A single model to segment brain tissues in any species

1998 

Abstract Submission 

romain valabregue¹, Katja Heuer², david meunier³, Ines Khemir⁴, Olivier Coulon⁵, Francois Rousseau⁶, Guillaume Auzias⁷, Roberto Toro⁸, eric bardinet¹

¹Centre de NeuroImagerie de Recherche–CENIR, Institut du Cerveau - Paris Brain Institute - ICM, Inser, paris, ²Institut Pasteur, Paris, ³Aix-Marseille Université, Institut de Neurosciences de la Timone, UMR 7289, marseille, marseille, ⁴Data Analysis Core facility, Institut du Cerveau - Paris Brain Institute - ICM, Inserm U 1127, CNRS, paris, ⁵Aix-Marseille Université, Institut de Neurosciences de la Timone, UMR 7289, Marseille, marseille, France, ⁶IMT Atlantique, LaTIM INSERM U1101, brest, ⁷Institut de Neurosciences de la Timone, UMR 7289, Aix Marseille Université, CNRS, Marseille, France, ⁸Institut Pasteur, Paris, France

romain valabregue  
Centre de NeuroImagerie de Recherche–CENIR, Institut du Cerveau - Paris Brain Institute - ICM, Inser
paris

Katja Heuer  
Institut Pasteur
Paris

david meunier  
Aix-Marseille Université, Institut de Neurosciences de la Timone, UMR 7289
marseille, marseille

Ines Khemir  
Data Analysis Core facility, Institut du Cerveau - Paris Brain Institute - ICM, Inserm U 1127, CNRS
paris

Olivier Coulon  
Aix-Marseille Université, Institut de Neurosciences de la Timone, UMR 7289, Marseille
marseille, France

Francois Rousseau  
IMT Atlantique, LaTIM INSERM U1101
brest

Guillaume Auzias  
Institut de Neurosciences de la Timone, UMR 7289, Aix Marseille Université, CNRS
Marseille, France

Roberto Toro  
Institut Pasteur
Paris, France

eric bardinet  
Centre de NeuroImagerie de Recherche–CENIR, Institut du Cerveau - Paris Brain Institute - ICM, Inser
paris

Atlas-based methods for brain segmentation rely on specific templates in order to define spatial priors. Automatic segmentation is thus available only for the few species mostly used in neuroimaging experiments, and adaptation to unseen species requires building of specific atlases, which use might in addition introduce biases in subsequent inter-species comparison studies. Having an efficient single model to segment the brain of many species would be of great interest for evolutionary biology [Milham 2022]. An example of today's limitation is the Brain Catalog database (https://braincatalogue.org) [Heuer 2019] where manual segmentation is still necessary in most cases.
Deep learning methods offer promising alternatives for brain tissue segmentation, but current models are trained for a specific species only. In this work, we evaluate the potential of training on synthetic data generated from a few species, and study whether the model can generalize to unseen species.

We used the synthetic framework [Billot 2023] because 1) it alleviates the need for real data for training and 2) the predictions are robust to variations in MRI contrast. We used the same generative process and training model (3D Unet) as described previously [Valabregue 2023]. In this study, we considered 5 different species for training: 1 adult human, 3 newborn humans from the dHCP, 3 fetal rhesus macaque, 1 adult rhesus macaque and 1 mouse lemur [Data]. The objective was to segment 13 structures in all these species (Head / CSF / Ventricles / Gray Matter / White Matter/ Cerebellum Gray Matter / 7 deep nuclei (Caudate Putamen, Pallidum, Thalamus, amygdala, accumbens, Substantia nigra ), all structures being present n the different species. The only exception was for fetal data where the deep nuclei are merged in a single structure. When such examples were seen during training, we merged the models predictions and computed the loss regarding the deep nuclei as a whole.
We validated our results on 5 different test sets : rhesus macaque (5 brains, private dataset), adult HCP (40 brains from the HCP project), dHCP Young (20 youngest subjects from the dHCP), dHCP Old (20 oldest subject from the dHCP) and new species (dog, horse, and 6 different primate species) [ref Data]. For the human test sets, we chose the Freesurfer segmentation as pseudo ground truth, for the macaque we used gray matter extracted with https://github.com/Macatools/macapype and for the new species we used a manual segmentation of the whole brain (without CSF). In this case, we added the predicted label to construct a brain mask for the evaluation.

Figure 1a) shows the Dice score for the Gray Matter (GM) for the different test sets. Although the performance of the model trained on multiple species was a bit lower compared to a model trained on a specific species, its performance was constant among all species in contrast to the species-specific models. Figure 1b) shows the Dice score for the brain mask, including the same test sets as above with the addition of the new species test set. We obtained similar results as for the GM but now the model trained on multiple species was the best-performing on new unseen species.
A full validation on each tissue for new species is a difficult task because of the lack of ground truth labels, we therefore show some examples of segmentations for a visual quality check (Cf Figure 2).

Having a unique model to segment the brain in any species is an ambitious objective but with a potential high impact for comparative anatomy and evolutionary neurobiology. We demonstrate its feasibility with synthetic training: the model shows good generalization to unseen species. Our preliminary results are promising and we hope they will motivate additional work towards universal brain segmentation methods.

Normal Brain Development: Fetus to Adolescence ²

Segmentation and Parcellation ¹

Atlasing

Machine Learning

Morphometrics

Segmentation

Other - Synthetic data; synthetic training