Description
Date depot: 8 juillet 2021
Titre: Deep learning for rating of atypical anatomical patterns on MRI data
Directeur de thèse:
Olivier COLLIOT (ICM)
Encadrante :
Claire CURY (Inria Rennes)
Encadrant :
Baptiste COUVY-DUCHESNE (ICM)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Incomplete hippocampal inversion (IHI) is an atypical anatomical pattern of the brain. Although quite common in the general population (15%-20%), it has been linked to several major neurological and psychiatric disorders, mainly epilepsy and schizophrenia. The causes of this atypical pattern remain mostly unknown. The presence of the IHI pattern can be assessed on magnetic resonance imaging (MRI) data. This is currently done using visual inspection using standardized scales. Such scales assess different features of IHI, their combination being used to determine the presence of absence of IHI. Such approach is reliable and reproducible. However, it does not scale to very large samples.
The aim of this project is to design and validate a deep learning method for automatic assessment of IHI from MRI data and apply it to study the genetic bases of IHI.
The first objective will be to develop a deep learning method for detection of IHI. We propose to set-up the problem as a joint training task, predicting simultaneously individual anatomical criteria as well as the overall presence of the whole IHI. In order to train and validate the model, we have a dataset of around 2000 subjects with annotations. Our collaborators also have annotated datasets in different diseases (schizophrenia, depression…). We will aim to extend the model to make it robust to variations of the MRI acquisition sequences.
Then, we propose to apply the model to several datasets in order to study the genetic bases of IHI. More specifically, we plan to apply the model to the UKBIOBANK comprising over 20,000 participants with MRI and genetic data, to the Queensland Twin Imaging Study as well as to some datasets of the ENIGMA consortium.
Doctorant.e: Hemforth Lisa