Assessing neurodegeneration of the retina and brain with ultra-widefield retinal imaging
Pead, Emma Jean Roberta
The eye is embryologically, physiologically and anatomically linked to the brain. Emerging evidence suggests that neurodegenerative diseases, such as Alzheimer’s disease (AD), manifest in the retina. Retinal imaging is a quick, non-invasive method to view the retina and its microvasculature. Features such as blood vessel calibre, tortuosity and complexity of the vascular structure (measured through fractal analysis) are thought to reflect microvascular health and have been found to associate with clinical signs of hypertension, diabetes, cardiovascular disease and cognitive decline. Small deposits of acellular debris called drusen in the peripheral retina have also been linked with AD where histological studies show they can contain amyloid beta, a hallmark of AD. Age-related macular degeneration (AMD) is a neurodegenerative disorder of the retina and a leading cause of irreversible vision loss in the ageing population. Increasing number and size of drusen is a characteristic of AMD disease progression. Ultra-widefield (UWF) retinal imaging with a scanning laser ophthalmoscope captures up to 80% of the retina in a single acquisition allowing a larger area of the retina to be assessed for signs of neurodegeneration than is possible with a conventional fundus camera, particularly the periphery. Quantification of changes to the microvasculature and drusen load could be used to derive early biomarkers of diseases that have vascular and neurodegenerative components such as AD and other forms of dementia.Manually grading drusen in UWF images is a difficult, subjective and a time-consuming process because the area imaged is large (around 700mm2) and drusen appear as small spots (< 125µm). An automatic approach to detecting drusen would overcome these challenges and facilitate investigations into drusen as a biomarker of neurodegeneration. In this thesis, an automatic system inspired by the recent successes of deep learning in medical image analysis was developed. As drusen are abundant in the retinas of people with AMD, a neural network was trained to classify patches in such a dataset of UWF images. This was compared to the manual gradings of two human observers. There was only a moderate agreement between observers (Kappa = 0.53, Average Dice Similarity Coefficient (DSC) = 0.38), reflecting the challenging and difficult nature of manually grading drusen in UWF images. Performances achieved for the automatic system (assessed using the area under curve (AUC) performance statistic) were 0.55-0.59, 0.62- 0.65 and 0.65-0.66 in the central, perimacular and peripheral regions of the retina, respectively. Highest performance was observed in a subset 8 images where observer agreement was at its highest (DSC > 0.8 and < 0.9), achieving AUC 0.55-0.59, 0.78-0.82 and 0.82-0.85 in the central, perimacular and peripheral zones, respectively. Measurements of the retinal vasculature appearing in UWF images of cognitively healthy (CH) individuals and patients diagnosed with mild cognitive impairment (MCI) and AD were obtained using a previously established pipeline. Following data cleaning, vascular measures were compared using multivariate generalised estimation equations (GEE), which accounts for the correlation between eyes of individuals with correction for confounders (e.g. age). The vascular measures were repeated for a subset of images and analysed using GEE to assess the repeatability of the results. When comparing AD with CH, the analysis showed a statistically significant difference between measurements of arterioles in the inferonasal quadrant, but fractal analysis produced inconsistent results due to differences in the area sampled in which the fractal dimension was calculated.When looking at drusen load, there was a higher abundance of drusen in the inferonasal region of the peripheral retina in the CH and AD compared to the MCI group. Using GEE analysis, there was evidence of a significant difference in drusen count when comparing MCI to CH (p = 0.02) and MCI to AD (p = 0.03), but no evidence of a difference when comparing AD to CH. However, given the low sensitivity of the system (partly the result of only moderate agreement between human observers), there will be a large proportion of drusen that are not detected giving an under estimation of the true amount of drusen present in an image. Overcoming this limitation will involve training the system using larger datasets and annotations from additional observers to create a more consistent reference standard. Further validation could then be performed in the future to determine if these promising pilot results persist, leading to candidate retinal biomarkers of AD.