Improving cross-modality alignment in medical foundation models

This PhD project tackles a key barrier to safe and generalisable medical AI, i.e. inadequate alignment between heterogeneous medical modalities, such as radiology images, multiomics/genomics data, and physiological time-series, and the language modality used by large multimodal language models. When these modalities are poorly aligned, models can generate ungrounded explanations, hallucinated findings, or clinically unreliable reports. This risk is amplified in data-scarce and high-complexity settings including rare diseases and personalised medicine, where robust grounding at the region, entity, or feature level is essential.
The student will investigate and develop new representation learning and alignment strategies that improve cross-modal grounding and attribution. Potential directions include self-supervised and contrastive pretraining, region- or entity-level alignment objectives, uncertainty-aware grounding, and data-efficient fusion techniques that better connect medical evidence (e.g., image regions, genomic variants, omics signatures, temporal patterns) to clinically meaningful language outputs. The work will also explore evaluation protocols for clinical faithfulness, including attention/attribution benchmarking and stress tests for hallucinations under distribution shift.
The project is supported by an interdisciplinary supervisory team spanning multimodal foundation models, self-supervised learning, statistical genetics, and medical AI (Dr. Ahmed, Dr. Demirkan, Dr. Awais), with established collaborations in AI and biosciences.

Applicants should demonstrate strong AI/ML fundamentals alongside working knowledge of medical data or healthcare applications. Prior research output is desirable, ideally evidenced by publications in leading AI/ML or medical AI venues (e.g., NeurIPS, ICML, ICLR, CVPR, MICCAI, or equivalent).

Open to any UK or international candidates. Up to 30% of our UKRI funded studentships can be awarded to candidates paying international rate fees. Find out more about eligibility.