Phd Position F/m Generative Models for Unsupervised Anomaly Detection in Spatio-temporal Data: Application to Medical Imaging

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat**: CDD

**Niveau de diplôme exigé**: Bac + 5 ou équivalent

**Fonction**: Doctorant

**Niveau d'expérience souhaité**: De 3 à 5 ans

**A propos du centre ou de la direction fonctionnelle**:
The Centre Inria de l’Université de Grenoble groups together almost 600 people in 22 research teams and 7 research support departments.

Staff is present on three campuses in Grenoble, in close collaboration with other research and higher education institutions (Université Grenoble Alpes, CNRS, CEA, INRAE,), but also with key economic players in the area.

The Centre Inria de l’Université Grenoble Alpe is active in the fields of high-performance computing, verification and embedded systems, modeling of the environment at multiple levels, and data science and artificial intelligence. The center is a top-level scientific institute with an extensive network of international collaborations in Europe and the rest of the world.

**Contexte et atouts du poste**:
Anomaly detection is a challenging task in contexts where abnormalities are not annotated and difficult to detect even for experts. This problem can be addressed through unsupervised anomaly detection (UAD) methods, which identify features that do not match with a reference model of normal profiles. In the context of Parkinson’s disease and newly diagnosed patients, the detection task is all the more challenging as abnormalities may be subtle and hardly visible in structural MR brain scans. The goal of this project is to further improve the reliability of the detection by leveraging additional information coming from longitudinal or temporal data.

**Mission confiée**:
In particular, we would like to investigate new very successful models based on generative diffusion models. Diffusion models have been used for anomaly detection in images on one hand, and on timeseries, on the other hand. In this internship, the goal is to study their use in both contexts, see [Yang et al 2024] for a recent survey on diffusions for spatio-temporal data. Such solutions are often computationally costly. More efficient approaches have been proposed, eg. [Livernoche et al 2024] and the goal is to study their scalability to detect anomalies in images evolving over time and in particular for longitudinal medical image data which present specific challenges, such as very sparse time points, possibly missing data and non-aligned times within the patient population.

**Principales activités**:
More specifically, longitudinal data [Hedeker & Gibbons 2006] consist in the repeated observations of patients over time. In practice, we expect to analyse image data at a few different times corresponding to successive visits of patients. Their analysis informs us on the progression of the disease through the evolution of abnormalities, both in size, numbers, or locations. More specifically, when applied to anomaly detection, the expectation is the confirmation of uncertain detections or the discovery of new ones, not visible at early stages.

Modelling longitudinal data presents different types of challenges. First are the methodological challenges related to the design of relevant models to handle all the data and disease’s characteristics in order to answer the statistical and medical questions. These modelling difficulties cannot be separated from challenges arising from data with very different modalities and time dependencies, in particular involving different acquisition time-sets and different scales of patient screening, resulting on possibly partially missing data [Couronne et al 2019].

Young et al. data [Young et al 2024] recently performed an exhaustive review of data-driven generative models of how a disease evolves over time. Such models use a generative disease progression model and a set of constraints informed by human insight to infer a data-driven disease time axis and the shape of biomarker trajectories along it.

Within this framework, Sauty et al. [Sauty et al 2022] recently investigated a way to model such longitudinal effects directly in the MR images by training a linear mixed effect model in the latent representation space of a longitudinal variational autoencoder. This design enables to combine the robustness of mixed-effects modelling of clinical biomarkers progression with missing data and, for any timepoint, with that of autoencoders both to learn efficient and compact representation of 3D images and reconstruct the image from the latent variable. This model was shown to successfully model based on 3D T1w MRI normal brains and disease progression in Alzheimer patients. However, it is not clear how to reproduce such results in particular on other images.

**Initial directions of research**:
Review the state-of-the art in the domain of deep generative progression models, e.g. based on the review by Young et al and Zhang et al 2024 or other recent works.

As a first direction of research, we propose to consi