| EPSRC Reference: |
EP/Y017803/1 |
| Title: |
From 2 Million to 20 Million: Scaling and Validating a Foundation Model for Ophthalmology |
| Principal Investigator: |
Keane, Dr PA |
| Other Investigators: |
|
| Researcher Co-Investigators: |
|
| Project Partners: |
|
| Department: |
Institute of Ophthalmology |
| Organisation: |
UCL |
| Scheme: |
Standard Research - NR1 |
| Starts: |
02 October 2023 |
Ends: |
01 April 2025 |
Value (£): |
468,372
|
| EPSRC Research Topic Classifications: |
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary on Grant Application Form |
Progress in artificial intelligence (AI) continues to accelerate at a dizzying pace. There has been particular excitement recently surrounding a type of AI model known as 'foundation models,' with ChatGPT and GPT-4 being prominent examples. Our research adapts this technology and applies it to eye scans, with the aim of improving the diagnosis of eye conditions as well as general medical conditions (using the eye as a window to the rest of the body). In particular, our work focuses on helping these models work better for minority ethnic groups and rare diseases.
Traditional AI models require large quantities of labeled data in order to 'train' a model. These labels are expensive and time-consuming to produce and the resulting model is usually only able to perform one specific task. In contrast, foundation models learn patterns from data without needing labels. Since unlabelled data is everywhere, it means the models can be trained on more data, with greater diversity. In this way, a model can learn general features that are applicable to a diverse range of tasks, a feature known as generalisability. Once developed, a foundation model only requires a relatively small quantity of labeled data to adapt to these different tasks, resulting in better label efficiency.
Although the most famous examples of foundation models use language data, the same idea can also be applied to imaging data. Since medicine is so reliant on images, we believe that foundation models have the ability to transform healthcare. A particular priority for us is to use these models to improve health equity, by applying them to different ethnic groups and to rare conditions.
Our research group has recently developed one of the first foundation models in medicine, which we named RETFound. We made important progress with this model, but we believe there is lots of scope for improvement. The central aim of our research is to enhance RETFound in three main ways. First, we will vastly increase the amount of data used during training from 2 million to 20 million images. Second, we will use newer types of models with better capacity to learn from the data. Third, we will explore innovative ways to use the data, with a focus on combining different kinds of imaging scans. Based on our understanding of how these models work in other fields, we expect these changes to lead to large improvements in performance, generalisability, and label efficiency. This will allow us to achieve our aim of improving eye care and health care for patients.
Finally, once developed we will be making our foundation model freely available for other researchers to use. This will maximize the impact of our work, by allowing others to build upon our findings. We also hope this will encourage other medical specialties which use imaging data to develop similar models using our approach.
|
|
Key Findings |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Potential use in non-academic contexts |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Impacts |
|
| Description |
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk |
| Summary |
|
| Date Materialised |
|
|
|
|
Sectors submitted by the Researcher |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
| Project URL: |
|
| Further Information: |
|
| Organisation Website: |
|