Deep learning method to predict original RNFL thickness contour using anatomical parameters and OCT imaging
Recommended Citation
Najafi A. Deep learning method to predict original RNFL thickness contour using anatomical parameters and OCT imaging. Invest Ophthalmol Vis Sci 2023; 64(8):2388.
Document Type
Conference Proceeding
Publication Date
6-1-2023
Publication Title
Invest Ophthalmol Vis Sci
Abstract
Purpose : This is a proof-of-concept study. The concept is that a contour (retinal nerve fiber layer (RNFL)) can be predicted using relevant paramerts (anatomical parameters (APs) of peripapillary vessel diameter, number of vessels, and vessel location) that make the contour. This can be done using deep learning algorithms of generative adversarial network (GAN). Methods : This is a study on simulated computer-generated images. All images used in this study are generated by a python-written-code program and analyzed by a GAN. We generated 1100 dimensionality-reduced optical coherence tomography (OCT) images of APs with their corresponding RNFL thickness contour using randomly chosen AP values ( expected images ). Next, we designed a GAN to predict the RNFL contour based on the APs ( generated images ). We used 1000 images for training with 100 epochs, and 100 unseen images for testing. Finally, we used the mean absolute error (MAE) for quantification of dissimilarity between the expected and generated images. MAE is a statistical measure of errors between paired observations on the same entity. In our study, each pixel of the RNFL contour was compared between the expected and the generated images. Results : MAE was 0.021 between the expected and the generated RNFL contours. In other words, our GAN model predicted the RNFL contour correctly in 97.9% of its entirety. Conclusions : Peripapillary vessel diameter, number of vessels, and the location of vessels collectively determine RNFL thickness contour, which can subsequently be predicted accurately (97.9%) using a GAN.
Volume
64
Issue
8
First Page
2388