|dc.description.abstract||Alzheimer’s disease (AD) is a neurodegenerative disease and the most common cause of dementia. According to the World Health Organization (WHO) in 2019, dementia affects around 50 million people worldwide and this number is still going to increase by 10 million every year. Currently, some treatments can delay the symptoms but no effective cure is available for AD. Part of the reason is that the current definitive diagnosis of AD can only occur after patients’ death, by finding two hallmarks, amyloid-β (Aβ) plaques and tau proteins, in patients’ brains. Although the early confirmation of amyloid in the brain can be achieved by using positron emission tomography (PET) or analyzing cerebral spinal fluid (CSF), both of them are invasive to human health, PET also involves expensive procedures.
Fortunately, the over-accumulation of one hallmark Aβ has also been found in the retinas of AD patients, which Dr. Campbell’s group has shown is naturally birefringent under polarized light. Considering the optical accessibility of the retina, our group has proposed that polarimetry imaging device could be a strong candidate as an early diagnostic method for AD. The presumed retinal amyloid deposits have been detected by our Mueller matrix polarimetry method ex vivo. Since thioflavin fluorescence is a biomarker for amyloid, we stained the retinal deposits with thioflavin dye and then imaged using fluorescence microscopy to determine the existence of a thioflavin fluorescence signal. To avoid the use of a dye in future in vivo live eye imaging, this thesis presents a method to predict the existence of thioflavin fluorescence of retinal deposits from their interactions with polarized light by combining polarimetry and machine learning. Three machine learning algorithms have been trained and tested, two oversampling methods have been applied to solve the problem of the low number of non-fluorescence deposits as polarimetry detects amyloid with high accuracy. The results suggest that the fluorescent retinal deposits can be differentiated from non-fluorescence deposits with high accuracy, and two polarimetric properties appear have high importance in predicting thioflavin fluorescence.
To ensure that the source of fluorescence is amyloid-β protein, the second research project in this thesis aims to differentiate pure amyloid-β protein deposits from another protein alpha-synuclein, which contains a β-sheet structure and amyloid-like fibrils thus also show positivity in thioflavin fluorescence. A powerful convolutional neural network model (CNN) -the residual neural network, also known as Resnet, has been applied to differentiate the polarization images of pure Aβ-42 protein deposits, which is the amyloid that most relevant for AD, from the alpha-synuclein pure protein. The performance of CNNs trained by images of different polarimetric properties is compared with the machine learning algorithm used before. The CNN models, which directly take the images of the polarimetric properties as input, have outperformed the machine learning algorithms tested in differentiating Aβ-42 and alpha-synuclein protein deposits.The results reported here may be useful to assist in the label-free detection of these two types of retinal amyloid deposits in live-eye imaging.||en