CNN, a type of deep learning neural network, has shown promising results in the field of image classification (Qin et al., 2020). CNN has the capability of extracting underlying non-linear structures of image data which are widely found in different medical images (Bogunovi et al., 2019). Hence, CNN is able to classify eye images that are considered complex and have multiple pixel dependencies throughout (Karthik and Mahadevappa, 2023). Accurate classification leads to accurate prediction of different diseases (Madbouly et al., 2020). Similar to the human brain, CNN contains several neurons that are arranged in layers (Ting et al., 2018; Zhao et al., 2020). The main three layers of CNN are the convolution layer, pooling layer, and fully connected layer (Ting et al., 2018). The convolution layer is responsible for primary filtering (Kernel filtering) of the input data (Orlando et al., 2018; Ting et al., 2018; Karthik and Mahadevappa, 2023). It takes the image as input and applies a kernel filter on it to give a convolved matrix as output (Chan et al., 2017; Bogunovi et al., 2019; Karthik and Mahadevappa, 2023). The main benefit of the convolution layer is to extract the major image features from the input image (Mohamed and Mohamed, 2021; Zobair et al., 2022). The convolution layer can also include a padding layer to control the edges of the convoluted matrix (Wang et al., 2023). The second layer is the pooling layer which applies a filter to reduce the size of the Convolved Feature, thus decreasing the computational power required to deal with it (Mohamed and Marzok, 2022; Wang et al., 2023). The third layer is the fully connected layer which is responsible for learning non-linear features of the output of the convolutional layer. The output of the convolution layer is the classification of the input images to a set of classes (Alqudah and Alqudah, 2021; Tayal et al., 2021).

The ability of CNN to extract non-linear features and semantic relations between neighbor pixels allows it to classify different diseases according to medical images (Mohamed and ElFatatry, 2015; Ting et al., 2018). In the field of ophthalmology, analyzing retinal images can predict eye diseases early (Tong et al., 2020). Such early prediction is critical for better treatment and avoids damaging tissues that leads to permanent degradation of sight (Mohamed and ElFatatry, 2019). The harmful impacts of eye diseases which are included in this work are discussed below.

In the human eye, the retina is responsible for sending images to the brain via the optic nerve (Ting et al., 2019). At the center of the retina there is the macula, which is responsible for gathering image information (Bogunovi et al., 2019; Ting et al., 2019). The retina processes this information and converts it to a neural signal to be passed to the brain. Several diseases can affect macular health, leading to sight problems (Storås et al., 2022). AMD, CNV, and DME are considered the major subjects of irreversible vision loss (Chan et al., 2017; Mohamed and ElFatatry, 2019; Ran and Cheung, 2021). Such diseases lead to geographic atrophy that affects the retina (Awais et al., 2017). The proposed work is motivated to use CNN to extract features from retinal images that predict the presence of these diseases.

According to WHO estimates, around 422 million people worldwide have diabetes (WHO, 2019). Diabetes is a set of metabolic illnesses characterized by high blood glucose levels (Le et al., 2020; Storås et al., 2022). Diabetic Macular Edema (DME) is a diabetic condition that can cause substantial damage to blood vessels in the retina, which can result in serious blindness (Akil et al., 2021). Another frequent disorder that affects the center of vision is Age-related Macular Degeneration (AMD; Prabhakar et al., 2021; Ran and Cheung, 2021). People in their 50's and 60's are frequently the first to be affected. It does not result in complete blindness. It has been associated with smoking, hypertension, being overweight, and having a family history of AMD (Zhang et al., 2020). The estimated number of persons infected with the illness was 196 million in 2020, predicted to rise to 288 million in 2040 (Pinazo-Durán et al., 2016). Choroidal Neovascular Membranes (CNVM) are new, potentially harmful blood veins that form beneath the retina. The choroid is where these blood vessels develop. They are capable of breaking through the choroid-retina barrier (Pinazo-Durán et al., 2016). They cause visual loss when they leak or bleed in the retina. CNVM may recur in 15–30% of individuals, resulting in a hemorrhagic or serous macular degeneration with accompanying vision loss (Boldt et al., 1990).

