Deep neural network based glaucoma detection using RNFL thickness map

Abstract

Purpose : Analysis of optical coherence tomography (OCT) data can be time consuming for a busy specialist. In this study, we propose an automated deep neural network (DNN) based approach for glaucoma detection using RNFL thickness maps extracted from optic nerve head (ONH) centered OCT scans.

Methods : DNNs were trained on data from 46 normal (740 scans) and 43 glaucoma (717 scans) subjects. ONH centered OCT cube scans were acquired on PRIMUS 200 (ZEISS, India). Retinal nerve fiber layer (RNFL) thickness maps were generated using the commercially available instrument software and used as inputs to a DNN. Data were split using 5-fold constrained cross-validation, ensuring that all scans from a particular subject would be in either the training or test set. Example normal and glaucoma thickness maps are shown in Fig. 1.
For each split, we trained six pre-trained DNNs: AlexNet, InceptionV1, LiteNet, MobileNet, SuperTinyNet and VGG16, using Tensorflow. Since data from all splits are mutually exclusive, results from all splits are averaged for the final network comparison. We performed validation using a separate data set containing 214 ONH scans acquired on CIRRUS^TM HD-OCT (ZEISS, Dublin, CA). 118 scans were normal and 96 scans were early glaucoma (defined as MD > -4 dB). Mean MD was 0.29±1.04 dB for normal subjects and -1.30+1.35 dB for glaucoma patients. For this binary classification task, we report Receiver Operating Characteristic (ROC) curves, which show true positive versus false positive rates for varying thresholds.

Results : Fig. 2 shows the ROC curves of all six DNNs for the average 5-fold scores. MobileNet achieved 95% AUC followed by SuperTinyNet, InceptionV1, VGG, AlexNet and LiteNet with AUCs 92.1%, 91.9%, 91.7%, 90% and 88.6% respectively. These values are similar to AUCs reported using traditional methods on the same validation data [Mwanza et al. TVST 2018; 7(2)]).

Conclusions : In this study we were able to train six different DNNs using data from one OCT (PRIMUS) to classify glaucomatous from normal eyes where the validation set was from a second OCT instrument (CIRRUS). AUCs were similar to those using standard methods trained solely on CIRRUS data. Further cross-training with CIRRUS data could improve results beyond what is possible with standard methods.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Fig. 1. Sample RNFL thickness maps

 

Fig. 2. ROC curves