Superpixel-ComBat multi-site harmonisation of unpaired T1W MRI data in Huntington’s disease

Le résumé fourni par la source

The significance of implementing magnetic resonance imaging (MRI) harmonisation methods has become increasingly pronounced in the context of large-scale multi-site clinical trials, especially when employing neuroimaging to assess biomarker responses to novel treatments. A critical challenge arises from scanner-related variability, stemming from variations in hardware design, field strengths, and imaging protocols across sites. This variability impacts image quality, affecting factors like noise, contrast, and signal inhomogeneity, which ultimately influences downstream analyses.1 Without a standardised imaging protocol across different sites, data variability and confounding factors can compromise the reliability and interpretability of study outcomes.2 Various complex statistical methods have been proposed to address scanner discrepancies at multiple levels of an imaging analysis pipeline in a single mega-analysis. Feature-level harmonisation involves applying techniques like ComBat to quantitative features extracted from images, such as cortical thickness or fractional anisotropy.2 While ComBat has several extensions for feature-level neuroimaging data,2,3 the impact of scanner differences can vary across individual voxels within an image.4 This understanding has driven the development of improved image-level harmonisation methods, focusing on individual voxels in MRI .5–7 The goal is to normalise images acquired from different scanners to reduce variations in characteristics like contrast and signal-to-noise ratio (SNR), making them appear as if they were all captured with a single, consistent imaging protocol.8 Examples include RAVEL, designed to remove between-scan unwanted variation at the voxel level,9 and MISPEL, a supervised voxel-level multi-scanner method requiring paired data.10 Notably, ComBat has also been adapted for image-level harmonisation, using individual voxels for diffusion tensor imaging (DTI)11 and structural imaging through superpixel parcellation for improved interpretability of image harmonisation (SP-ComBat).12 The application of such advanced harmonisation techniques is especially pertinent for understanding complex neurodegenerative conditions like Huntington's disease (HD), where subtle brain changes are key to tracking progression and treatment efficacy. Integration of structural MRI data across multiple HD studies is needed to enable more comprehensive characterisation of neuroimaging biomarkers across the entire disease course. Achieving this remains challenging due to scanner-related variability in structural MRI data at the image-level, reducing statistical power and generalisability of research outcomes, even with standardised acquisition protocols.12 Voxel-based morphometry (VBM), an unbiased whole-brain imaging analytical technique, allows for data-driven identification of regional brain differences13 and has been extensively applied in HD cohorts, consistently showing significant striatal grey-matter loss decades before clinical motor onset and wide spread white-matter loss as the disease progresses.14–19 However, intensity signal differences between scanners may influence downstream VBM results2, and conventional statistical correction methods may be insufficient to correct for the differences.2 While meta-analyses are commonly used robust methods for integrating MRI data across various scanners and reducing site-specific biases, they do not directly harmonise the raw voxel intensity values that differ between scanners due to hardware and software variations.20 Super-pixel ComBat (SP-ComBat) functions as an analytic pipeline to integrate ComBat modelling with a 3D superpixel parcellation algorithm for image-level harmonisation.12 It is specifically designed to estimate and quantify the parametric location and scale effects in the relative signal distributions of clusters in T1-weighted (T1W) images. The SP-ComBat method led to post-harmonised T1W images exhibiting more consistent signal profiles and similar contrasts across scanners qualitatively. The SNR in white matter (WM) and CSF substantially improved after harmonization; while grey matter (GM) SNR was not significantly improved, it became more consistent across scanners. The method achieved a significant reduction in the variation of volumetric measures of brain tissues, specifically reducing the Coefficient of Variation (CV) of GM volume by an average of 40.3% compared to raw images.12 This reflects a substantial improvement in overall image quality and a reduction in technical variability, aligning with the goal of rendering image characteristics indistinguishable between scanners. However, SP-ComBat, like many other advanced image-level harmonisation techniques, relies on the availability of paired data from traveling subjects (i.e., the same individuals scanned across different scanners to harmonise) for reliable estimation of inter-scanner differences. This requirement of ~16–20 traveling subjects21 makes its application to unpaired retrospective datasets, which are common in real-world multi-site studies, underexplored. This study will directly address this limitation by developing and evaluating a novel pseudo-pairing approach for SP-ComBat, enabling a mega-analysis with unpaired structural MRI data in HD. Since obtaining travelling subjects retrospectively is not feasible, we propose to leverage a novel pseudo-pairing approach using demographically matched healthy controls to generate pseudo-travelling subjects to learn inter-scanner differences. While identifying 16–20 demographically matched controls per scanner across diverse cohorts is challenging, identifying 4-8 per scanner appears feasible. One strategy we will explore is to select 4 well-matched controls per scanner and apply bootstrapping within each scanner group to generate 25 synthetic scans for pseudo-pairing: enough pseudo-pairs for reliable estimation of inter-scanner differences. These pseudo-paired samples can then be used to estimate SP-ComBat parameters for image harmonisation. Alternatively, we will also investigate pseudo-pairing the required 16-20 available controls across scanners without strict demographic matching. Comparing these strategies will allow us to assess the trade-offs between demographic alignment and statistical robustness in parameter estimation for unpaired data. If the harmonisation negatively impacts the imaging data, we will focus on optimising a VBM meta-analysis approach as an alternative harmonisation strategy to statistically remove scanner-effect bias without altering the original images. The current study, therefore, aims to adapt and evaluate SP-ComBat for use with unpaired structural MRI data in HD and establish a harmonised VBM protocol. The goal is to enable valid VBM analyses, which can investigate neuroanatomical volume measures in combination with biomarkers of interest. This will ultimately support the development of robust imaging-biomarker pipelines by enhancing the consistency of biological signals across scanners, aiding in the development of targeted therapeutic interventions and providing valuable markers for disease monitoring and prognosis.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.