Harmonization

Harmonization removes scanner-induced variability from diffusion magnetic resonance imaging (dMRI) datasets using adaptive dictionary learning to preserve biological variability in tissue microstructure metrics.

Key Features:

• Adaptive Dictionary Learning: Uses adaptive dictionary learning to iteratively construct overcomplete dictionaries from multiple datasets simultaneously with an adaptive regularization criterion that separates scanner-induced variability from genuine biological differences.
• No Requirement for Paired Datasets or Spatial Registration: Operates without paired training datasets, spatial registration, or the need to match spatial resolutions across scans.
• Preservation of Biological Variability: Removes scanner-specific confounds while preserving diffusion metrics that reflect tissue microstructure and true biological variation.
• Native-space to Standardized Scanner-space Mapping: Produces mappings that are applied in each subject's native space to transform data into a standardized scanner-space without altering inherent biological information.

Scientific Applications:

• Multicenter Studies: Mitigates scanner-specific confounds to enable pooling of dMRI data across multiple centers and increase statistical power.
• Preservation of Diffusion Metrics: Maintains the effect size of key diffusion metrics for accurate clinical and research analyses.
• Robustness to Unseen Variations: Preserves experimentally introduced modifications such as a free water compartment even when those alterations were not present in the training data.

Methodology:

Learn overcomplete dictionaries from all available datasets concurrently; adaptively adjust dictionaries using an adaptive regularization criterion to remove scanner-attributable variability; apply the resulting mappings in each subject's native space to transform data into a standardized scanner-space while preserving biological information.