Computational framework for the discovery of retinal microvascular biomarkers of diabetes and renal disease

Abstract

The retina is a thin layer at the back of the eye unique in allowing easy observation of blood vessels using simple and non-invasive instruments. The visualisation of the in vivo vasculature using retinal imaging has thereby become a key modality for the study of retinal phenotypes that can lead to the development of biomarkers for early detection not only of eye diseases but also of systemic diseases that may affect the cardiovascular system and the central nervous system.

In the last two decades, thanks to the advance in technology, new retinal imaging modalities have been introduced, optical coherence tomography (OCT) and OCT angiography (OCT-A), as new methods of visualising the neuro-retinal landscape. In particular, OCT-A, a fast and efficient technology, allows the visualisation of the vasculature at the capillary level. Considering changes in microvessel structure a common characteristic in cardiovascular disease, neurodegenerative disease, and diabetes, among others, this technology offers a unique opportunity to investigate and identify alterations associated with diseases and contribute to timely targeting patients at risk.

This thesis proposes a fully automated approach for the analysis of OCT-A images with the aim of providing novel retinal microvascular biomarkers for early detection of vascular changes and identification of disease status, realising the full potential of this imaging technology.

The first step of the OCT-A computational framework involves image processing and vessel segmentation. This is a crucial procedure in OCT-A analysis since results may be highly dependent on the accuracy of this task. To date, however, no standard blood vessel segmentation for OCT-A images has been established. This work investigates the best image processing pipeline for this type of technology by creating an original OCT-A dataset (publicly available) and by scrutinising a large set of blood vessel segmentation methods. State-of-the-art handcrafted filters and deep learning approaches (DL) for medical images are surveyed and evaluated by using standard performance metrics and newly customised OCT-A vascular measures. Results show the preeminence of DL methods over handcrafted filters and the susceptibility of clinical measurements to each of the segmentation approaches, suggesting that precaution should be taken in performing meta-analyses.

The second step of the analysis concerns the processing of the segmented OCT-A images for the discovery of candidate retinal biomarkers. Despite the popularity of deep learning approaches, these methods require a large amount of labeled data for training purposes which are not always available. Therefore, using a network-based framework, this work takes advantage of a feature engineering approach instead. Structural and functional retinal characteristics are proposed based on the current clinical knowledge. Furthermore, novel microvascular metrics based on geometric and topological properties of the graph representation of the vasculature are implemented to cover the full spectrum of possible retinal measurements, enabling the hypothesis-free discovery of new clinically relevant biomarkers of diseases.

Finally, using the extracted OCT-A microvascular phenotypes, it is possible to explore associations with ocular and systemic dysfunctions and elucidate potential biological and physiological mechanisms linked to diseases. The same OCT-A measurements can also help to identify diseased patients by using machine learning models based on modestly sized image datasets and, contrarily to many deep learning approaches, without compromising features interpretability.

This thesis presents the application of the OCT-A computational framework in three case studies: diabetic retinopathy (DR), chronic kidney disease (CKD), and living kidney donation. This framework enables to reproduce previous findings but also to show novel unreported changes. In the DR study, in accord with other investigations, results indicate a reduction in vascular skeleton density and an enlargement of the intercapillary spaces in the superficial layer of the retina. In the CKD study, changes to vessel radius and curvature of the foveal avascular zone are reported in patients with CKD when compared to healthy controls. The classification of participants, based on the retinal measurements extracted using the OCT-A computational framework, demonstrates that patients with DR are easier to identify than participants with diabetes without DR, whereas changes in patients with CKD are more subtle and challenging to detect.

The final application of the OCT-A computational framework involves participants who have donated a kidney. Despite living kidney donors being considered near-healthy patients, they are at higher risk of developing CKD and cardiovascular disease, and identifying those donors most at risk of these complications remains an open clinical challenge. The present research explores, for the first time, the possibility of using non-invasive imaging of the retinal microvasculature as a tool to improve the targeting of patients at-risk, and suggests that OCT-A microvascular phenotypes may shed light on the timing of vascular changes in kidney donors.

Studies over the past decades have provided important information about how changes in the retina vasculature can occur due to diseases years before other signs become apparent. This has led to an increasing interest in mining the retinal landscape for biomarkers of both ocular and systemic disease, and the emergence of OCT-A as a key modality for the study of the retinal microvascular system in vivo. This research proposes an automated framework for the analysis of OCT-A images and demonstrates the wide spectrum of clinical applications of this technology, showing the full potential of OCT-A imaging modality.