Advancing human-centric AI for robust X-ray analysis through holistic self-supervised learning
8 affiliations
- 1FAIR at Meta, Paris, France. [email protected]
- 2MICS, CentraleSupélec, Université Paris-Saclay, Gif-sur-Yvette, France. [email protected]
- 3FAIR at Meta, Paris, France
- 4Department of Radiology, Hôpital Cochin, AP-HP. Centre Université Paris Cité, Paris, France
- 5Université Paris Cité, Paris, France
- 6MICS, CentraleSupélec, Université Paris-Saclay, Gif-sur-Yvette, France
- 7Cancer Data Science Unit, CentraleSupélec, Gustave Roussy, INSERM, IHU PRISM, Université Paris-Saclay, Gif-sur-Yvette, France. [email protected]
- 8MICS, CentraleSupélec, Université Paris-Saclay, Gif-sur-Yvette, France. [email protected]
Abstract
Foundational artificial intelligence models are gaining traction in various applications, including medical fields like radiology. However, medical foundation models are often tested on limited tasks, leaving their generalisability and biases unexplored. Here we introduce RayDINO, a large-scale self-supervised visual encoder for chest X-rays trained on 840,000 images and evaluated on 82,000 images sourced from 12 publicly available datasets. We compare RayDINO to previous state-of-the-art models across nine radiology tasks, from classification and dense segmentation to text generation, and provide an in-depth analysis of population, age and sex biases of our model. Our findings suggest that self-supervision enables patient-centric analysis useful in clinical workflows, allowing for a more holistic interpretation of X-rays. With RayDINO and small task-specific adapters, we reach state-of-the-art results and improve generalization to unseen populations while mitigating bias, illustrating the versatility and robustness of foundation models.
Related
- Enhancing self-supervised representation learning through lightweight learnable data augmentation
- Urea supplementation is associated with holistic changes in suprathreshold non-volatile flavour perception through oral microbiome carbon metabolism
- Validation Study of ASA Physical Status Prediction Model Developed Through Holistic Integration of Lung Images, Electrocardiogram Data and Clinical Data
- Holistic and proactive machine learning approaches to predict cyberattacks in blockchain networks
- Effectiveness of AI-Supported Holistic Health Management Through PREVERY's Digital Health Promotion and Prevention Program: A Prospective Before-and-After Observational Study
BibTeX
@article{moutakanni2026advancing,
title = {{Advancing human-centric AI for robust X-ray analysis through holistic self-supervised learning}},
author = {Théo Moutakanni and Piotr Bojanowski and Guillaume Chassagnon and Céline Hudelot and Armand Joulin and Yann LeCun and Matthew J. Muckley and Maxime Oquab and Marie‐Pierre Revel and Maria Vakalopoulou},
journal = {Nature Communications},
year = {2026},
month = aug,
doi = {10.1038/s41467-026-76076-4},
url = {https://doi.org/10.1038/s41467-026-76076-4},
note = {CC BY},
}
RIS
TY - JOUR
TI - Advancing human-centric AI for robust X-ray analysis through holistic self-supervised learning
AU - Théo Moutakanni
AU - Piotr Bojanowski
AU - Guillaume Chassagnon
AU - Céline Hudelot
AU - Armand Joulin
AU - Yann LeCun
AU - Matthew J. Muckley
AU - Maxime Oquab
AU - Marie‐Pierre Revel
AU - Maria Vakalopoulou
PY - 2026
DA - 2026/08/20
JO - Nature Communications
DO - 10.1038/s41467-026-76076-4
UR - https://doi.org/10.1038/s41467-026-76076-4
AB - Foundational artificial intelligence models are gaining traction in various applications, including medical fields like radiology. However, medical foundation models are often tested on limited tasks, leaving their generalisability and biases unexplored. Here we introduce RayDINO, a large-scale self-supervised visual encoder for chest X-rays trained on 840,000 images and evaluated on 82,000 images sourced from 12 publicly available datasets. We compare RayDINO to previous state-of-the-art models across nine radiology tasks, from classification and dense segmentation to text generation, and provide an in-depth analysis of population, age and sex biases of our model. Our findings suggest that self-supervision enables patient-centric analysis useful in clinical workflows, allowing for a more holistic interpretation of X-rays. With RayDINO and small task-specific adapters, we reach state-of-the-art results and improve generalization to unseen populations while mitigating bias, illustrating the versatility and robustness of foundation models.
SN - 2041-1723
ER -