Harmonization of cerebral blood flow measurements by multi-delay 3D gradient and spin echo, and single-delay 2D echo planar imaging

J. B. Lewis, Chunwei Ying, Michael M. Binkley, Melanie E. Fields, Igor Dedkov, Slim Fellah, Jingyi Zhang, Amy Mirro, Joshua S. Shimony, Yasheng Chen, Jin‐Moo Lee, Andria L. Ford, Hongyu An, Kristin P. GuilliamsmedRxiv 2025 · DOI

Abstract

Our study demonstrates the feasibility of imputing 3D-GRASE multi-post-label delay (PLD) cerebral blood flow (CBF) from a 2D-EPI single-PLD technique. This statistical "translation" method promotes data sharing and harmonization, allowing researchers to combine legacy datasets with modern research to increase statistical power.

The Challenge of Evolving MRI Standards

Measuring Cerebral Blood Flow (CBF) is a cornerstone of neuroimaging, providing critical insights into brain health, particularly in pediatric populations and rare diseases. For years, the gold standard for non-invasive CBF measurement has been pseudo-continuous arterial spin labeling (PCASL). However, the field has recently undergone a technological shift from older "single-delay 2D-EPI" readouts to more sophisticated "multi-delay 3D-GRASE" techniques.

While 3D-GRASE offers superior signal-to-noise ratios and better handling of Arterial Transit Time (ATT)—the time it takes for labeled blood to reach brain tissue—it is not directly comparable to older 2D data. These methods differ in spatial resolution, noise profiles, and underlying physics. This discrepancy creates "data silos," preventing the integration of valuable legacy datasets with contemporary research. To address this, we developed a statistical framework to harmonize these two disparate imaging modalities.

Methodology: A Paired-Data Approach

To build a robust translation model, we conducted a retrospective study involving 50 participants (ages 8–45). This cohort was intentionally diverse, including 17 individuals with Sickle Cell Disease (SCD). Patients with SCD often exhibit significantly higher CBF and lower hemoglobin levels than healthy controls, providing a rigorous "stress test" for our harmonization model across a wide physiological spectrum.

Each participant received both types of PCASL scans in a single session. We compared a 2D Echo Planar Imaging (EPI) sequence using one post-label delay (1PLD) against a 3D Gradient and Spin Echo (GRASE) sequence with five delays (5PLD). The technical differences between these two protocols are summarized below:

Comparison of acquisition and processing parameters for the 2D-EPI and 3D-GRASE sequences.

Instead of applying a simple scaling factor, we employed multiple linear regression to "impute" (predict) the 3D-GRASE values from the 2D-EPI data. Crucially, we included hemoglobin (hb) and age as predictors. Hemoglobin levels directly influence the magnetic properties of blood ($T1_{blood}$), while age serves as a proxy for developmental changes in both blood flow and transit time.

Key Findings: Accuracy and Reliability

The results of the imputation model were highly encouraging. We validated the model using leave-one-out cross-validation (LOOCV) and assessed agreement using the Intraclass Correlation Coefficient (ICC).

  • Gray Matter (GM): The model achieved an adjusted $R^2$ of 0.82 and an ICC of 0.89, indicating excellent agreement. While the older 2D method showed a fixed overestimation bias of approximately 37% in GM, the regression model successfully corrected for this.
  • White Matter (WM): The model achieved an adjusted $R^2$ of 0.78 and an ICC of 0.87. Although the data in white matter was noisier, the fixed bias was minimal (4%).

One of the most significant takeaways was that the error in the imputation model (approximately 9-11%) is comparable to the inherent "test-retest" variability of scanning the same individual twice with the same method. This suggests that statistical conversion is nearly as reliable as a physical measurement.

The Role of Biological Covariates

The study highlighted that biological metadata is just as important as the imaging data itself for successful harmonization. Including hemoglobin levels significantly improved the model's fit, particularly in gray matter. Furthermore, the model successfully captured the relationship between the two methods regarding ATT. The older 2D method often underestimates flow in "watershed" regions with long transit times; the multi-delay 3D method corrects for this, and our regression model effectively bridged that gap.

Conclusion and Future Directions

This study demonstrates that legacy 2D-EPI data can be harmonized with modern 3D-GRASE datasets with high reliability. While the current model focuses on spatially averaged regions of interest (whole gray or white matter), it provides a vital tool for longitudinal studies and research into rare conditions where maximizing sample size is paramount.

Future work will explore whether these models remain robust across different scanner manufacturers (e.g., GE or Philips) and whether they can be applied to smaller, more specific brain subregions. For now, this "translation" method offers a powerful way to break down data silos and advance our understanding of cerebral hemodynamics.

Summary updated

Abstract

Our study demonstrates feasibility to impute 3D-GRASE multi-PLD CBF from a 2D-EPI single-PLD technique, which promotes data sharing and harmonization.
BibTeX
@misc{lewis2025harmonization,
  author = {J. B. Lewis and Chunwei Ying and Michael M. Binkley and Melanie E. Fields and Igor Dedkov and Slim Fellah and Jingyi Zhang and Amy Mirro and Joshua S. Shimony and Yasheng Chen and Jin‐Moo Lee and Andria L. Ford and Hongyu An and Kristin P. Guilliams},
  title = {{Harmonization of cerebral blood flow measurements by multi-delay 3D gradient and spin echo, and single-delay 2D echo planar imaging}},
  year = {2025},
  doi = {10.1101/2025.06.20.25328792},
  howpublished = {medRxiv},
}
Figure