A Feature Agnostic Based Glaucoma Diagnosis from OCT Images with Deep Learning Technique

Nahida Akter

Abstract

Purpose

In ophthalmology, artificial intelligence (AI) is becoming common for screening, image-interpretation, early diagnosis, and synthesizing large amount datasets. Optical coherence tomography (OCT) based glaucoma diagnosis from optic nerve head (ONH) features such as temporal-superior-nasal-inferior-temporal (TSNIT) retinal layers, average retinal nerve fiber layer (RNFL) thickness, disc topography, and macular RNFLT those cause variations in glaucomatous retinal changes became well-known clinical standard features to train them in artificial neural network (ANN). Here, we aimed to develop an automated glaucoma detection model from OCT retinal images using deep learning technique at clinician accuracy levels to accelerate the diagnosis process and synthesis extensive clinical data.

Methods

In this study, a transfer learning based modified convolutional neural network (CNN) has been used to classify eyes as normal or glaucoma directly from raw OCT images. Total 410 raw OCT (Spectralis OCT, Heidelberg Engineering) images of TSNIT retina layers (peripapillary retinal scans with circle diameters of 3.5 and 4.7) were collected from the two subject groups: 206 images for normal and 204 for glaucoma patients seen at the Centre for Eye Health. The entire dataset was split into 70% as training and 30% as validation. Two modified transfer learning architectures Alexnet and Inception-v3 were trained and evaluated the performance based on accuracy, sensitivity, and specificity. The raw images were resized to 227×227 pixels and 299×299 pixels for Alexnet and Inception-v3 respectively and no other pre-processing or segmentation has been done.

Results

The classification accuracy was measured by confusion matrix and found the accuracy for validation data 93% and 96%, sensitivity 87% and 92%, specificity (both 100%) for Alexnet and Inception-v3 respectively. Additionally, 20 random raw images were tested with the trained network and for testing data, the accuracy was 100% and in future, the data size will be increased to get the optimal result.

Conclusion

Our proposed model that was trained from raw OCT images using transfer-learning based CNN is more efficient for the diagnosis compared to the previous studies and TSNIT profile of retinal layers could be considered as a novel clinical imaging feature to train them in ANN for automated glaucoma detection.

Details

Year: 2019

Program Number: 190116

Resource Type: Scientific Presentation: Paper first choice, Poster second

Author Affiliation: School of Optometry and Vision Science, UNSW Sydney

Co-Authors: Adam Li; Rocky Shi; Jack Phu; Stuart Perry; John Fletcher; Maitreyee Roy

Co-Author Affiliation: n/a

Room: Tangerine WF3/4