Back to Projects

AI-enabled Imaging, Emerging Imaging

Frugal AI: Enabling the use of AI in small molecular imaging datasets through self-supervised learning

Project ID: 2023_004

1st Supervisor: Dr Jorge Cardoso, King’s College London
2nd Supervisor: Prof Sebastien Ourselin, King’s College London
Clinical Supervisor: Prof Vicky Goh, King’s College London

 

Aims of the Project: 

  • Create self-supervised AI models that can help humans label medical imaging data faster
  • Allow models to adapt to new labelling tasks with only a few examples
  • Create AI models that can learn from multiple raters and replicate their labelling style

 

Lay Summary:

Artificial Intelligence is revolutionising medical image analysis. While AI tools can be used to tackle many problems, supervised tasks, such as image segmentation, classification and object detection, are the current primary target of AI models. Developing supervised AI models often requires large labelled datasets, which are often not available or are too expensive to acquire. For example, when developing novel PET imaging tracers, disease-specific imaging datasets are acquired, but these tend to be small due to the rarity of certain diseases and the costs associated with acquiring this data. These limited-sized datasets cause many practical problems, such as overfitting and bias, causing the developed AI models to underperform in rare patient presentations. Thus, novel AI models need to be developed to be able to cope with small datasets and generalise to new unseen data.

To mitigate these limitations, this project will focus on 3 primary challenges with key applications in molecular imaging:

  • Creation of large pretrained foundational models from multimodal data
  • Develop a self-supervised training procedure to adapt the pretrained models to new data
  • Create a training mechanism that can learn new tasks using only a few examples

Expected academic background: The preferred background of the student is computer science or biomedical engineering, with prior advanced knowledge of programming, algorithms, and data processing). While prior knowledge in medical imaging would be ideal, the CDT training programme should be sufficient to make the student proficient on these topics.

 

Back to Projects