Benoit Dufumier
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Benoit Dufumier has a total of 40 co-authors. Here are the top co-authors by number of shared publications:
Edouard Duchesnay
Université Paris-Saclay, CEA, CNRS, UMR9027 Baobab, NeuroSpin, Saclay, France
6 shared publicationsAntoine Grigis
Université Paris-Saclay, CEA, CNRS, UMR9027 Baobab, NeuroSpin, Saclay, France
5 shared publications- Jean-François Mangin
Université Paris-Saclay
4 shared publications Pietro Gori
Universit é Paris-Saclay CEA CNRS UMR9027 Baobab NeuroSpin Saclay France
4 shared publications- Joël Chavas
Université Paris-Saclay, CEA, CNRS, Neurospin, UMR 9027 BAOBAB, GAIA Laboratory, Gif-sur-Yvette, France
3 shared publications Julie Victor
NeuroSpin CEA Saclay Universit é Paris-Saclay France
3 shared publications
Latest Publications
- 7 April 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Apprentissage des représentations en neuroimagerie : transfert de connaissance à partir de larges jeux de données contrôles vers de petites cohortes cliniques
Representation learning in neuroimaging : transferring from big healthy data to small clinical cohorts: La physiopathologie des maladies mentales telles que la schizophrénie et le trouble bipolaire est encore mal comprise, cependant l'émergence de grandes bases de données transdiagnostiques d'images cérébrales offre une occasion unique d'étudier les signatures neuroanatomiques de ces maladies.Le développement de modèles d'apprentissage profonds pour l'imagerie médicale a ouvert la voie à des applications complexes comme la segmentation d'images. Néanmoins, l'applicabilité de telles méthodes aux problèmes de prédiction à l'échelle individuelle à partir d'IRM anatomique reste encore inconnue. Dans cette thèse, nous étudions d'abord la performance des réseaux de neurones actuels en fonction de la quantité de données disponibles. Nous comparons ces performances avec les modèles linéaires régularisés ainsi que les machines à vecteurs de support avec noyau. Nous constatons un problème de sur-ajustement important sur les jeux de données cliniques ainsi qu'une courbe d'apprentissage similaire aux modèles linéaires pour les tailles d'échantillon actuellement accessible en recherche clinique. Nous montrons que cet effet de sur-ajustement est en partie dû au biais induit par les scanners IRM et les protocoles d'acquisition (effet site).Ainsi, nous proposons une nouvelle solution d'apprentissage des représentations sur de grands jeux de données multi-site d'imagerie de la population saine, basée sur l'apprentissage auto-supervisé par contraste. En transférant ces connaissances à de nouveaux jeux de données cliniques, nous démontrons une amélioration des performances de classification et une plus grande robustesse à l'effet site. Par ailleurs, nous fournissons des garanties théoriques de généralisation de ces modèles pour les tâches de classification.Enfin, pour une meilleure reproductibilité et comparaison des modèles profonds en neuroimagerie, nous introduisons un nouveau jeu de données multi-site à large échelle: OpenBHB. Cette base de données est spécialement conçue pour la prédiction de l'âge cérébrale (tâche supervisée) ainsi que la suppression de l'effet site dans les représentations des modèles profonds. Nous proposong également un défi, accessible en ligne, pour l'apprentissage des représentations avec OpenBHB ainsi qu'une nouvelle méthode pour évaluer le biais dans les représentations des modèles soumis.
- 7 April 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Apprentissage des représentations en neuroimagerie : transfert de connaissance à partir de larges jeux de données contrôles vers de petites cohortes cliniques
Representation learning in neuroimaging : transferring from big healthy data to small clinical cohorts: La physiopathologie des maladies mentales telles que la schizophrénie et le trouble bipolaire est encore mal comprise, cependant l'émergence de grandes bases de données transdiagnostiques d'images cérébrales offre une occasion unique d'étudier les signatures neuroanatomiques de ces maladies.Le développement de modèles d'apprentissage profonds pour l'imagerie médicale a ouvert la voie à des applications complexes comme la segmentation d'images. Néanmoins, l'applicabilité de telles méthodes aux problèmes de prédiction à l'échelle individuelle à partir d'IRM anatomique reste encore inconnue. Dans cette thèse, nous étudions d'abord la performance des réseaux de neurones actuels en fonction de la quantité de données disponibles. Nous comparons ces performances avec les modèles linéaires régularisés ainsi que les machines à vecteurs de support avec noyau. Nous constatons un problème de sur-ajustement important sur les jeux de données cliniques ainsi qu'une courbe d'apprentissage similaire aux modèles linéaires pour les tailles d'échantillon actuellement accessible en recherche clinique. Nous montrons que cet effet de sur-ajustement est en partie dû au biais induit par les scanners IRM et les protocoles d'acquisition (effet site).Ainsi, nous proposons une nouvelle solution d'apprentissage des représentations sur de grands jeux de données multi-site d'imagerie de la population saine, basée sur l'apprentissage auto-supervisé par contraste. En transférant ces connaissances à de nouveaux jeux de données cliniques, nous démontrons une amélioration des performances de classification et une plus grande robustesse à l'effet site. Par ailleurs, nous fournissons des garanties théoriques de généralisation de ces modèles pour les tâches de classification.Enfin, pour une meilleure reproductibilité et comparaison des modèles profonds en neuroimagerie, nous introduisons un nouveau jeu de données multi-site à large échelle: OpenBHB. Cette base de données est spécialement conçue pour la prédiction de l'âge cérébrale (tâche supervisée) ainsi que la suppression de l'effet site dans les représentations des modèles profonds. Nous proposong également un défi, accessible en ligne, pour l'apprentissage des représentations avec OpenBHB ainsi qu'une nouvelle méthode pour évaluer le biais dans les représentations des modèles soumis.
