Champollion : a foundation model for cortical folding
Poster No:
1728
Submission Type:
Abstract Submission
Authors:
Julien Laval1, Joël Chavas1, Antoine Dufournet1, Vanessa Troiani2, William Snyder3, Marisa Patti4, Mylène Moyal5, Marin Plaze5, Arnaud Cachia6, Federica Santacroce7, Giorgia Committeri7, Zhong Yi Sun8, Kevin De Matos9, Lisa Hemforth9, Baptiste Couvy-Duchesne10, Claire Cury11, Olivier Colliot9, Vincent Frouin12, Pietro Gori13, Denis Rivière14, Jean-François Mangin15
Institutions:
1NeuroSpin, CEA Saclay, Gif-sur-Yvette, France, 2Geisinger Autism and Developmental Medicine Institute, Lewisburg, PA, 3National Institute of Mental Health Intramural Research Program, Bethesda, MD, 4A.J. Drexel Autism Institute, Philadelphia, PA, 5GHU Paris, Paris, France, 6Université de Paris, LaPsyDÉ, CNRS, Paris, France, 7Università degli Studi, Chieti, Italy, 8Institut du Cerveau, Paris, France, 9Institut du cerveau, Paris, France, 10The University of Queensland, Brisbane, Queensland, 11Université de Rennes, INRIA, Rennes, France, 12CEA/Neurospin, Gif-sur_Yvette, France, 13LTCI, Télécom Paris, Paris, France, 14CEA, SACLAY, France, 15Université Paris-Saclay, Gif-sur-Yvette, France
First Author:
Co-Author(s):
Introduction:
Structural MRI can capture the brain's folds, made of bumps called gyri, separated by sulci. While the main sulci can be identified across all individuals, their shapes vary widely. Although computational tools already exist to characterize sulci through morphological measurements, they are limited to simple features such as length or depth. Sulcal shape variability has not been extensively studied and may contain a number of unknown biomarkers. Indeed, folding patterns were linked to prematurity (Laval, J., 2024), epilepsy (Mellerio, C., 2014), and psychiatric disorders (Isomura, S., 2017). In this study, we present Champollion, a foundation model for characterizing the human brain folding variability, trained on 42,000 subjects from the UkBioBank dataset (Bycroft, 2018).
Methods:
We develop a self-supervised learning model based on BarlowTwins (Zbontar, 2021) to encode the cortical folding variability of 42,000 subjects from UkBioBank. For better explainability, we crop the brain into 60 regions of interest (ROI) centered on the main sulci (30 per hemisphere) and train 60 regional models. This way, we obtain 60 independent 32-dimensional representation spaces that encode the variability specific to each region.
We find a unique training strategy with the exact same hyperparameters to train each regional model to avoid tremendous tuning work. We ensure the generalizability of the training strategy to all regions by concurrently optimizing the model in 3 different regions per hemisphere. To evaluate the quality of the representation, we project the ROIs of subjects from external databases into the representation spaces of their corresponding regional models and perform a simple linear classification on hand-labeled patterns, which are typical patterns of interest to neuroscientists in the field, and are the following: presence or absence of the paracingulate sulcus in the anterior cingulate region (Cachia, 2014), classification of 4 interruption types in the orbitofrontal cortex region (Troiani, 2022), and presence or absence of interruption in the intraparietal sulcus (Santacroce, 2024). This way, we select the best backbone, augmentation strategy and hyperparameters.
We then assess the representation quality in two additional regions, the central region (consisting of the central sulcus and the sylvian sulcus) on a shape-related task (Sun, Z.Y., 2012), and the collateral region (consisting of the collateral, rhinal, and occipito-temporal lateral sulci) in an interruption detection task (Kim, H., 2008).
We find a unique training strategy with the exact same hyperparameters to train each regional model to avoid tremendous tuning work. We ensure the generalizability of the training strategy to all regions by concurrently optimizing the model in 3 different regions per hemisphere. To evaluate the quality of the representation, we project the ROIs of subjects from external databases into the representation spaces of their corresponding regional models and perform a simple linear classification on hand-labeled patterns, which are typical patterns of interest to neuroscientists in the field, and are the following: presence or absence of the paracingulate sulcus in the anterior cingulate region (Cachia, 2014), classification of 4 interruption types in the orbitofrontal cortex region (Troiani, 2022), and presence or absence of interruption in the intraparietal sulcus (Santacroce, 2024). This way, we select the best backbone, augmentation strategy and hyperparameters.
We then assess the representation quality in two additional regions, the central region (consisting of the central sulcus and the sylvian sulcus) on a shape-related task (Sun, Z.Y., 2012), and the collateral region (consisting of the collateral, rhinal, and occipito-temporal lateral sulci) in an interruption detection task (Kim, H., 2008).
Results:
After optimization, we obtain a linear classification ROC-AUC of : 86% for the paracingulate sulcus detection, 80% weighted average for the binary classification of the orbitofrontal interruption types, and 85% for the intraparietal sulcus interruption. The test in the collateral region yields a ROC-AUC of 94% for the interruption between the collateral sulcus and the rhinal sulcus, further demonstrating the generalizability of our framework. Furthermore, using linear regression on continuous shape descriptors of the central sulcus, we find correlations of up to 60%. Since all of these tasks were performed on external databases with different populations and age ranges, Champollion provides robustness to site effects.
Conclusions:
We design a foundation model for characterizing cortical folding shapes in 60 predefined regions spanning the entire cortex and show its ability to retrieve hand-labeled patterns in multiple brain regions through a simple linear classification in the representation space, demonstrating its semantic richness. Champollion could provide automatic labeling to researchers in the field, alleviating the need to annotate cortical patterns through time-consuming visual inspection. Moreover, since the representation space contains a complete description of the variability, we hope to move towards a more systematic and comprehensive study of the cortical folding patterns as biomarkers.
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping 1
Novel Imaging Acquisition Methods:
Anatomical MRI 2
Keywords:
Computational Neuroscience
Cortex
Machine Learning
STRUCTURAL MRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Provide references using APA citation style.
Cachia, A. (2014). The shape of the ACC contributes to cognitive control efficiency in preschoolers. Journal of Cognitive Neuroscience 26(1), 96–106
Isomura, S. (2017). Altered sulcogyral patterns of orbitofrontal cortex in a large cohort of patients with schizophrenia. Schizophrenia 3
Kim, H. (2008). Basal temporal sulcal morphology in healthy controls and patients with temporal lobe epilepsy. Neurology 70 (Issue 22, Part 2) 2159-2165
Laval, J. (2024). Self-supervised contrastive learning unveils cortical folding pattern linked to prematurity. Medical Imaging with Deep Learning
Mellerio, C. (2014). The Power Button Sign: A Newly Described Central Sulcal Pattern on Surface Rendering MR Images of Type 2 Focal Cortical Dysplasia. Radiology 274
Santacroce, F. (2024). Human intraparietal sulcal morphology relates to individual differences in language and memory performance. Communications Biology, 7(1):520
Sun, Z.Y. (2012). The effect of handedness on the shape of the central sulcus. NeuroImage 60(1), 332–339
Troiani, V. (2022). Variability and concordance of sulcal patterns in the orbitofrontal cortex: A twin study. Psychiatry Research: Neuroimaging 324, 111492
Zbontar, J. (2021) Barlow twins: Self-supervised learning via redundancy reduction. International Conference on Machine Learning p. 12310–12320