Bertrand Thirion

This paper compares functional alignment methods and template construction strategies in fMRI to improve group-level decoding accuracy and interpretability.

The paper develops TVBase, a platform that projects biomedical knowledge onto a 3D brain atlas to support neuroimaging analysis and hypothesis generation.

NeuroConText improves brain meta-analysis by using contrastive learning and large language models to better retrieve and reconstruct brain activation maps from neuroscience texts.

This study uses fMRI to identify distinct brain networks for language, mathematics, and social knowledge, showing similar activation patterns across written and spoken modalities.

The paper identifies a nonlinear neural code for event probability in the human brain using fMRI, showing distinct tuning curves in prefrontal and parietal cortices.

This study shows that deep convolutional neural networks encode approximate numerosity across diverse objects and scenes, suggesting a potential mechanism for number sense in both biological and artificial systems.

This study benchmarks predictive models for long-term stroke outcomes using a global consortium, revealing mechanistic diversity and key predictors across motor, cognitive, and emotional domains.

This paper evaluates denoising strategies in resting-state fMRI using fMRIPrep and Nilearn, finding that scrubbing combined with global signal regression is effective but may disrupt continuous sampling.

This paper benchmarks five functional alignment methods for inter-subject decoding across four datasets, finding that alignment improves accuracy but the best method depends on context.

This paper uses machine learning on brain imaging and sociodemographic data from the UK Biobank to create proxy measures of psychological constructs, which are as useful or more useful than traditional assessments for predicting health behaviors.

This paper develops a multidomain brain decoder to identify cognitive states from fMRI data with high accuracy across multiple cognitive domains.

This paper introduces MSNR, a multi-scale network regression method that models brain networks at multiple scales, showing improved interpretability and performance in linking brain connectivity to phenotypes.

The paper uses machine learning to rank cognitive domains based on brain data to predict schizophrenia in a multi-site dataset.

The paper compares statistical inference and prediction methods in biomedical data, showing they often diverge in performance and interpretation.