Abstract
Purpose :
Alzheimer’s disease (AD) is characterized by the deposition of amyloid protein in the brain and neural retina. Here, we present several machine learning methods with the aim of predicting when retinal deposits found in association with the disease are positive in thioflavin fluorescence, based on their polarization properties.
Methods :
Eyes (N=31) were obtained from donors with a primary or secondary diagnosis of AD. The retinas were stained with 0.1% Thioflavin-S and counterstained with DAPI. Flat-mounted retinas were scanned under crossed polarized light and each deposit was then imaged with fluorescence and Mueller matrix polarimetry microscopy. Imaged deposits were then categorized by the presence or absence of a Thioflavin signal. For each imaged deposit, 6 polarization properties were examined and compared between deposits with and without fluorescence signal. The data was split into two sets for training and testing. Three algorithms, linear discriminant analysis (LDA), supporting vector machine (SVM) and random forest (RF) were applied to train the classifiers to automatically determine the existence of a fluorescence signal. The accuracy, sensitivity and specificity of each classifier were computed to evaluate the performance of classification. The accuracy was obtained from 10-fold cross-validation; sensitivity and specificity were calculated from the confusion matrix of the test set.
Results :
The RF classifier had the best results for the presence or absence of a fluorescence signal, with an accuracy of 94.2%, a sensitivity of 96.7% and a specificity of 96.2%. The accuracy, sensitivity and specificity of the SVM classifier were 93.9%, 95.9% and 95.4%, respectively. Finally, the LDA classifier had an accuracy of 86.7%, a sensitivity of 84.5% and a specificity of 91.1%, which is not as powerful as the first two.
Conclusions :
The existence of Thioflavin positivity in retinal amyloid deposits can be predicted from polarization properties imaged without the use of a dye. The machine learning classifiers, random forest (RF) and supporting vector machine (SVM), gave high accuracy. Thus, polarimetry imaging can accurately identify thioflavin positive deposits in the retina. This method in turn could be used for the diagnosis of AD.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.