional use is best suited to unilateral vascular steno-occlusive disease (SOD), such that results can be normalized to contralateral normal hemispheres, limiting applicability for bilateral or ambiguous SOD. Normative CVR atlases with Z-score mapping could allow group reference, but suppress valuable inter-individual baseline variability. We recently developed a machine learning method using custom denoised BOLD temporal shifts during resting-state to learn within-subject normal voxel signatures, built upon a previously validated custom pipeline for dynamic CVR analysis. The following tests the hypothesis that integrated atlas- and subject-based normalization improves identification of CVR impairment. Methods: Twenty-two patients with angiographic unilateral SOD underwent 28 BOLD-CVR studies under provocation with acetazolamide. Images were registered to MNI-152 space, and CVR maps ( Fig 2B ) were generated using a previously validated framework for BOLD-CVR. A healthy voxel CVR atlas ( Fig 1 ) was built from unaffected hemispheres by rescaling each map to normative group baselines. Additionally, a dedicated machine learning model recently developed in our group was trained on structural tissue and BOLD time course (temporal shift and maximal correlation) features to yield auto-normalized, in-subject healthy voxel predictions ( Fig 2C ), enabling individualized relative CVR (rCVR) maps. Lastly, voxel-wise Z-score maps ( Fig 2E ) were generated from integrated atlas and per-subject normalization and varied across thresholds (-2 – 0). For unilateral disease subjects, an asymmetry index (Fig 3A) was used for summarization. Among an additional population of nine bilateral SOD subjects, chi-squared distance (Fig 3B) was used to compare CVR in abnormal and normal voxels. Results: In unilateral SOD, consistently greater delineation and volume of abnormal voxels and greater asymmetry indices were demonstrated by integrated normalization than by atlas-only assessment ( Fig 3A ). Among nine bilateral subjects, chi-squared analysis showed superior discrimination between CVR of abnormal and normal voxels at all thresholds below –1.0, where regions of greater impairment are emphasized ( Fig 3B ), based on Z-score risk map thresholds. Conclusion: Integrated normalization enhances detection of abnormalities in BOLD-CVR, facilitating extension of BOLD-CVR to bilateral SOD.