Causal modelling of disease progression in medical images and associated clinical data

Brief description of the context of the research including potential impact

The project aims to combine disease progression modelling techniques with deep structural causal models to counterfactually reason about patient characteristics and synthesise realistic images which represent this causal modelling - for example, to predict how a patient might respond to a given intervention, or how their prognosis might change if they stop smoking. Ultimately, this could inform clinical inference, as well as be used to better educate doctors and their patients about disease trajectories.


Aims and Objectives
-The specific objectives are to:



Exploit structural causal models to constrain inference of long-term disease trajectory in Alzheimer's disease.

Models will be able to forecast disease progression from multiple clinical covariates according to existing hypotheses, and of synthesising realistic images that reasonably match predicted progression(s).

Apply structural causal modelling in datasets which are irregularly sampled, and/or represent short-term longitudinal natural histories, or are even cross-sectional to produce realistic predictions. Establish effective techniques for multi-modal datasets.

Synthesize both image and associated healthcare data which is realistic enough to support clinical inference.

Uncover the effects of various clinical factors on patient trajectories for applications such as neurological or respiratory diseases.





Novelty of Research Methodology


The Deep Structural Causal Model (DSCM) uses normalising flows and variational inference to enable tractable inference of noise variables. This represents a development over existing causal learning methods with deep learning components. The DSCM has been validated on synthetic and real-world datasets for its ability to represent all three levels of causation; namely association, intervention, and counterfactuals. Use of this and associated techniques can allow for realistic, causal image modelling of complex disease, which can ultimately allow for a better understanding of the condition of interest.





Alignment to EPSRC's strategies and research areas

The project aligns with EPSRC's Healthcare technologies theme and the following research areas:

Artificial intelligence technologies

Clinical technologies

Image and vision computing

Medical imaging

Statistics and applied probability



Any companies or collaborators involved


Microsoft Research Cambridge