In order to avoid the danger of late diagnoses of the listed eye diseases, the proposed work designs and implements a deep-learning-based diagnostic system. The diagnostic system analyzes OCT images to detect common curable blinding retinal disorders (Le et al., 2020). OCT is a commonly used integral imaging technique (Wang et al., 2021). The main reason for using OCT is its ability to provide a cross-sectional retinal image with high resolution. Such high-resolution images of biological tissues allow CNN to extract features that indicate the presence of abnormal retina and thus allows early detection of eye diseases (Akil et al., 2021; Ran and Cheung, 2021). Figure 1 presents different OCT images of the three mentioned diseases and the normal retina. The OCT technique uses light wave's reflections to provide cross-sectional images of the macula and optic nerve as well as a clear image of blood vessels. The normal eye is characterized by a normal retinal shape without any fluid and normal blood vessels shape. AMD is characterized by retinal thickness, intraretinal fluid, and subretinal fluid. Fluid appears in OCT images as black holes on the white surface of the retina. DME images are characterized by the presence of interretinal fluid pockets, while CNVM images show abnormal sizes of blood vessels.

CNN transfer learning methodology is employed by using convolutional layers to train a neural network with a training set of data containing OCT images (Wang et al., 2021). This work contributes to designing CNN layers to be able to extract features that indicate the presence of each of the three mentioned diseases from retina images. The proposed design overcomes the complexity and non-linearity of features composing retinal images as well as the semantics of neighbor pixel relationships. A manual analysis of the OCT images with specialists is time consuming and is subject to human errors.

This main objective of this work is to diagnose early three eye disorders: DME, CNM, and AMD. Such early and accurate detection of the mentioned diseases is important for several reasons. First, it can help ensure that patients receive the most effective treatment possible. Second, it can help to rule out potentially serious conditions such as blindness. Finally, making an accurate diagnosis can provide peace of mind for both patients and their families.

The rest of the paper is structured as follows: Section 2 discusses the related work. Section 3 provides information on the Efficient Perturbation Inference and Expandable Network (EPIE-Net), while Section 4 presents experiments. In Section 5, we summarize our work and go through possible future research areas.

Several studies have been conducted to develop expert systems that automate disease diagnoses processes (Hasan et al., 2021). The objective of such systems is to obtain an accurate and fast response. In the area of medical images, CNN shows promising results. The complexity of retinal images attracted several researchers to develop machine learning systems to deal with them. In Waudby et al. (2011) and Peissig et al. (2012) researchers used written data to perform Natural Language Processing techniques to detect cataract disease without surgery. Electronic reports were used to describe the data associated with each patient including specific fields that can predict cataract disease. They concluded that eye diseases can be predicted using different Machine learning algorithms.

An early work to use OCT is presented in Ginsburg (2006). Authors used OCT imaging data for detecting intraocular lenses and refractive surgery. The authors used machine learning to identify specific eye diseases. After that, several machine learning solutions were employed for the detection of certain eye diseases such as age-related macular degeneration and automatic detection of diabetic retinopathy and automatic optical disc localization using image classification with support vector machines (Farooq and Sattar, 2015).

In Al-Mohtaseb et al. (2021) a study was conducted to find the relationship between signs and diagnoses of dry eyes disease. The researchers employed Independent Component Analysis (ICA) and Pearson correlations. Each component of the ICA indicates the negligible remaining data. Hence, no steady relationship was found among the foremost regularly utilized signs and indications.

In Ting et al. (2017) designed a deep learning system diagnostic for detecting referable AMD; their study used Fovea-centered images without macula segmentation collected from 38,189 patients. However, few patients in their sample had referable AMD. Moreover, their CNN lacks a step for image processing that includes filters to extract deep features of OCT images.

In Burlina et al. (2017) reported high diagnostic accuracy by pre-segmenting the macula region prior to training and testing step. They used the National Institutes of Health AREDS data set that contains color fundus photographs that were captured from patients. The used images included stereo pairs taken from both eyes. Both AlexNet and OverFeat models for CNN were employed to classify the input images; the authors clarified that AlexNet showed better performance.

