Harmonizing different diffusion MRI acquisitions

White matter changes are increasingly implicated in neurological disease progression, and diffusion weighted magnetic resonance imaging (DW-MRI) has been included in many national-scale studies. Yet, quantitative investigation of DW-MRI data is hindered by a lack of consistency due to variation in acquisition protocols, sites, and scanners. DW-MRI enables quantification of brain microstructure and facilitates structural connectivity mapping. Substantial recent progress has been made with calibration and harmonization to reduce inter-subject variance and improve interpretability of computed measures. Yet, the fundamental challenge remains that clinical application of DW-MRI (as currently implemented) is confounded by inter-scanner and inter-site effects. There is thus a strong need to harmonize diffusion MRI data to allow reliable combination of datasets provided by different imaging sites in order to increase statistical power and sensitivity of research and clinical studies.

However, different sites have different scanners and associated hardware with different image acquisition settings, which may lead to differences in quantitative results and interpretation. Specifically, there is a need to harmonize preprocessing of diffusion MRI datasets in order to ensure similar quantitative metrics are derived from each site, including (1) voxel-wise microstructure measures, (2) features of white matter fiber bundles, and (2) connectomics measures. In this challenge, participants are provided raw data from two scanners, with two different acquisition protocols, and asked to preprocess the data in order to minimize scanner differences while retaining biological variation (i.e., maximize intraclass correlation coefficient with the scanners as the rater).

This challenge builds off the successful SuperMUDI (MICCAI CDMRI 2020) and MUSHAC (MICCAI CDMRI 2017/2018) challenges. The key innovations are (1) we assess bundles and tractography and connectomics in the context of harmonization for the first time, (2) we assess connectomics in the context of harmonization for the first time, and (3) we have 10x additional subjects over MUSHAC and 100x over SuperMUDI. Additionally, the data that form the basis of this challenge represent a difficult clinical scenario for  harmonization and are part of a much larger twins study, which could provide rich context for continuing validation / extension of this challenge's findings. Given a fixed [post-processing] pipeline (tractography, bundle extraction, and connectome construction), our challenge seeks a pre-processing harmonization method that generates consistent pipeline outputs between different scanners.