Méthodes d'analyses de données massives de neuro-imagerie

Brain MRI (Magnetic Resonance Image) can provide direct insights into the brain structures, regions or functions associated with behaviour or disorders of the brain. The field of neuroimaging is at a turning point, as several large datasets are becoming available. For example, the UK Biobank has currently released MRI from 40,000+ volunteers and is planning to reach 100,000 MRI images in the next years. Such large data promise a finer understanding of the brain association with behaviour and disorders, though they also raise computational and methodological challenges : how to analyze thousands of individuals and hundreds of thousands of brain measurements. In this context, global measures of association—such as morphometricity, which quantifies the proportion of variance in a trait explained by brain structure—have emerged as important tools for quantifying brain-trait relationships. In addition, prediction of behavioural or clinical traits from brain data is increasingly central to both scientific discovery and translational applications. These approaches demand scalable models and computational strategies that can operate at a reasonable cost. T1w brain MRI images provide a detailed mapping of the grey-matter structure and can be processed using different imaging software. Currently, there are no guidelines to select an image processing pipeline, and there is a growing concern that the choice of image processing could introduce a significant variability in brain representation, contributing to the reproducibility crisis in neuroimaging. This thesis first investigates the impact of different grey-matter processing methods on brain-trait associations. Using a cohort of 42,272 participants from the UK Biobank, we evaluated five high-dimensional grey-matter representations assessing their sensitivity to confounding variables and their ability to relate (via association or prediction) to 19 behavioral and clinical traits. Moreover, to address the limitations of analyses conducted on a single pipeline, we applied a multiverse analytical framework that combines results from different processing pipelines. This approach enabled the identification of novel, robust, and generalisable brain-trait associations. Our analyses demonstrate that results obtained on one processing can be applied to other representations of the grey matter, but that multiverse approaches increase power and maximise prediction accuracy of brain predictors. Further, we introduced a novel approach for estimating morphometricity and grey-matter correlation directly from brain association maps, bypassing the need for individual-level data and enabling cross-trait analyses, not currently possible. We validated our methods using synthetic and real phenotypic data from the UK Biobank. We applied our approach to brain association maps from several cohorts. They revealed novel associations, including a link between maternal smoking around birth and risk of Alzheimer's disease in later life. Overall, this work contributes to the advancement of neuroimaging research by assessing how image processing choices impact both association and prediction, and by offering practical guidelines to enhance the reproducibility of neuroimaging findings. The development of multiverse and ensemble strategies, alongside innovative techniques leveraging brain association maps, establishes a foundation for more robust, powerful, computationally efficient, and reproducible brain-trait analyses of large-scale neuroimaging studies.