You are here : Home > News > AI to Reveal the Neurodevelopmental Component of Psychiatric Disorders

Scientific result | Artificial intelligence | MRI | Brain

AI to Reveal the Neurodevelopmental Component of Psychiatric Disorders


Researchers from NeuroSpin (BAOBAB and UNIACT units) demonstrate that cortical folding, analyzed using deep learning, holds promising potential for predicting psychiatric disorders. They highlight the value of pre-training to build foundation models on large cohorts and the benefits of a regional approach.

Published on 7 April 2026

​Le p​​​lissement cortical : un marqueur du neurodéveloppement​​​

Cortical folding begins during the last trimester of pregnancy and remains relatively stable throughout life from birth (at term). This process leads to the formation of the brain's sulci and gyri, which are unique to each individual and whose structure reflects fundamental neurodevelopmental mechanisms. Growing evidence suggests that psychiatric disorders—such as schizophrenia, bipolar disorder, and autism spectrum disorders—arise from complex interactions between early disruptions in brain development and later environmental influences. Thus, patterns of cortical folding could serve as stable biomarkers, revealing the neurodevelopmental component of these conditions.

Researchers from the GAIA laboratory (BAOBAB unit /NeuroSpin) conducted a study in collaboration with the Psychiatry team at UNIACT (NeuroSpin/Henri Mondor Hospital, AP-HP) to explore the use of deep learning in extracting meaningful representations of cortical folding patterns from structural MRI images. The goal is to demonstrate that these representations enable individual prediction of major psychiatric disorders, paving the way for a better understanding of the neurodevelopmental origins of these diseases.

The researchers developed and compared three distinct approaches:

  • A global model trained from scratch, i.e., a neural network trained directly on clinical cohorts, serving as a reference framework;
  • A foundation model, using a neural network pre-trained on a large general population database (UK Biobank) to capture robust anatomical features, which are then transferred for clinical applications;
  • A regional approach: a set of specialized models (generated by the "Champollion" algorithm developed by the GAIA laboratory) that analyzes specific brain regions, offering better interpretability. Regional predictions are then combined for a global prediction.

The study's findings, published in the Journal of Neural Transmission, show that pre-training on large cohorts like the UK Biobank significantly improves predictive performance for psychiatric disorders such as bipolar disorder, schizophrenia, and autism spectrum disorder. The regional approach identifies key brain areas, such as the superior temporal sulcus, already known to be involved in autism spectrum disorders.

Contact​​ at Frédéric-Joliot Institute for Life sciences:

Edouard Duc​​​hesnay (edouard.duchesnay@cea.fr) ​

This text was translated with the assistance of Mistral AI.


Top page