Convolutional Neural Network Using RNFL Thickness Maps for Diagnosis of Glaucoma

Abstract

Purpose : RNFL analysis from OCT imaging is an important tool in the diagnosis of glaucoma. In this study, we developed a deep learning model using a Convolutional Neural Network (CNN), which extracts local features and combines them to diagnose open angle glaucoma based on patterns of RNFL thickness maps.

Methods : RNFL thickness maps of 273 glaucoma patients and 205 control patients in the Chinese American Eye Study and Los Angeles Latino Eye Study were obtained with Zeiss Cirrus 4000. Glaucoma diagnosis was based on review of clinical exam and visual field by expert panel, independent of OCT data. After excluding poor quality maps, 284 RNFL thickness maps from glaucoma patients and 360 maps from controls were used in training the CNN model, and another 37 maps with glaucoma and 42 control maps were used for testing. The CNN network was constructed with two convolutional layers and two fully connected layers. Pooling and dropout of nodes in the network were used to boost the performance. The network’s output was binary (glaucoma versus not).

Results : The CNN model achieved an overall diagnostic accuracy of 85% on repeated testing.

Conclusions : We have established a working CNN model to diagnose glaucoma with high accuracy. Further improvements could be made with blocking the optic disc from the RNFL map, reducing dimensions of the RNFL thickness map to prevent overfitting, and further increasing the size of the database.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Fig1. Comparison of RNFL thickness maps of a normal person(left) and a glaucoma patient(right).

 

Fig2. Representation of the convolutional neural network for diagnosing glaucoma with RNFL thickness maps.