- Open Access
- 2 March 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 203, 78-84
Robust brain age estimation from structural MRI with contrastive learning
- 2 March 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 203, 78-84
Open AccessRobust brain age estimation from structural MRI with contrastive learning
- Open Access
- 7 February 2026
- Published by Springer Nature in Journal of Neural Transmission
Decoding cortical folding with deep learning: toward neurodevelopmental biomarkers of psychiatric disorders
Cortical folding is primarily determined during prenatal and early postnatal brain development and remains relatively stable throughout life. Accumulating evidence suggests that psychiatric disorders, including schizophrenia (SCZ), Bipolar Disorder (BD), and Autism Spectrum Disorder (ASD), result from intricate interactions between early neurodevelopmental disruptions and subsequent environmental influences. Therefore, cortical folding patterns are promising candidates as stable imaging biomarkers reflecting the neurodevelopmental component of psychiatric conditions. This study aims to demonstrate that deep learning can be used to learn meaningful representations of cortical folding patterns that enable the individual-level prediction of major psychiatric disorders, namely SCZ, BD, and ASD. We introduce a dedicated deep learning architecture leveraging self-supervised pre-training on large datasets of the general population to extract subject-level representations of cortical folding from structural magnetic resonance imaging (MRI) data. We demonstrate that the learned representations allow significant single-subject prediction of clinical status for SCZ, BD, and ASD. The proposed folding-based representation presents a novel approach for identifying imaging biomarkers related to the neurodevelopmental origins of psychiatric disorders. It opens the possibility of inferring early-life brain alterations from adult MRI data, offering potential tools for stratification and precision psychiatry.
- 7 February 2026
- Published by Springer Nature in Journal of Neural Transmission
Open AccessDecoding cortical folding with deep learning: toward neurodevelopmental biomarkers of psychiatric disorders
Cortical folding is primarily determined during prenatal and early postnatal brain development and remains relatively stable throughout life. Accumulating evidence suggests that psychiatric disorders, including schizophrenia (SCZ), Bipolar Disorder (BD), and Autism Spectrum Disorder (ASD), result from intricate interactions between early neurodevelopmental disruptions and subsequent environmental influences. Therefore, cortical folding patterns are promising candidates as stable imaging biomarkers reflecting the neurodevelopmental component of psychiatric conditions. This study aims to demonstrate that deep learning can be used to learn meaningful representations of cortical folding patterns that enable the individual-level prediction of major psychiatric disorders, namely SCZ, BD, and ASD. We introduce a dedicated deep learning architecture leveraging self-supervised pre-training on large datasets of the general population to extract subject-level representations of cortical folding from structural magnetic resonance imaging (MRI) data. We demonstrate that the learned representations allow significant single-subject prediction of clinical status for SCZ, BD, and ASD. The proposed folding-based representation presents a novel approach for identifying imaging biomarkers related to the neurodevelopmental origins of psychiatric disorders. It opens the possibility of inferring early-life brain alterations from adult MRI data, offering potential tools for stratification and precision psychiatry.
