Navigation – Plan du site

AccueilLes journées de la SAP37(S)Résumés des journées 2025Automatic skull segmentation of M...

Résumés des journées 2025

Automatic skull segmentation of MRI volume, a deep learning model trained on synthetic data

Segmentation automatique du crâne sur les volume IRM, modèle d’apprentissage profond entrainé sur des données synthétiques
Romain Valabregue, Mélanie Didier, Mathieu Santin, Éric Bardinet, Victor Giolland et Antoine Balzeau

Texte intégral

1The neuro-imaging community has focus on brain, and therefore the first preprocessing step is often to perform a brain mask, throwing out everything outside the brain. There is then a lack of tool for segmenting the skull, which is a tissue of interest for Paleoanthropology. We propose in this work to train a deep learning model for multiple tissue segmentations of the whole head. The task objective is to segment the MRI input volume in different labels. We keep the usual labels: GM (gray matter), WM (white matter), CSF (cerebrospinal fluid), dGM (deep gray matter nucleus) and cerebellum. We also include the skull and the head (regrouping all other tissues outside the skull). We obtain the training labels as follow from 5 patients who had a MRI session and a CT session. Both session were co-register, the CT scan was use to properly segment the skull, and the MRI volume was segmented (with sota tools) to obtain other labels. We then use the synthetic training strategy (SynthSeg) proposed by Billot which we reimplemented. We used pair of labels and synthetic image for supervise training of a standard UNET. Synthetic images are generated with random contrast starting only from the label. Each tissue is assigned to a random signal value, with gaussian distribution and we add standard data augmentation: noise motion affine and elastic deformation. The result is a model that can segment any MRI contrast into the defined tissue classes. The quality of segmentation are similar to previous results, but we now have an automatic segmentation of the skull. This tool was applied to obtain the 3D models of 75 volunteers imaged with MRI during the paleobrain project in order to study jointly their brain, endocast and skull.

Haut de page

Pour citer cet article

Référence électronique

Romain Valabregue, Mélanie Didier, Mathieu Santin, Éric Bardinet, Victor Giolland et Antoine Balzeau, « Automatic skull segmentation of MRI volume, a deep learning model trained on synthetic data »Bulletins et mémoires de la Société d’Anthropologie de Paris [En ligne], 37(S) | 2025, mis en ligne le 16 janvier 2025, consulté le 09 février 2025. URL : http://journals.openedition.org/bmsap/15459 ; DOI : https://doi.org/10.4000/133po

Haut de page

Auteurs

Romain Valabregue

CENIR, Institut du Cerveau (ICM) – Paris Brain Institute, Inserm U 1127, CNRS UMR 7225, Sorbonne Université, Paris, France ; romain.valabregue[at]upmc.fr

Articles du même auteur

Mélanie Didier

CENIR, Institut du Cerveau (ICM) – Paris Brain Institute, Inserm U 1127, CNRS UMR 7225, Sorbonne Université, Paris, France

Articles du même auteur

Mathieu Santin

CENIR, Institut du Cerveau (ICM) – Paris Brain Institute, Inserm U 1127, CNRS UMR 7225, Sorbonne Université, Paris, France

Articles du même auteur

Éric Bardinet

CENIR, Institut du Cerveau (ICM) – Paris Brain Institute, Inserm U 1127, CNRS UMR 7225, Sorbonne Université, Paris, France

Articles du même auteur

Victor Giolland

UMR 7194 – Histoire Naturelle de l’Homme Préhistorique (HNHP), PaleoFED team, Muséum National d’Histoire Naturelle, CNRS, UPVD, Musée de l’Homme, Paris, France

Articles du même auteur

Antoine Balzeau

UMR 7194 – Histoire Naturelle de l’Homme Préhistorique (HNHP), PaleoFED team, Muséum National d’Histoire Naturelle, CNRS, UPVD, Musée de l’Homme, Paris, France

Articles du même auteur

Haut de page

Droits d’auteur

CC-BY-NC-ND-4.0

Le texte seul est utilisable sous licence CC BY-NC-ND 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.

Haut de page
Search OpenEdition Search

You will be redirected to OpenEdition Search