Longitudinal neuroimaging plays a central role in the follow-up of neurological diseases such as Alzheimer’s disease (AD) and glioblastoma (GBM). In clinical routine, physicians do not only interpret a scan as a static snapshot, but also assess how anatomy and pathology evolve over time. For example, in AD, disease progression is reflected by gradual structural atrophy, while in glioblastoma, follow-up MRI is used to monitor tumour evolution after treatment. However, longitudinal imaging datasets are often sparse, irregularly sampled, and expensive to curate, which makes temporal prediction difficult and limits the development of robust AI tools for disease progression modelling. Recent work has shown that diffusion-based and latent diffusion-based generative models can synthesize realistic medical images, but longitudinal prediction remains challenging when temporal evolution is treated only as an additional conditioning signal rather than as an explicit latent dynamical process [1,2]. World models have recently re-emerged as an important research direction in AI, motivated by the idea that systems should learn how structured states evolve over time rather than only predict the next token (current LLMs). Large language models and other generative models alone are unlikely to be sufficient for robust planning and general intelligence. This motivates exploring predictive latent-state modelling in medical imaging, where temporal structure and clinically meaningful progression are central. A promising alternative is offered by latent world models, in which the model first learns a compact representation of patient state and then predicts how this latent state evolves over time.

In particular, the Joint-Embedding Predictive Architecture (JEPA) [3-4] performs prediction directly in representation space rather than in image space, encouraging the model to focus on clinically meaningful and predictable structure while ignoring unpredictable low-level image detail. Recent developments in longitudinal medical image generation, conditional latent diffusion, and world models suggest that latent dynamics modelling is becoming an important research direction for medical imaging [5–7]. These developments make a JEPA-style latent dynamics model conditioned on patient metadata a promising basis for longitudinal neuroimaging forecasting.

The main objective is to develop and validate a JEPA-style latent dynamics model for longitudinal brain imaging forecasting based on patient metadata, with primary focus on Alzheimer’s disease and optional extension to glioblastoma depending on dataset availability.
Specific objectives are to: (1) design a structured latent prediction framework that models temporal evolution directly in representation space (2) train the model using baseline image-derived latent representations and patient metadata to evaluate whether the predicted latent states preserve clinically meaningful temporal information across irregular follow-up intervals and (4) derive and assess clinically relevant biomarkers or progression metrics from the learned latent representations, such as whole-brain or hippocampal atrophy trends in AD, or volumetric evolution indicators in GBM.

The developments will be performed in close collaboration with the university hospital UZ Brussel. The research will mainly address the exploration of novel generative AI methodologies for temporal modelling in longitudinal neuroimaging, with focus on latent dynamics rather than direct image prediction. This thesis will expose the student to a broad set of tasks and competences, covering experimental work and algorithm development and testing. The following tasks can be distinguished:
•
Literature review on longitudinal neuroimaging prediction, JEPA/world models, latent diffusion, and state-of-the-art generative methods for medical image progression modelling.
•
Familiarization with the curated dataset prepared within the related project, with likely focus on AD and optional extension to GBM depending on data readiness.
•
Implementation of a JEPA-style latent dynamics model in Python (PyTorch and MONAI), starting from latent representations obtained from a pretrained static generative model.
•
Integration of patient metadata such as age, sex, disease stage, and time interval between scans as conditioning variables for latent prediction.
•
Training and validation of the latent dynamics model on longitudinal data.
•
Evaluation of latent prediction quality using latent-space metrics and downstream clinical relevance.
•
Comparison with suitable longitudinal baselines from the literature.
•
Full thesis writing.