GCC Aware Glaucoma Detection Using Macula OCT Image Analysis Based on Deep CNN

Abstract

Glaucoma is a major public health challenge as it is the second leading cause of blindness next to cataract. Since vision loss caused by glaucoma is unrecoverable, an early, reliable diagnosis is desirable. Although complete eye examination is recommended for assessment of both structural and functional states of the disease, glaucomatous structural changes precede functional changes. Recently, classifying glaucomatous images taken from different modalities based on Deep Learning (DL) is increasingly being studied. Most of the researchers, however, focused on images generated from fundus camera and others on OCT scans taken from the optic nerve head (ONH). While others focused on specific information derived from the OCT machine itself including thickness and deviation maps of macular and ONH scans, and en-face images. However, the glaucomatous eye can be more effectively detected by analyzing the degeneration of the ganglion cell complex (GCC) by using original OCT complete scans as input. The current study used deep segmentation models to extract the GCC region which is composed of the retinal nerve fiber layer and ganglion cells with the inner plexiform layer. Convolutional Neural Network (CNN) based classifiers were used for detecting glaucomatous pathologies by paying attention to the GCC region of the macula Spectral Domain OCT (SD-OCT) scans. Model training and validation was carried out on a dataset composed of 1,262 locally acquired macula SD-OCT B-scans (432 non-glaucomatous and 830 glaucomatous) from four different regions of the macula: superior outside, inferior outside, inferior inside and central macula regions. Transfer learning was employed for segmentation as well as classifying the dataset. Three deep segmentation models, namely SegNet, PSPNet, and RAG-Net$_{v2}$ were employed for segmentation and five CNN models, namely VGG16, VGG19, ResNet50, EfficientNetV1 and InceptionV3 were used for classification purpose. SegNet showed the best performance for retina layer segmentation with 97.89% accuracy, 87.0% recall, 87.5% F1-score, 88.0% precision, 89.0% mean dice coefficient, and 81.0% mean_IOU. In terms of classification of glaucomatous and normal images, the maximum accuracy of 94.3% was obtained using VGG16 computed on the superior outside macula region, with 93.3% precision, 91.7% recall, 91.8% F1-score and 91.7% AUC. The study demonstrated that using GCC aware DL model based on macular B-scans shows great promise in the accurate screening of glaucoma and suggests that incorporating DL into macula SD-OCT for glaucoma assessment has the potential to fill some gaps in current practices and clinical workflow. To the best of our knowledge, such a DL scheme that considers the effect of the different GCC regions for the purpose of glaucoma screening hasn’t been reported in the literature before.