Self-Supervised Learning for the Classification of Diabetic Retinopathy

Lead Research Organisation: University College London
Department Name: Institute of Health Informatics

Abstract

Over the past decade, the increasing scale of computation, combined with datasets of unprecedented size, has driven the rise of deep learning. For the most part however, the development of deep learning systems has relied on supervised learning. In the medical domain, obtaining the detailed labels required for these supervised algorithms can be impractical. Additionally, when these labels are available, they may be unreliable due to inter expert disagreement.

Self-supervised learning has the potential to overcome this barrier. Self-supervised learning algorithms aim to learn the underlying representation of the unlabeled data through a pretext task. The representations can then be transferred to downstream tasks, such as classification or image segmentation, requiring far fewer labels.

This research will explore the application self-supervised learning to the medical domain, specifically for ophthalmic disease using electroretinography and retinal fundus photography. The research will begin by exploring the application of contrastive predictive coding to the classification of diabetic retinopathy, with a focus on achieving high performance with few labelled training examples. Later this work will be extended to explore results using a range of datasets, as well as exploring alternative implementations of self-supervised learning. This will include exploring the generalisability of learned representations to different tasks, such as different disease classifications and image segmentation. The robustness of the approaches will also be explored by looking at how well the algorithm transfers to different datasets with different demographics. Additionally, the work will aim to explore technical innovations that could enhance the interpretability of features learned by the different approaches to self-supervised learning, aiding human-AI collaboration. The work will also seek to undertake a detailed evaluation of the potential benefits of these approaches for the practical development of clinical AI systems.

For the supervisory team, Professor Daniel Alexander will provide support for algorithmic development and implementation; Dr Pearse Keane will provide a range of anonymised ophthalmic datasets with ethical approvals already in place, and expertise in the clinical evaluation of outputs. In particular, data will be provided by the Moorfields Reading Centre.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021612/1 01/04/2019 30/09/2027
2418776 Studentship EP/S021612/1 28/09/2020 30/09/2024 Madeline Kelly