Cross-scanner and cross-protocol multi-shell diffusion MRI data harmonization: Algorithms and results.

Area of Science:

• Medical Imaging
• Neuroscience
• Computational Biology

Background:

• Cross-scanner and cross-protocol variability in diffusion MRI (dMRI) hinders multi-site clinical studies.
• Harmonization algorithms are crucial for aggregating dMRI data from diverse acquisition settings.
• Existing methods often focus on single b-value data, limiting their application to multi-shell dMRI.

Purpose of the Study:

• To evaluate 19 algorithms for harmonizing multi-shell dMRI data across scanners and protocols.
• To assess the effectiveness of these algorithms in reducing data variability in a benchmark dataset.
• To provide guidance for data harmonization in future multi-site dMRI studies.

Main Methods:

• Evaluation of 19 dMRI harmonization algorithms using a benchmark database.
• Algorithms employed diverse approaches: rotational invariant spherical harmonics, deep neural networks, and hybrid methods.
• Data acquired from the same subjects on two scanners with different gradient strengths and protocols.

Main Results:

• Harmonization algorithms successfully reduced cross-scanner and cross-protocol variability to levels comparable to scan-rescan variability.
• LinearRISH algorithm demonstrated the lowest variability for fractional anisotropy (FA), mean diffusivity (MD), mean kurtosis (MK), and RISH features.
• DIAMOND, SHResNet, DIQT, and CMResNet algorithms showed improved harmonization for return-to-origin probability (RTOP).

Conclusions:

• Computational harmonization is effective in mitigating variability in multi-shell dMRI data.
• Algorithm performance varies depending on the specific diffusion metric being harmonized.
• Findings offer valuable insights for selecting appropriate harmonization strategies in multi-site dMRI research.