- Open Access
- 6 June 2024
- Published by Elsevier in NeuroImage
- Vol. 296, 120665
Exploring the potential of representation and transfer learning for anatomical neuroimaging: Application to psychiatry
- 6 June 2024
- Published by Elsevier in NeuroImage
- Vol. 296, 120665
Open AccessExploring the potential of representation and transfer learning for anatomical neuroimaging: Application to psychiatry
- Open Access
Exploring the potential of representation and transfer learning for anatomical neuroimaging: application to psychiatry
The perspective of personalized medicine for brain disorders requires efficient learning models for anatomical neuroimagingbased prediction of clinical conditions. There is now a consensus on the benefit of deep learning (DL) in addressing many medical imaging tasks, such as image segmentation. However, for single-subject prediction problems, recent studies yielded contradictory results when comparing DL with Standard Machine Learning (SML) on top of classical feature extraction. Most existing comparative studies were limited in predicting phenotypes of little clinical interest, such as sex and age, and using a single dataset. Moreover, they conducted a limited analysis of the employed image pre-processing and feature selection strategies. This paper extensively compares DL and SML prediction capacity on five multi-site problems, including three increasingly complex clinical applications in psychiatry namely schizophrenia, bipolar disorder and Autism Spectrum Disorder (ASD) diagnosis. To compensate for the relative scarcity of neuroimaging data on these clinical datasets, we also evaluate three pre-training strategies for transfer learning from brain imaging of the general healthy population: self-supervised learning, generative modelling and supervised learning with age. Overall, we find similar performance between randomly initialized DL and SML for the three clinical tasks and a similar scaling trend for sex prediction. This was replicated on an external dataset. We also show highly correlated discriminative brain regions between DL and linear ML models in all problems. Nonetheless, we demonstrate that self-supervised pre-training on large-scale healthy population imaging dataset (N ≈10k), along with Deep Ensemble, allows DL to learn robust and transferable representations to smaller-scale clinical datasets (N ≤ 1k). It largely outperforms SML on 2 out of 3 clinical tasks both in internal and external test sets. These findings suggest that the improvement of DL over SML in anatomical neuroimaging mainly comes from its capacity of learning meaningful and useful abstract representations of the brain anatomy, and it sheds light on the potential of transfer learning for personalized medicine in psychiatry.
Open AccessExploring the potential of representation and transfer learning for anatomical neuroimaging: application to psychiatry
The perspective of personalized medicine for brain disorders requires efficient learning models for anatomical neuroimagingbased prediction of clinical conditions. There is now a consensus on the benefit of deep learning (DL) in addressing many medical imaging tasks, such as image segmentation. However, for single-subject prediction problems, recent studies yielded contradictory results when comparing DL with Standard Machine Learning (SML) on top of classical feature extraction. Most existing comparative studies were limited in predicting phenotypes of little clinical interest, such as sex and age, and using a single dataset. Moreover, they conducted a limited analysis of the employed image pre-processing and feature selection strategies. This paper extensively compares DL and SML prediction capacity on five multi-site problems, including three increasingly complex clinical applications in psychiatry namely schizophrenia, bipolar disorder and Autism Spectrum Disorder (ASD) diagnosis. To compensate for the relative scarcity of neuroimaging data on these clinical datasets, we also evaluate three pre-training strategies for transfer learning from brain imaging of the general healthy population: self-supervised learning, generative modelling and supervised learning with age. Overall, we find similar performance between randomly initialized DL and SML for the three clinical tasks and a similar scaling trend for sex prediction. This was replicated on an external dataset. We also show highly correlated discriminative brain regions between DL and linear ML models in all problems. Nonetheless, we demonstrate that self-supervised pre-training on large-scale healthy population imaging dataset (N ≈10k), along with Deep Ensemble, allows DL to learn robust and transferable representations to smaller-scale clinical datasets (N ≤ 1k). It largely outperforms SML on 2 out of 3 clinical tasks both in internal and external test sets. These findings suggest that the improvement of DL over SML in anatomical neuroimaging mainly comes from its capacity of learning meaningful and useful abstract representations of the brain anatomy, and it sheds light on the potential of transfer learning for personalized medicine in psychiatry.
- Open Access
Integrating Prior Knowledge in Contrastive Learning with Kernel
Data augmentation is a crucial component in unsupervised contrastive learning (CL). It determines how positive samples are defined and, ultimately, the quality of the learned representation. In this work, we open the door to new perspectives for CL by integrating prior knowledge, given either by generative models -- viewed as prior representations -- or weak attributes in the positive and negative sampling. To this end, we use kernel theory to propose a novel loss, called decoupled uniformity, that i) allows the integration of prior knowledge and ii) removes the negative-positive coupling in the original InfoNCE loss. We draw a connection between contrastive learning and conditional mean embedding theory to derive tight bounds on the downstream classification loss. In an unsupervised setting, we empirically demonstrate that CL benefits from generative models to improve its representation both on natural and medical images. In a weakly supervised scenario, our framework outperforms other unconditional and conditional CL approaches.
Open AccessIntegrating Prior Knowledge in Contrastive Learning with Kernel
Data augmentation is a crucial component in unsupervised contrastive learning (CL). It determines how positive samples are defined and, ultimately, the quality of the learned representation. In this work, we open the door to new perspectives for CL by integrating prior knowledge, given either by generative models -- viewed as prior representations -- or weak attributes in the positive and negative sampling. To this end, we use kernel theory to propose a novel loss, called decoupled uniformity, that i) allows the integration of prior knowledge and ii) removes the negative-positive coupling in the original InfoNCE loss. We draw a connection between contrastive learning and conditional mean embedding theory to derive tight bounds on the downstream classification loss. In an unsupervised setting, we empirically demonstrate that CL benefits from generative models to improve its representation both on natural and medical images. In a weakly supervised scenario, our framework outperforms other unconditional and conditional CL approaches.
