Back to Projects

AI-enabled Imaging, Emerging Imaging

Causal Generative Modelling for PET data decomposition

Project ID: 2023_003

1st Supervisor: Dr Jorge Cardoso, King’s College London
2nd Supervisor: Prof Sebastien Ourselin, King’s College London
Industry Supervisor: Renee Miller, GE Healthcare


Aims of the Project

  • To develop AI methods that can non-parametrically decompose PET images into phenotype, function model of PET abnormalities and associated statistical assessment.
  • Model the causal relationship of covariate of interest and introduce the ability to intervene on these covariates.
  • Extend the above models to multiple tracers, both to understand tracer-tracer interaction but also to allow for pre-trained model generalisation.


Lay Summary

Cancers can have a highly heterogeneous pattern of anomalies in positron emission tomography (PET). These specific patterns of anomaly are important to detect, stage and predict the evolution of disease. In clinical practice, PET images are mostly analysed visually, an approach which greatly depends on the observer’s experience. The quantitative analysis of PET images, and improving our understanding of what and why is a certain tracer binding to specific tissues would improve our ability to diagnose and understand disease.

One way to achieve a better understanding and quantification is to decompose the PET images into its multiple contributing factors. For example, a tracer such as FDG will have healthy binding that is not disease specific, which is a function of delivered dose. The image will also vary dependent on the time between tracer injection and image acquisition, and as a function of exercise (cardiac uptake) and renal function (kidney and bladder uptake), resulting in the final observed image in healthy subjects. Pathology then adds to this healthy signal though pathological binding of FDG. A model able to decompose a PET image into its multiple components would allow for confound independent analysis of the data and to trivially separate healthy from unhealthy binding.

This PhD project will work on conditional generative models, such as the ones used in stable diffusion (, to achieve this decomposition. However, differently from stable diffusion, decomposing imaging data will require the learning and modelling of the causal relationships between variables. As an example of causal relationship is that tracer uptake is a function of the patient weight (as we give higher dose to heavier patients), which then interacts with renal function, and thus changes the observed patterns in the bladder. The candidate will then use the decomposed images and demonstrate that these can be used to improve disease diagnosis and easy clinical interpretation.

The candidate should have a background in one of the following: biomedical engineering, applied mathematics/physics or computer science. The candidate should also have a keen interest in advanced Artificial Intelligence algorithms, and modelling of high-dimensional system.


Back to Projects