Convolutional neural networks (CNN) were proposed by Yann LeCun[17] in 1989. He described them as a biological-inspired adaptation of multilayer perceptrons. In 2012, the ImageNet competition showed the vast potential of CNNs in various fields. Their success was mainly due to better and fast computational resources. Recently, different models based on deep learning have been used in the analysis of medical images, especially in the detection of diseases such as cancer [18], diabetic retinopathy[19], glaucoma[20], etc. Several studies have proposed deep learning-based classification models for the detection of glaucoma in fundus images.
In 2015, Chen et al.[21] proposed a six-layer deep-learning model to detect glaucoma in fundus images. Input images were preprocessed to obtain regions of interest. Then, data augmentation was performed to extract random patches of size 224 X 224 for CNN training. The model provided an area under the curve (AUC) values of 0.832 and 0.887 for the ORIGA and SCES datasets, respectively.
Alghamdi et al.[22] used two sequential deep learning architectures to detect optic disc abnormalities in fundus images. The author used multiple classifiers and deep CNNs for extracting optic disc regions. These were further given as input to the second layer of deep CNN for identifying whether an image was healthy or not. The model achieved an accuracy of 86.52% on the HAPIEE dataset and 97.76% on the PAMDI dataset.
Deep learning-based glaucoma detection was also done by Abbas[23]. He used a convolutional neural network architecture to extract the features from the fundus images and then used a deep belief network to hand-pick the most discriminating features. The model achieved 84.50% average sensitivity, 98.01% specificity and 99% accuracy on a dataset consisting of PRV-Glaucoma datasets, DRIONS-DB, HRF and sjchoi86 HRF.
Orlando et al. [24] used two convolutional neural networks, OverFeat and VGG-S, for detecting glaucoma. The authors also used preprocessing techniques such as vessel repair, adaptive equalization of the contrast-limited histogram, and clipping around the optic nerve head for better classification. The model obtained AUC values of 0.7212 and 0.6655 when tested on the Drishti-GS1 dataset on OverFeat and VGG-S, respectively.
Andres Diaz-Pinto[24] compared the performance of five distinct ImageNet-trained models (Xception, Inception V3, VGG16, ResNet-50 and VGG19) for the automatic detection of glaucoma in fundus images. The model was evaluated on five datasets, namely Drishti-GS1, HRF, sjchoi86-HRF, RIM-ONE and ACRIMA. The authors showed that the Xception model outperformed the rest of the models with an accuracy of 80% on the HRF dataset.
Sertan Serte & Ali Serener[25] also compared the glaucoma classification performance of ResNet-50, ResNet-150 and GoogLeNet architectures using five datasets, namely HRF, Drishti-GS1, RIM-ONE, sjchoi86-HRF and ACRIMA. ResNet-152 obtained the highest accuracy of 77% on the RIM-ONE dataset compared to other networks understudy.
The recent literature discussed above shows the potential of using deep learning to automatically detect glaucoma using fundus images. Such deep learning models are more effective than traditional approaches and can also assist ophthalmologists in the early diagnosis and treatment of the disease.