Effect of a consistent reconstruction algorithm on inter-scanner reproducibility in diffusion MRI

Diffusion MRI (dMRI) enables non-invasive characterization of brain microstructure and connectivity. However, multi-center studies face reproducibility challenges due to inter-scanner variability, which arises from differences in hardware, acquisition protocols, and image reconstruction algorithms. While prior harmonization efforts have focused on standardizing protocols and post-processing methods, the impact of using a consistent reconstruction algorithm across scanners on inter-scanner reproducibility remains unexplored.

To evaluate the impact of consistent reconstruction algorithms on cross-vendor, inter-scanner reproducibility in diffusion MRI (dMRI) microstructure and tractography-derived measures.

Identical single-shell dMRI protocols were used on two clinical 3T scanners (Siemens Prisma and GE Premier) using simultaneous multi-slice (SMS) EPI sequences. Five healthy volunteers were scanned twice for capturing within-scanner variability and also on both scanners for computing cross-scanner variability (total of 20 scans). Three MRI image reconstruction methods were assessed: vendor-provided online reconstruction (Product), offline Split slice-GRAPPA (Split-GRAPPA), and offline L1-wavelet regularized SENSE (L1-ESPIRiT). Microstructure measures that were estimated included fiber-specific fractional anisotropy (FA) and mean diffusivity (MD) (from a multi-tensor UKF tractography model) and FA and MD (from a diffusion tensor imaging (DTI) model). Tractography measures included the number of streamlines and the volumetric overlap (weighted Dice coefficient, wDice). Standard error (SE) and wDice were used to evaluate within- and inter-scanner variability. Additional analyses included voxelwise noise estimation using a homomorphic filtering algorithm and bootstrapped quantification of uncertainty in FA/MD using a residual-resampling approach.

Offline Split-GRAPPA significantly reduced the inter-scanner SE of FA in both the multi-tensor and DTI models compared to Product (p-value < 0.001, Wilcoxon rank-sum test). MD values showed similar inter-scanner variability across all reconstruction methods. For tractography measures, the SE in the number of streamlines and wDice values (∼0.8) were similar across reconstruction algorithms. Noise analysis confirmed that Split-GRAPPA achieved the lowest noise levels, as well as consistently lower FA variability. Notably, for both microstructural measures and tractography measures, inter-scanner variability remained significantly higher than within-scanner variability.

Offline Split-GRAPPA reconstruction algorithm reduced inter-scanner variability in FA but not MD. Overall, a consistent reconstruction (with matched acquisition parameters) did not improve inter-vendor reproducibility in dMRI measures or tractography results using other reconstruction methods. These findings highlight the need for further harmonization at the acquisition level (i.e. sequences) to achieve robust cross-vendor comparability in dMRI studies.