Automated In-line Normalization Procedure for BOLD-CVR Using the Resting-State Temporal Shift with Machine Learning.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (Y.Z., J.R.P., S. Dehkharghani), Stanford University School of Medicine, Palo Alto, CA, USA; .Department of Radiology (Y.Z., S. Dogra, S. Dehkharghani), New York University Grossman School of Medicine, New York, NY, USA and Department of Radiology (X.W.), Weill Cornell Medical College, New York, NY, USA.
- From the Department of Radiology (Y.Z., J.R.P., S. Dehkharghani), Stanford University School of Medicine, Palo Alto, CA, USA; .Department of Radiology (Y.Z., S. Dogra, S. Dehkharghani), New York University Grossman School of Medicine, New York, NY, USA and Department of Radiology (X.W.), Weill Cornell Medical College, New York, NY, USA. [email protected] [email protected].
Abstract
Cerebrovascular reactivity (CVR) is commonly used to estimate hemodynamic impairment. Conventional use is best suited to unilateral vascular disease, such that CVR can be normalized to reference values from the contralateral hemisphere or to posterior circulation territories; however, major confounds have been identified that leave implementation difficult in more common cases of bilateral disease, even despite common cerebellar normalization. Recently, we reported data-driven identification of candidate <i>healthy</i> voxel signatures learned from contemporaneous imaging data. Here, we introduce an entirely inline, automated approach exploiting the dynamics of resting-state BOLD (rs-BOLD) signal from the BOLD baseline, hypothesizing prediction to within ten percent error relative to ground truth <i>healthy-voxel</i> CVR values. 22 subjects with strictly unilateral intracranial steno-occlusive disease (SOD) underwent 28 CVR studies under pharmacologic provocation using acetazolamide with BOLD-MRI (ACZ-BOLD). Separate affected and unaffected hemispheric masks were segmented to train machine learning models to learn signatures of the unaffected hemisphere using the rs-BOLD baseline, as well as anatomic and vascular parameters. Twenty additional healthy subjects from the Human Connectome Project supplemented training, wherein all voxels were classified <i>normals</i>. 32 distinct time-delays were computed voxelwise, with 32 maximum correlation values constrained to each of 32 paired arterial territories. Performance in prediction of ground-truth reference CVR was computed and compared. The ensembled model achieved AUC of 0.81 in predicting candidate unaffected voxels, demonstrating strong performance in estimation of normal-hemisphere CVR (median absolute percent error [95%CIs] 7.28[3.48-10.34] and 5.61[2.90-9.86] to predict median and mean reference CVR), exhibiting significant improvements over naïve whole-brain voxel selection (P=0.005 and P=0.048, respectively) or conventional cerebellar normalization (26.4[10.1-40.3] median and 27.6[23.7-33.2] mean). In nine bilateral cases assessed to illustrate potential use, the proportion of candidate voxels and corresponding volumes predicted by the ensembled model was significantly lower than in most healthy hemispheres, but yielded subjectively improved delineation of putatively abnormal regions. We demonstrate feasibility of learning unaffected reference voxel CVR signatures for BOLD-CVR MRI. The approach facilitates extension of brain CVR beyond existing constraints in subjects with bilateral disease.