Automated In-line Normalization Procedure for BOLD-CVR Using the Resting-State Temporal Shift with Machine Learning.

Area of Science:

• Neuroimaging
• Vascular Neurology
• Machine Learning in Medicine

Background:

• Cerebrovascular reactivity (CVR) assessment is crucial for estimating hemodynamic impairment.
• Conventional CVR methods face challenges in bilateral disease due to confounding factors.
• Previous work identified healthy voxel signatures from imaging data.

Purpose of the Study:

• To introduce an automated, inline approach for predicting healthy voxel CVR using resting-state BOLD (rs-BOLD) signals.
• To hypothesize accurate prediction of healthy-voxel CVR within 10% error of ground truth.
• To overcome limitations of current CVR assessment in bilateral cerebrovascular disease.

Main Methods:

• Developed machine learning models trained on rs-BOLD baseline, anatomic, and vascular parameters from 22 unilateral steno-occlusive disease (SOD) patients and 20 healthy controls.
• Computed 32 distinct time-delays and 32 maximum correlation values voxelwise, constrained to paired arterial territories.
• Evaluated prediction performance against ground-truth reference CVR values.

Main Results:

• The ensembled model achieved an AUC of 0.81 in predicting unaffected voxels, with a median absolute percent error of 7.28% for median reference CVR.
• Demonstrated significant improvement over naive whole-brain voxel selection (P=0.005) and conventional cerebellar normalization (median error 26.4%).
• In bilateral cases, the model showed reduced candidate voxels but improved delineation of abnormal regions.

Conclusions:

• Feasibility of learning unaffected reference voxel CVR signatures using BOLD-CVR MRI is demonstrated.
• This automated approach extends brain CVR assessment capabilities beyond current limitations, particularly for bilateral disease.
• The method offers a more robust estimation of CVR in complex cerebrovascular conditions.