Deep Learning for Retinal Disease Detection Surveys

Abstract

The accurate imaging of retinal tissues is crucial for the detectinon/treatment of retinal diseases. The direct inspection of retina was first developed by [14]. After that, many models were developed for the retinal anatomical structures. Fundus images is an effective way for early detection and screening of the main causes of blindness (e.g., glaucoma, macular degeneration, and DR). The twodimensional (2-D) representation of retinal using the "primary fundus cameras" struggled the depth, which caused inaccurate detection process of specific retinal pathology, which can be addressed using "Tomography".
On the other hand, "Optical Coherence Tomography (OCT)" is a technique that is able to capture "3-D cross sectional maps" of retina [15].
The literature includes a lot of surveys and reviews that consider retinal diseases such as detecting retinal landmark, pathology segmentation, and classification of retinal diseases [1] [13]. The contribution of this work is to comprehensively review the deep learning techniques that are mainly used for retinal image analysis, which is a lack in the literature. The main focus of this review is to cover the "2-D fundus" and " 3-D OCT" retinal images based on advanced deep learning approaches that are used for autoidentification of pathology, retinal landmarks, and classification. The metrics used in the analysis of the approaches are: sensitivity, specificity, area under ROC curve, and accuracy. Hence, the main objective of this work is present the state-of-the-art in ophthalmology based on the recent deep learning techniques.
The rest of this paper is organized as follows: Section 2 presents the principles of deep learning and the main concepts used for implementing deep learning approaches. Section 3 illustrates the approaches in deep learning that are used in retinal images analysis and the state-of-the-art in this specific field. Section 4 discusses the presented works in terms of providing general observations for researchers who work in this area. Finally, this paper is concluded in Section 5.

П. Deep Learning Techniques
Generating self-thinking programs/machines has become one of the main goal of human beings. The advent of sophisticated software tools, techniques, and proficient programming languages, the focus has increased towards designing smart models that are able to overcome our everyday challenges. Deep learning is currently considered the most efficient way to apply Artificial Intelligence (AI) in our life. Deep learning can be applied using the " Deep Neural Networks.
DNN, which is a special form of Artificial Neural Network (ANN) has the ability to perform the following: -Hierarchical feature extraction.
-Overcome input image pre-processing limitations.
Deep Neural Networks (DNN) contains three layers: input, hidden, and output. Using the DNN, features can be learned using the supervised or the unsupervised approaches and both can be used in detecting retina diseases. Figure 2 is plotted to clarify how a DNN detects retina diseases passing through network layers. Datasets on retinal diseases images are available in many sources and can be listed below: • STARE: image size of (700x605), 400 retinal images, 20 colored and 20 hand labeled images as "Ground Truth".
• ARIA: image size of (768 X 576), and 212 color fundus images where 92 are AMD, 59 are DR, and 61 are regular images.
• EyePacs DR: captured using different kinds of cameras including 35,126 images.
In fact, the datasets are not limited to the above mentioned list, there are many datasets available for researchers such as Kaggle datasets [17].

П.Ӏ Supervised-Based Learning
By supplying labeled training data, it is possible to perform supervised learning using a DNN. After then, the network attempts to learn the labels using a specific approach. Typically, supervised learning is accomplished through the use of classification algorithms that deal with labeled data. On the other hand, a response can be made by CNNs' neurons to a particular area in the input image and is termed "receptive field", which is a region in the input

Ⅱ. Ⅱ Unsupervised-Based Learning
Unsupervised learning approaches are widely involved in pattern recognition tasks. This kind of learning does not need labeled data during the learning process. The unsupervised DNNs also have the same three layers of the supervised approach and can be fully or partially tied.
Usually, input images can be compressed in unsupervised DNN. Also, noise is introduced into the input images, and then a stacked de-noising auto-encoder is used to rebuild the original image from the compressed noisy image [17].

Ⅲ. Deep Learning Challenges
Although deep learning is widely used in a variety of applications, it is still required to develop a concrete theoretical background for the tuning processes of parameters and performance evaluation of the features selected. Another challenge that struggle deep learning is the hardware resources required to perform the heavy computations of deep learning. Therefore, it is needed to adopt or develop efficient approaches that have the ability to deal with the aforementioned challenges [16].

Ⅳ. Ӏ Supervised Approaches
A supervised deep learning method for DR classification, together with an automated screening algorithm, is suggested by [17]. To this end, they designed a DNN approach. This method is run through Kaggle's DR Detection dataset to check if it's working properly. The dataset consists of 5000 patient photos taken from over 10,000 photographs. As seen in Table 1, the algorithm had an AUC of 94.6%, sensitivity of 96.2%, and 66.6% as determined.
The authors in [18] proposed a neural network to identify and classify all kinds of DR into four categories: "mild,  Table 1.
Another study performed Abramoff et al. [19] suggested an updated version of their prior work, using deep learning for categorization of both DR and ME. An improved statistical method was developed to automate the process of identifying and classifying DR into moderate, severe non proliferative DR (NPDR), PDR, and ME. The "IDX-DR X2.1" automated system was involved. The "EyeCheck" project dataset was used to train the device. To evaluate the system, the Messidor-2 dataset was used. Table 1 presents the statistical findings of the approach. Gulshan et al. [20] developed an approach to identify diabetic macular edema (DME) and diabetic retinopathy (DR) using fundus images.
They involved DNN v3 architecture based on Inception V3.
Training and testing were performed on the EyePACS-1 and MESSIDOR-2 datasets. The ImageNet dataset was utilized to set the initial weights of the network. Distributed stochastic gradient learning methods and batch normalization were both utilized in learning the weights, which resulted in an efficient training and Table 1 [22] advocate for the use of a deep CNN to detect ROP. They were the first to propose an end-to-end system based on deep learning approach for the early detection of ROP. They optimized the pre-trained GoggleNet and employed it as a ROP detector. A Bayesian framework was involved aiming to improve the accuracy.
Besides, they used pre-processing and enhancement of retinal images. A private database was used to train and test the network. As indicated in Table 1 Table 1.
Burlina et al. [24] developed their previous work by identifying a subclass of AMD. No AMD was classified as Class_1, early AMD was classified as Class_2, intermediate AMD was classified as Class_3, and advanced AMD was classified as Class_4. They determined the severity score of AMD using OverFeat deep CNN. The algorithm's efficacy was evaluated using the NIH AREDS dataset. Table 1 summarizes the performance metrics acquired. The OCT images extracted automatically and utilized to train and test the network. The Xavier algorithm [25] was used to establish the weights, and stochastic gradient descent was used to optimize them. After downscaling and histogram equalization, the input images were fed into the network.
The results are summarized in Table 1  This was performed aiming to improve the efficiency of computation. The model was capable of extracting detailed features due to the fact that its layers contain layered
Preprocessing of the input photos removes noise and adjusts contrast. The system's effectiveness was evaluated using the ARIA dataset. Table 1 summarizes the findings.

Ⅴ. Discussions
The field of retinal image analysis using DNNs is still in its infancy. Although research into the extraction of retinal diseases has been undertaken, the apex of this technology is going to grow. However, unsupervised