Cerebral small vessel disease (SVD) causes up to 45% of dementias and 20% of all strokes worldwide. The affected vessels in the brain are too small to be visible with routinely used hospital equipment. However, the small vessels at the back of the eye are closely related to those in the brain and can be seen in fine detail with high definition retinal imaging cameras. In this project, we will investigate how the retinal vascular characteristics are related to SVD onset and progression in the same patient’s brains. This powerful approach will help us to diagnose future SVD cases earlier and at much lower cost than current methods.
Explaining the science
OCTA diagnostic potential in the assessment of pathological conditions relies on the quality of the image analysis. A first step in many image analysis algorithms is image segmentation. However, automatic segmentation of OCTA images is still an open problem. In this project, we will establish a standard for OCTA image segmentation with particular emphasis on its applicability to routinely acquired clinical data. We will develop pixelwise Convolutional Neural Network (CNN) classifiers for the automated segmentation of microvascular angiograms and will compare them with existing approaches based on handcrafted filters.
A key challenge to be overcome is the modest size of the OCTA datasets currently available for CNN training (order of hundreds of images). We will leverage data augmentation and transfer learning to overcome this limitation. Based on our segmentations, we want to develop novel vascular structural phenotyping metrics which are physiologically informed and allow us to argue about the mechanisms responsible for the loss of vascular efficiency. We propose leveraging a formal approach using network science and machine learning, and focusing on both geometrical and topological traits as well as metrics describing temporal network evolution.
Alongside structural phenotyping, haemodynamic analysis of retinal microvascular networks would allow us to establish functional phenotyping of the retina and investigate associations with compromised brain haemodynamics. However, only qualitative methods for OCTA haemodynamic analysis exist. In previous work, we proposed the first-ever non-invasive method for the assessment of haemodynamics in the parafoveal region of the retina.
Our methodology combines adaptive optics scanning laser ophthalmoscopy (AOSLO) and Computational Fluid Dynamics modelling. In the current project, we will adapt this methodology to OCTA images. The main challenges to overcome are the definition of patient-specific boundary conditions and the estimation of uncertainty, which we will address by leveraging data assimilation techniques based on partial differential equation constrained optimisation.
We hypothesise that such structural and functional biomarkers have the potential to become a tool for characterising disease mechanisms and predicting the onset/progression of vascular conditions such as vascular dementia, stroke, diabetic retinopathy, and SVD of the brain and other tissues vascular beds.
The main aim of this project is to discover associations between the compromised brain haemodynamics observed in SVD and retinal structural and functional phenotypes derived from Optical Coherence Tomography Angiography (OCTA) retinal images.
This aim will be investigated based on delivering the following objectives:
- To establish a standard for OCTA image segmentation with particular emphasis on its robustness and applicability to routinely acquired data.
- To develop novel metrics to characterise the structure and temporal evolution of microvascular networks based on the principles of network science and machine learning.
- To establish quantitative methods for haemodynamic characterisation of the retinal microvasculature.
- To investigate associations between OCTA-derived retinal microvascular phenotypes and compromised brain haemodynamics (cerebrovascular reactivity, cerebral blood flow, and blood brain barrier leak as well as cross-sectional and longitudinal lesion and diffusion tensor imaging changes.)
At the end of the project, we will be in position to deliver the algorithms and software pipelines necessary to enable large-scale clinical studies dedicated to establishing the diagnosis potential of the proposed retinal biomarkers compared to current gold standard diagnosis approaches.
In this project, we will take advantage of data already acquired in two ongoing longitudinal studies on cerebral SVD: one in patients with obstructive sleep apnoea (an SVD risk factor) where patients are studied before and after treatment, and the other in patients with lacunar stroke (an SVD stroke) assessing SVD mechanisms and medical and lifestyle factors associated with disease worsening.
We aim to develop a non-invasive technology for SVD detection and patient management that can be integrated into medical imaging devices. The current proposal will bring this technology to technology readiness level (TRL) 1. We have identified an unmet clinical need, proposed a novel solution, and will demonstrate feasibility of OCTA retinal phenotyping for SVD prediction/association studies. At the end of the project, we will deliver a proof of concept of OCTA-based SVD detection based on prospectively acquired existing data, which will constitute transition to TRL 2.