Label-free Learning from Large-Scale Multi-Modal Medical Images and Its Application to Automated Diagnosis of Eye Diseases

Artificial intelligence (AI) (particularly deep learning) has become the fundamental part of computer-aided medical diagnostics to aid clinical decision makings. As the disease has multiple risk factors, modern AI algorithms heavily rely on large well-annotated multimodal datasets. However, curating human labelled data at scale is expensive, daunting, and subject to individual bias. 

This project proposes to include multiple modalities and aims to develop a novel SSL framework, capable of integrating the cross-modal information in the learned representation from unlabelled multi-modal images, with initial focus on its application to automated diagnosis of eye diseases. 

AIMS AND OBJECTIVES  

This project aims to include multiple modalities and develop a novel SSL framework, capable of integrating the cross-modal information in the learned representation from unlabelled multi-modal images, with initial focus on its application to automated diagnosis of eye diseases. 

The objectives are to:  

1. Conduct the comprehensive literature review in relation to SSL and computer-aided disease diagnosis 
2. Design a set of pretext tasks to learn representations and model parameters from unlabeled multimodal data, which will be transferred to down-stream tasks (e.g. image classification or segmentation) 
3. Develop a novel SSL framework enabling learning powerful cross-model information and apply it to eye disease diagnosis using existing datasets from collaborators and publicly available datasets 

For more information visit https://www.mmu.ac.uk/research/research-study/scholarships#ai-71440-1