We are seeking two highly motivated PhD students to work at the intersection of academia and industry through the CHAI Hub (Causality in Healthcare AI), a newly founded EPSRC AI hub.

The CHAI Hub is a consortium of six UK universities, the NHS, industry partners, and government bodies, focused on revolutionising healthcare through causal AI.

These PhDs are in collaboration with Canon Medical Research Europe and focus on developing causal AI solutions for rethinking how we develop and use medical imaging, and contribute to an exciting area where AI meets imaging.

The candidates will join Prof. Sotirios Tsaftaris’ team [1] and CHAI [2], with regular visits to Canon Medical's Edinburgh-based R&D center [6] (part of the Canon Inc conglomerate) and potentially visit other CHAI partners. You’ll gain valuable experience working across both academic and industry environments, with strong mentorship and training opportunities, from leading experts.

Projects overview

The appearance of a medical image depends on acquisition factors such as scan modality (e.g. CT, MRI, X-Ray), scanner properties (e.g. detector size and characteristics), scan acquisition parameter choices (e.g. radiation dose), any tissue enhancement techniques (e.g. injected contrast), any phenomena giving rise to artefacts (e.g. metal implants causing streak artefact), and the position and pose of the patient. Thus, even for the same patient at the same timepoint, one image may have a different appearance to another; this variation makes it challenging for both human experts and automatic algorithms to interpret a scan.

Causal theory concerns the problem of modelling variables and their directional relationships, helping to answer questions such as: “If I change (intervene on) X, what will be the (size of) the effect on Y?” Causal models have been demonstrated in computer vision for scene understanding, to allow domain generalisation when there are changes in generative factors e.g. camera viewpoint, spatial object configuration [3]. Specifically, they have been studied in the context of deep learning on medical images, focussing on data collection, annotation, preprocessing, and learning strategies [4] with some preliminary investigation of robust learning in the presence of causal and domain-related factors [5]. In these projects we aim to model the causal relationships between scanner acquisition parameters, the subsequent acquired images, and the predictions of deep learning models trained or deployed on these images. We will additionally consider patient-related factors where available, such as prior images and clinical information.

Modelling causal relationships will enable simulations to test the robustness of deep learning solutions, as well as guiding the development of methods to mitigate the effect of these changes, either during training or deployment. Methods should be designed to learn from retrospective data; there may be opportunity to acquire new data under new conditions e.g. new combinations of scan acquisition parameter values.

Please note that this advert might close once the positions are filled. Please apply as soon as possible to avoid disappointment.