Telecom Paris is looking for a highly motivated candidate for a 1-year post-doc on Disentangled and Controllable Latent Representations for Computer Vision and Medical Imaging using Diffusion Models.
Keywords: Deep learning, Generative Models, Diffusion Models, Contrastive Analysis, Medical Imaging, Neuroimaging, Generative models
Many of the most impressive advances in deep learning have been based on models which incorporate the notion of a latent space (autoencoders, GANs, diffusion models). This space represents data in a more compact fashion. However, much recent work on these models has been dedicated to imposing additional structure of these latent spaces. In particular, disentanglement is the goal of constructing a latent space such that each separate axis corresponds to an independent image attribute (such as head pose of a facial image, smile etc).
The goal of this postdoc is to apply/create tools for structuring and disentangling latent spaces for the understanding of medical images, in particular from neuroimaging. This will likely build on previous works of the team which have been done on latent spaces for facial and medical images. Furthermore, given the incredible results of diffusion models, we aim to use these as the model of choice. Indeed, many of these models use a latent space for synthesis, thus it should be possible to apply the aforementioned tools in this case.
The unsupervised separation of the healthy latent patterns from the pathological ones is not a trivial task in medical imaging. In neuroimaging, pathological brain signatures of psychiatric or neurodevelopmental disorders are not easily visible with the naked eye, even for experienced radiologists. The automatic identification of prognostic brain signatures of clinical courses would pave the way towards personalised medicine in psychiatry.
In this project, following our recent works in contrastive analysis (CA) [1,2], we wish to discover in an unsupervised way the salient imaging patterns that characterize a target dataset of psychiatric patients compared to a control dataset of healthy subjects, as well as what is common between the two domains. Current SOTA methods are based on VAE. However, they all ignore important constraints/assumptions and the generated images have a rather poor quality, typical of VAEs, which decrease their interpretability and usefulness.
- Study and understand the recent advances in disentanglement of latent spaces;
- Review literature on diffusion models with latent spaces;
- Adapt more recent, well-performing models, such as diffusion models, to the CA framework for neuroimaging
Deadline to candidate: 15/11/2024. Starting before: 01/01/2025.
This project will be carried out under the supervision of A. Newson (ISIR, Sorbonne Université) and P. Gori (Télécom Paris, IPParis) in collaboration with researchers and clinicians from NeuroSpin (CEA).
Salary will depend on experience and academic background (Starting salary: ~35K euros/year).
Required background PhD in applied mathematics, statistics, computer science, engineering with a good knowledge of Python and deep learning.
Candidates are invited to send a CV to anewson@isir.upmc.fr and pietro.gori@telecom-paris.fr detailing their academic background and publications.
[1] Louiset, R., Duchesnay, E., Grigis, A., Dufumier, B., and Gori, P. SepVAE: a contrastive VAE to separate pathological patterns from healthy ones. In Workshop on Interpretable ML in Healthcare at International Conference on Machine Learning (ICML) (2023).
[2] Louiset, R., Duchesnay, E., Grigis, A., and Gori, P. Separating common from salient patterns with contrastive representation learning. In International Conference on Learning Representations (ICLR) (2024).