Apprentissage profond pour l'évaluation des modèles anatomiques atypiques sur des données d'IRM

Incomplete hippocampal inversion (IHI), sometimes referred to as hippocampal malrotation, is an atypical anatomical pattern of the hippocampus found in around 20% of the general population. IHI can be assessed visually on coronal sections of T1-weighted magnetic resonance images (MRI), using a composite score that combines four anatomical criteria. IHI has been studied in association with several brain disorders (epilepsy, schizophrenia), however, these studies were based on small samples, limiting their statistical power. In addition, the factors (genetic or environmental) that contribute to the genesis of IHI are largely unknown. Large-scale studies are therefore necessary. However, visual assessment is time-consuming and tedious, justifying the need for an automatic method. The first objective of this thesis is therefore defined as the development of an automatic method to annotate IHIs on large cohorts. The second objective is then to study the genetic background of IHIs on large databases, as well as its genetic links to other phenotypic traits such as those described above. In this work, we propose, for the first time, to automatically evaluate IHI. We proceed by predicting four anatomical criteria, which are then summed to form the IHI score, which has the advantage of being interpretable. We carried out an in-depth experimental study of different machine learning methods and learning strategies. We performed automatic scoring using a variety of deep learning models (“conv5-FC3”, ResNet and “SECNN”) as well as linear models. We studied the generalization of our models using different cohorts and performed multi-cohort learning. We showed that deep learning models outperformed linear models. We demonstrated that the performance of the “conv5-FC3” network was at least as good as that of more complex networks, while maintaining low complexity and computation time. We have shown that training on a single cohort can lack variability, whereas training on several cohorts improves generalization. We applied this conv5-FC3-based automatic IHI assessment method to conduct a genome-wide association study (GWAS) on 39,432 UKBiobank participants using mixed models that maximize statistical power and limit false positives caused by population structures. We identified six novel top SNPs associated with IHI in the left hemisphere and seven in the right hemisphere, three of which were replicated in an independent sample in the left and right hemispheres. Functional annotation using gene-based assays such as mBAT-combo, h-MAGMA and multi-SMR linked the previous results to several genes that are linked to development and various traits such as Alzheimer's disease or schizophrenia (DAAM1, GPR135, CREB3L4, SLC39A1, SLIT2, LEKR1, CCDC50, IL1RAP and GMNC). The differential expression of these genes during prenatal development coincides with the hypothesis of incomplete rotation (after 12 to 25 weeks' gestation). We reported heritability using twin studies (H2 left=0.34(se=0.05); H2 right=0.37(se=0.05)) as well as SNP heritability (h2 left=0.24(se=0.01); h2 right=0.20(se=0.01)). A moderate genetic correlation (p<0.05) between IHI and brain volume and bipolar disorder (BIP) was found. However, no genetic correlation was found for any of the traits investigated in previous studies.