Representation learning for automated diagnosis and biomarker discovery in age-related macular degeneration
File(s)
Author(s)
Holland, Robert
Type
Thesis
Abstract
Representation learning promises to revolutionise the practise of medicine by automating the analysis of medical images. This thesis advances representation learning for the purpose of clinical applications to ophthalmology, focusing particularly on age-related macular degeneration (AMD). As AMD is projected to significantly impact global populations by 2040, innovative solutions are essential for early diagnosis and effective management. We develop three main representation learning techniques with applications in the practise of ophthalmology.
Firstly, we make advancements in self-supervised learning for label-efficient diagnosis. Self-supervised models have the potential to reduce the demand for labelled data by orders of magnitude, but standard approaches developed for natural images introduce several systemic issues in their application to longitudinal medical image datasets. We investigate opportunities to extend these methods to medical datasets through the use of patient metadata. However, the clinical utility of automatic detection systems is limited by the diagnostic and prognostic value of the pre-defined disease stages.
In our second work, we propose the use of self-supervised contrastive learning to accelerate the rate of clinical research into new imaging biomarkers for AMD. Data-driven approaches could enable the discovery of spatiotemporal biomarkers that were previously intractable for clinicians to identify among tens of thousands of patients and across multiple years of disease progression. By empowering ophthalmologists with self-supervised models, we investigate the extent to which cluster-based approaches can lead to the identification of novel biomarkers. We then assess if interpretable, AI-derived biomarkers can improve risk stratification for patients.
In our third and final application, we explore the potentially extensive benefits of integrating large language models (LLMs) with image-based models for facilitating ophthalmological tasks. Specifically, we introduce a specialisation curriculum for training image-based clinical decision-makers that write imaging reports for disease staging and patient referral at the level of experienced practitioners.
Firstly, we make advancements in self-supervised learning for label-efficient diagnosis. Self-supervised models have the potential to reduce the demand for labelled data by orders of magnitude, but standard approaches developed for natural images introduce several systemic issues in their application to longitudinal medical image datasets. We investigate opportunities to extend these methods to medical datasets through the use of patient metadata. However, the clinical utility of automatic detection systems is limited by the diagnostic and prognostic value of the pre-defined disease stages.
In our second work, we propose the use of self-supervised contrastive learning to accelerate the rate of clinical research into new imaging biomarkers for AMD. Data-driven approaches could enable the discovery of spatiotemporal biomarkers that were previously intractable for clinicians to identify among tens of thousands of patients and across multiple years of disease progression. By empowering ophthalmologists with self-supervised models, we investigate the extent to which cluster-based approaches can lead to the identification of novel biomarkers. We then assess if interpretable, AI-derived biomarkers can improve risk stratification for patients.
In our third and final application, we explore the potentially extensive benefits of integrating large language models (LLMs) with image-based models for facilitating ophthalmological tasks. Specifically, we introduce a specialisation curriculum for training image-based clinical decision-makers that write imaging reports for disease staging and patient referral at the level of experienced practitioners.
Version
Open Access
Date Issued
2024-07
Date Awarded
2024-10
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Rueckert, Daniel
Menten, Martin J.
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)