|
PhD project title |
Deep learning with limited labelled data: theory, algorithm and application to medical data |
|
Outline description, including interdisciplinary, intersectoral and international dimensions |
In recent years, deep learning has shown stunning progress and been applied to many domains, e.g., medical application. However, current deep learning models heavily rely on a significant amount of labelled data and extensive parameters tuning to achieve high performance. Collecting, cleaning and labelling such a large amount of training data is costly, time-consuming, and even infeasible. Specifically, for medical application, data collection requires professional equipment, strict granting procedure and limited data source, e.g., in some cases, only a few or no examples at all may be available. Furthermore, data cleaning and labelling are also challenging since they only can be done by well-trained experts. To address these issues, in this PhD project, we mainly focus on: (1) Theoretical research on zero/few-shot learning, transfer learning (e.g., domain adaptation), meta-learning, etc; (2) Guiding by theoretical outputs, designing novel algorithms to solve general problems in deep learning with limited labelled data; (3) Applying these algorithms on medical data to make real-world impacts, e.g., earlier detection of diabetic retinopathy by identifying the earliest signs of retinal vessel change using a new non-invasive imaging modality (i.e., Optical Coherence tomography angiography) and other domain data. This project focuses on cutting-edge research, which has high potential impacts on both international academics and real-world applications, e.g., medical application. And from the track-record of the first and second supervisors, almost all publications are at international venues. Therefore, for this project, all the outputs will be published in top-tier international conferences and journals. Furthermore, we also have several international partners, who can provide more user scenarios and data for the proposed project.
|
|
Fit to CITI-GENS theme(s) |
|
|
Supervisor Information
|
First Supervisor: Dr. Yang Hua School: School of Electronics, Electrical Engineering and Computer Science Second Supervisor: Dr. Ruth Hogg School: School of Medicine, Dentistry and Biomedical Sciences & Institute for Health Sciences & Centre for Public Health Third Supervisor: Professor Tunde Peto Company: Belfast Health and Social Care Trust |
|
What costs are associated with the project and how will they be funded?
NB: The COFUND research grant supports the financing of student fees and the salary of the ‘Fellows.’ Additional overheads (e.g. specialist training, equipment) are not provided for |
- One high-performance computing workstation. It will be supported by one collaboration project (funded by Diabetes UK) between Dr Ruth Hogg (PI) and Dr Yang Hua (Co-PI).
|
|
Name of non-HEI partner(s) |
Belfast Health and Social Care Trust, UK |
|
Contribution of non-HEI partner(s) to the project:
|
The Belfast Health and Social Care Trust will support the Doctoral Programme in the following ways:
Ophthalmology department which will enhance their interaction with Clinicians and identify new research questions and needs
|
|
Profile of the non-HEI partner and the nature of the relationship.
|
The Belfast Trust is a Health and Social Care Trust covering Belfast, Northern Ireland. The Trust is one of five new NHS trusts which were created on 1 April 2007 by the then Department of Health, Social Services and Public Safety (DHSSPS). The Belfast Trust employs 22,000 staff. It has responsibility for services to over 340,000 patients, previously provided by Belfast City Hospital, The Royal Hospitals, the Mater Hospital, Green Park Healthcare Trust, North and West Belfast and South and East Belfast HSS Trusts. |
|
Research centre / School |
ECIT / EEECS |
|
Subject area |
Artificial Intelligence |