PhD opportunity

Robustness and Uncertainty Evaluation of Medical Vision Language Models

Funding availability

Unfunded

Application deadline

31 January 2027

Principal Supervisor

Perley Xu

View all supervisors

  • Funding – China Scholarship Council PhD Programme
  • Application deadline 31st January 2027
  • Standard Entry date - September 2027
  • Applicants are expected to have a degree (equivalent of Honours or Masters) in a relevant discipline.

Medical vision-language models combine images and text to support tasks such as image classification and visual question answering. However, a model may give a confident but incorrect answer when images are degraded, prompts are reworded, or data differ from those used in training. This PhD will develop methods to evaluate these failures and enable models to recognise when they should abstain from answering. It extends research on adversarial robustness in medical imaging towards uncertainty-aware multimodal prediction. 

The project will address three questions. How do image perturbations, changes in question wording and distribution shifts affect prediction accuracy and uncertainty? Can uncertainty estimates identify unreliable answers across these conditions? How can selective prediction, which allows a model to abstain, reduce error while retaining useful coverage? 

The research will initially focus on chest radiographs and bounded classification or visual question-answering tasks, using public datasets subject to their access and licensing conditions. The student will establish reproducible baselines using pretrained models, patient-level data splits and controlled image and text perturbations. Experiments will compare clean and perturbed performance, calibration, error detection and risk-coverage curves, with uncertainty intervals and independent test data where available. 

Building on these findings, the student will develop and test a new method that combines evidence from image-text consistency with uncertainty estimation and selective prediction. Robust training or lightweight adaptation will be investigated where justified by the failure analysis. Any statistical guarantees will state their assumptions and will be tested for sensitivity to distribution shift. The intended contributions are an evaluation protocol, an uncertainty-aware prediction method and evidence about their limitations. 

The student will gain training in Python, PyTorch, multimodal learning, adversarial evaluation, statistical calibration and reproducible research. There is scope to explore collaboration with medical imaging researchers and healthcare AI developers, subject to agreement. Initial work will use existing datasets rather than require new patient recruitment. The results will inform assessment of medical AI reliability; clinical deployment is outside the project scope.

How to apply

Please apply through the link below, you may contact the principal project supervisor to discuss your interest further, see supervisor details below.

Apply now

Supervisors

Principal supervisor

Related PhD programme