Post-Doctorant F/H Generation of connectivity models, application to population connectivity studies

Le centre de recherche Inria de Saclay a été créé en 2008. Sa dynamique s’inscrit dans le développement du plateau de Saclay, en partenariat étroit d’une part avec le pôle de l’Université Paris-Saclay et d’autre part avec le pôle de l’Institut Polytechnique de Paris . Afin de construire une politique de site ambitieuse, le centre Inria de Saclay a signé en 2021 des accords stratégiques avec ces deux partenaires territoriaux privilégiés.

Le centre compte 40 équipes-projets , dont 32 sont communes avec l’Université Paris-Saclay ou l’Institut Polytechnique de Paris. Son action mobilise plus de 600 personnes, scientifiques et personnels d’appui à la recherche et à l’innovation, issues de 54 nationalités.

Le centre Inria Saclay - Île-de-France est un acteur essentiel de la recherche en sciences du numérique sur le plateau de Saclay. Il porte les valeurs et les projets qui font l’originalité d’Inria dans le paysage de la recherche : l’excellence scientifique, le transfert technologique, les partenariats pluridisciplinaires avec des établissements aux compétences complémentaires aux nôtres, afin de maximiser l’impact scientifique, économique et sociétal d’Inria.

Dans le cadre d’un partenariat projet PEPR Brein Health Trajectories

The main goal of this project is thus to develop a generative approach for covariance models that is consistent with the Riemannian framework for connectivity modeling in fMRI [8]. We will learn to generate such samples from large populations and to fine-tune models for smaller populations. We will then assess the quality of the generated samples using a variety of metrics (authenticity, coverage, recall, GAN-train, GAN-test). Finally, we will evaluate the utility of the generative approach on longitudinal and cross-sectional connectivity-based diagnostic problems.

Des déplacements réguliers sont prévus pour ce poste ? Non

Technical developments The candidate will implement different generative models for covariance matri-
ces that are used in functional connectivity analysis.
• The so-called R-CNN network embedded in the score-based generative modeling can learn real images
at the pixel level, resulting in highly realistic covariance matrices [9].• Graph neural networks can generate good representations of node relationships. Whether the generated
covariance matrices appear realistic is still an open question.
• A well-known approach called stable diffusion [10] maps data to latent space via an encoder and then
uses a diffusion model. For specific neuroimaging tasks, many details need to be explored. This can be
combined with Riemannian score-based generative modeling [11, 12] for further experimentation.
The generative model has to be a conditional one: it should be tuned to some input information: age, sex
and characteristics of the target population, such as a disease status.
In a second step, the candidate will then systematically evaluate the quality of the generated connectivity
models using qualitative evaluation and the metrics described in [13] to assess the coverage (i.e. do the
generated covariance matrices cover the entire distribution of the observed covariance matrices ?), recall (i.e.
are all the generated samples close enough to the input distribution ?), and fidelity (i.e. are the generated
samples distinct enough from the input samples ?).
Validation on brain imaging datasets We will train the generative model on the large-scale UKBiobank
time series [14] (40k samples). The validation will consist of i) assessing whether the effects of some covariates
or treatments (age, disease, education) can be captured and reproduced by the generated data, for example
when creating explicit counterfactuals ii) assessing the utility of the data generation in downstream prediction
tasks.
When the quality is good enough, we will apply it to augment datasets from other cohorts with similar
population profiles (CamCAN [15, 16], 1000 brains[17]).
In particular, we will study the adaptation of the generators to the specific settings of new cohorts. To
handle potential covariate shifts, we will consider Riemannian approaches [18, 8] as well as optimal transport
techniques [5].

The project will produce experimental code in Python, available as an open source project, for the sake of
scientific reproducibility. Care will be taken to have good compatibility with existing frameworks such as
Nilearn.
The best parts of the code could be incorporated into Nilearn or Geomstats.
The experiments will be carried out using data available on MIND servers, that are protected.

Compétences techniques et niveau requis : Expertise in Python, expertise in deep learning, expertise in brain imaging.

Langues : English

Compétences relationnelles : Team work

Compétences additionnelles appréciées : -

• Restauration subventionnée
• Transports publics remboursés partiellement
• Congés: 7 semaines de congés annuels + 10 jours de RTT (base temps plein) + possibilité d'autorisations d'absence exceptionnelle (ex : enfants malades, déménagement)
• Possibilité de télétravail et aménagement du temps de travail
• Équipements professionnels à disposition (visioconférence, prêts de matériels informatiques, etc.)
• Prestations sociales, culturelles et sportives (Association de gestion des œuvres sociales d'Inria)
• Accès à la formation professionnelle
• Sécurité sociale

According to profile