- Open Access
Benchmarking CNN on 3D Anatomical Brain MRI: Architectures, Data Augmentation and Deep Ensemble Learning
Deep Learning (DL) and specifically CNN models have become a de facto method for a wide range of vision tasks, outperforming traditional machine learning (ML) methods. Consequently, they drew a lot of attention in the neuroimaging field in particular for phenotype prediction or computer-aided diagnosis. However, most of the current studies often deal with small single-site cohorts, along with a specific pre-processing pipeline and custom CNN architectures, which make them difficult to compare to. We propose an extensive benchmark of recent state-of-the-art (SOTA) 3D CNN, evaluating also the benefits of data augmentation and deep ensemble learning, on both Voxel-Based Morphometry (VBM) pre-processing and quasi-raw images. Experiments were conducted on a large multi-site 3D brain anatomical MRI data-set comprising N=10k scans on 3 challenging tasks: age prediction, sex classification, and schizophrenia diagnosis. We found that all models provide significantly better predictions with VBM images than quasi-raw data. This finding evolved as the training set approaches 10k samples where quasi-raw data almost reach the performance of VBM. Moreover, we showed that linear models perform comparably with SOTA CNN on VBM data. We also demonstrated that DenseNet and tiny-DenseNet, a lighter version that we proposed, provide a good compromise in terms of performance in all data regime. Therefore, we suggest to employ them as the architectures by default. Critically, we also showed that current CNN are still very biased towards the acquisition site, even when trained with N=10k multi-site images. In this context, VBM pre-processing provides an efficient way to limit this site effect. Surprisingly, we did not find any clear benefit from data augmentation techniques. Finally, we proved that deep ensemble learning is well suited to re-calibrate big CNN models without sacrificing performance.
Open AccessBenchmarking CNN on 3D Anatomical Brain MRI: Architectures, Data Augmentation and Deep Ensemble Learning
Deep Learning (DL) and specifically CNN models have become a de facto method for a wide range of vision tasks, outperforming traditional machine learning (ML) methods. Consequently, they drew a lot of attention in the neuroimaging field in particular for phenotype prediction or computer-aided diagnosis. However, most of the current studies often deal with small single-site cohorts, along with a specific pre-processing pipeline and custom CNN architectures, which make them difficult to compare to. We propose an extensive benchmark of recent state-of-the-art (SOTA) 3D CNN, evaluating also the benefits of data augmentation and deep ensemble learning, on both Voxel-Based Morphometry (VBM) pre-processing and quasi-raw images. Experiments were conducted on a large multi-site 3D brain anatomical MRI data-set comprising N=10k scans on 3 challenging tasks: age prediction, sex classification, and schizophrenia diagnosis. We found that all models provide significantly better predictions with VBM images than quasi-raw data. This finding evolved as the training set approaches 10k samples where quasi-raw data almost reach the performance of VBM. Moreover, we showed that linear models perform comparably with SOTA CNN on VBM data. We also demonstrated that DenseNet and tiny-DenseNet, a lighter version that we proposed, provide a good compromise in terms of performance in all data regime. Therefore, we suggest to employ them as the architectures by default. Critically, we also showed that current CNN are still very biased towards the acquisition site, even when trained with N=10k multi-site images. In this context, VBM pre-processing provides an efficient way to limit this site effect. Surprisingly, we did not find any clear benefit from data augmentation techniques. Finally, we proved that deep ensemble learning is well suited to re-calibrate big CNN models without sacrificing performance.
- Open Access
- 17 September 2022
- Published by Elsevier in NeuroImage
- Vol. 263, 119637
OpenBHB: a Large-Scale Multi-Site Brain MRI Data-set for Age Prediction and Debiasing
- 17 September 2022
- Published by Elsevier in NeuroImage
- Vol. 263, 119637
Open AccessOpenBHB: a Large-Scale Multi-Site Brain MRI Data-set for Age Prediction and Debiasing
- 15 September 2022
- Published by Springer Nature
- p. 77-87
Unsupervised Representation Learning of Cingulate Cortical Folding Patterns
- 15 September 2022
- Published by Springer Nature
- p. 77-87
Unsupervised Representation Learning of Cingulate Cortical Folding Patterns
- 21 September 2021
- Published by Springer Nature
- p. 63-72
Detection of Abnormal Folding Patterns with Unsupervised Deep Generative Models
- 21 September 2021
- Published by Springer Nature
- p. 63-72
Detection of Abnormal Folding Patterns with Unsupervised Deep Generative Models