The concept of segmentation was also applied to blood vessels to allow early detection of eye diseases. In Moccia et al. (2018) authors used a machine-learning methods model that performed a vessel segmentation process to extract features that lead to the prediction of eye diseases. Their work adopted a segmentation algorithm to overcome the drawbacks of noise and low contrast when dealing with eye images.

In An et al. (2019) introduced a machine learning technique for detecting glaucoma. The authors used three-dimensional (OCT) data as well as color fundus images to detect the abnormal features of eye retina. A segmentation algorithm was used to generate thickness and deviation maps. Then (CNN) transfer learning was applied to a set of input images such as gray-scale optic disc fundus image, retinal nerve fiber layer, and retinal ganglion cell complex. CNN is trained using data augmentation. Then a random forest (RF) was trained by combining the results of each CNN model. Their model showed high accuracy to detect glaucomatous subjects based on features extracted from images.

Prasad et al. (2019) introduced a deep neural network model that is able to recognize features associated with diabetic retinopathy and glaucoma in the early stages. They aimed to advise the patients to consult an ophthalmologist to be sure of the presence of the mentioned eye disease. However, they did not focus on obtaining high accuracy for their work.

Thus, the related work show the need for a model that can efficiently predict different widely spread types of eye disease. Such a model should be able to use famous and commonly used retinal images devices such as OCT images.

This section clarifies the design of the proposed CNN to detect the three mentioned types of diseases from CRT images. The proposed model starts with image processing for image enhancement. Then the enhanced images are fed to a CNN to apply deep learning methodology to predict and classify input images to four categories. Hence the output of the model is to classify the image as a normal retinal image or as a retinal image that suffers from DME, CNM, or Drusen. The diagnoses and classification criteria depend on the features extracted from OCT images. The proposed model is trained to classify each image that shows intraretinal fluid or subretinal fluid to be classified as an AMD image. If the fluid is in the form of several distributed pockets, the image is classified as DME. However, if the image shows larger sizes of blood vessels than the normal size the image is considered CNM. Otherwise, the image is considered a normal eye image. The training data set focuses on the previous features to classify the retinal images. The proposed model contributes to predicting the three most common eye diseases in only one model with higher accuracy and performance to other models up to now. Such a higher performance is achieved with enhancement of the image before connecting images to the implemented CNN. Image enhancement techniques were based on two basic concepts. The first is the characteristics of OCT images which depend on light reflection. The second is the features that should be abstracted to predict eye disease are blood vessel edges and fluid shape. The model also enhances performance and accuracy by performing fine tuning and optimization to the model output.

The proposed work depends on CNN to implement a deep learning model to classify eye diseases from OCT images. The used data is collected from publicly available data sets for OCT images. The system block diagram is presented in Figure 9. The proposed model was implemented using Keras, a high-level API of “Tensor Flow” platform on a PC with Intel core i7 CPU 12700H 14 cores and 32 GB RAM and NVIDIA GTX 1070 GPU.

The proposed work shows the ability of CNN to predict different eye diseases early using OCT images with an accuracy that reached 97%. The experimental results showed the higher accuracy of the proposed model over other models listed in the literature. The model classifies retinal OCT images into four classes: normal eye retina, DME, NVM, and AMD. Early eye disease detection prevents dangerous effects of eye diseases that may lead to complete loss of sight. The deep-learning model was implemented using Python libraries including Keras, OpenCV, NumPy, and matplotlib. The proposed model has the potential to be evolved by adding new diagnoses and new symptoms.

Future work will focus on adding new diseases and their symptoms to the model. We will also try to deeply classify the infection of mentioned diseases. The classification will include elementary and urgent cases. Urgent cases are cases that require urgent treatments or surgeries.

Publicly available datasets were analyzed in this study. This data can be found at: https://www.kaggle.com/datasets/paultimothymooney/kermany2018.

All authors have participated in (a) conception and design or analysis and interpretation of the data, (b) drafting the article or revising it critically for important intellectual content, and (c) approval of the final version.