ON-Harmony: A multi-site, multi-modal travelling-heads resource for brain MRI harmonisation with integration of UK Biobank scanners

ON-Harmony: A multi-site, multi-modal travelling-heads resource for brain MRI harmonisation with integration of UK Biobank scanners Shaun Warrington1, Andrea Torchi1, Olivier Mougin2, Jon Campbell3, Asante Ntata1,4, Martin Craig1, Stephania Assimopoulos1, Fidel AlfaroAlmagro3, Stephen M Smith3, Adam J Lewandowski5,6, Karla L Miller3, Mark Jenkinson3,7, Paul S Morgan1, and Stamatios N Sotiropoulos1 1Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham, UK. 2Sir Peter Mansfield Imaging Centre, School of Physics, University of Nottingham, UK. 3Oxford Centre for Integrative Neuroimaging, University of Oxford, UK. 4National Physical Laboratory, UK. 5Nuffield Department of Population Health, University of Oxford, UK. 6UK Biobank Ltd, UK. 7Australian Institute for Machine Learning (AIML), School of Computer and Mathematical Sciences, The University of Adelaide, Australia. Introduction: MRI quantifiability is hindered by non-biological sources of variability, e.g scanner hardware/software1–3. Several approaches aim to standardise/harmonise acquisition and processing4–6, but lack of harmonisation is an open challenge. We ran one of the most comprehensive, freely accessible, multi-modal travelling heads studies, ON-Harmony7–9 (Fig1): 20 subjects, each scanned in up to 8 3T scanners, 3 vendors (Siemens/Philips/GE) and 5 modalities (T1w/ T2w/dMRI/swMRI/fMRI), plus within-scanner/ within-subject repeats, enabling within-scanner, between-scanner and between-subject variability to be mapped across multi-modal imaging-derived phenotypes (IDPs). We have recently scanned 11 of the subjects at the Stockport UK Biobank (UKB) imaging centre (Reading UKB centre also planned), enabling linkage of UKB populationlevel imaging data with various clinical scanners. Here, we showcase the data and reuse scenarios for assessing harmonisation efficacy. Methods: ON-Harmony consists of 2 primary phases (10 subjects each), plus the UKB extension. Acquisition protocols were aligned with the UKB imaging study10, while respecting best practices and hardware limitations (i.e. parameters not simply nominally-matched)8. Each subject was scanned in at least 6 different scanners (out of a collection of 8 scanners) and 9 subjects had 5 additional within-scanner repeats. 11 participants were re-scanned at the UKB centre (1 subject with 5 within-scanner repeats). All data underwent quality control through visual inspection and then using MRIQC11 (T1w/T2w/fMRI) and eddyQC12 (dMRI). Data were processed with a modified version8 of the UKB pipeline13. Hundreds of IDPs were derived for each session, allowing us to quantify IDP variability. Results: Fig2a shows a subject’s raw data, depicting sessions across scanners (columns) and modalities (rows). We assessed between-session IDP similarity for within-scanner, between-scanner, between-subject pools (Fig2b). Between-subject variability has little overlap with scan-rescan variability, however, it overlaps quite substantially with between-scanner variability. We explored how ON-Harmony can be used to assess harmonisation efficacy, e.g. ComBat14,15 (explicit harmonisation). Pre/post harmonisation between-scanner variability was compared to within-scanner variability (Fig3) for dMRI tract-wise FA measures. Reductions in between-scanner variability following harmonisation were revealed but did not match the within-scanner baseline. ON-Harmony can also be used to assess pipeline/tool generalisability across scanners (implicit harmonisation). Fig4 shows generalisability of tractography for FSL-XTRACT16. For each scanner, subject-averaged tract maps were obtained and correlated against a UKB atlas, with moderate-high correlation values across all scanners and generally consistent trends across vendors/phases, although with some exceptions (GE-A was a low gradient system with a single-shell protocol). Such comparisons showcase how ON-Harmony can be used to assess susceptibility of tools/pipelines to between-scanner effects. Conclusion: We have presented a comprehensive harmonisation resource (ON-Harmony) for multimodal neuroimaging data, based on a travelling-heads paradigm. ON-Harmony is openly released, freely available9 and can be used to assess harmonisation efficacy and to develop new vendor agnostic tools/pipelines. The novel UK Biobank extension will enable direct linkage of a representative sets of scanners from all vendors with one of the largest population-level studies. References: [1] Pinto et al. Front. Neurosci. 14 (2020). [2] Han et al. NeuroImage 32:180 (2006). [3] Friedman et al. Hum. Brain Mapp. 29:958 (2008). [4] Potvin et al. NeuroImage Clin. 24:101943 (2019). [5] Chen et al. NeuroImage Clin. 24:101943 (2019). [6] Layton et al. Magn. Reson. Med. 77:1544 (2017). [7] Warrington et al. Sci. Data 12:609 (2025). [8] Warrington et al. Imaging Neurosci. (2023). [9] ON-Harmony, https://openneuro.org/datasets/ds004712/ versions/2.0.1. [10] Miller et al. Nat. Neurosci. 19:1523 (2016). [11] Esteban et al. PLOS ONE 12:e0184661 (2017). [12] Bastiani et al. NeuroImage 184:801 (2019). [13] Alfaro-Almagro et al. NeuroImage 166:400 (2018). [14] Fortin et al. NeuroImage 167:104 (2018). [15] Fortin et al. NeuroImage 161:149 (2017). [16] Warrington et al. NeuroImage 217: 116923 (2020).