Deep Learning and Medical Image Processing Techniques for Diabetic Retinopathy: A Survey of Applications, Challenges, and Future Trends

Diabetic retinopathy (DR) is a common eye retinal disease that is widely spread all over the world. It leads to the complete loss of vision based on the level of severity. It damages both retinal blood vessels and the eye's microscopic interior layers. To avoid such issues, early detection of DR is essential in association with routine screening methods to discover mild causes in manual initiation. But these diagnostic procedures are extremely difficult and expensive. The unique contributions of the study include the following: first, providing detailed background of the DR disease and the traditional detection techniques. Second, the various imaging techniques and deep learning applications in DR are presented. Third, the different use cases and real-life scenarios are explored relevant to DR detection wherein deep learning techniques have been implemented. The study finally highlights the potential research opportunities for researchers to explore and deliver effective performance results in diabetic retinopathy detection.


Introduction
DR happens to a person afected by diabetes, and consequently, the eyes get afected. In diabetic patients, irregular blood sugar levels are observed. Generally, glucose in the human body needs to be transformed into some form of energy while performing daily activities. When blood sugar levels exceed their normal control limit, organs such as the retina, heart, nerves, and kidneys begin to sufer. In the presence of high blood sugar levels in the human body, there is a serious risk of eye blood vessels being afected to the point where they become completely obstructed, causing blood leakage, the retinal vessels swell, and even new tiny blood vessels are formed on the eye retina. Tis is known as hyperglycemia; diabetes is categorized mainly into two types: Type 1 and Type 2 diabetes. Type 1 diabetes is an incurable disease that results from abnormalities in the insulin hormone level, which is responsible for controlling blood glucose levels. Te evidence is that Type 1 diabetes reduces the secretion of insulin hormones and hinders the ability of patients to maintain proper glucose levels in the body, leading to multiple health issues. In a Type 2 diabetes situation, the insulin hormone does not function in converting glucose to energy. On the contrary, the human body needs an adequate amount of energy to sustain a healthy and normal life and hence becomes dependent on medication to support the defciency in the insulin level. In both cases of Type 1 or Type 2, regardless of the blood sugar levels, the patient may be afected by diabetic retinopathy, and furthermore, it may lead to complete loss of eyesight [1].
Te initial symptoms of DR include blurred vision and eye foaters, and at an advanced stage, patients can lose complete vision if the disease is chronically sustained for a longer period. Tis disease is indicated by symptoms of diferent types of eye retinal lesions. Tese are comprised of hemorrhages (HM), microaneurysms (MA), hard exudates, and soft exudates. Patients who are aficted with the illness often fail to avail themselves of early diagnosis by identifying initial symptoms. Hence, it is of utmost importance for a person to ensure an eye examination is conducted at least twice a year to eliminate or proactively avoid disease complications. Considering the disease's severity, diabetic retinopathy is categorized, namely, as one that has a low severity and is recognized at an early stage. On the other hand, PDR is of a higher severity level, being detected at an advanced stage [2].
Various techniques are implemented for the diagnosis of DR at an early stage, and machine learning has predominantly contributed to the same. A subset of artifcial intelligence is a machine learning technique implemented to improve the learning experience without human intervention. Deep learning is also a technique that is predominantly used in various types of disease prediction. It is a part of artifcial intelligence and machine learning, which contribute to numerous developments in areas such as computer vision techniques, image coloring, image captioning, medical image analysis, drug discovery, neural networks, and even fraud detection systems. In deep learning, large datasets are used on which feature extraction techniques are implemented to identify the most signifcant attributes, and then various data mining algorithms are implemented to perform data analysis and predictions. Tese techniques work towards achieving optimized, accurate predictive outcomes without human intervention, and hence, disease predictions become trivial at early stages saving the lives of millions of patients. Te following Figure 1 shows the relationship between deep learning, machine learning, and artifcial intelligence.

Global Statistics on Diabetic Retinopathy. Te American
Diabetes Association Position (ADAP) has observed signifcant improvement in diabetic retinopathy detection and treatment due to the prevalent use of optical coherence tomography and intraretinal pathology. Tese results enable the evaluation of the thickness of retinal fundus images and generate results about microvascular lesions with enhanced accuracy [3]. Te statistical report also validates the fact that diabetic retinopathy is a major public health concern in today's society. Te prevalence of diabetes mellitus (DM) since 1980 men's rates have risen by 110 percent, while women's rates have risen by 58 percent, which will further increase by 9 and 7.9 percent globally in 2014. It is estimated that there will be an increase of almost 422 million patients sufering from diabetes worldwide in the next couple of years, which might further increase to almost 629 million by 2045 [4]. Similarly, a computer-aided diagnosis system can be used to reduce the burden on ophthalmologists for performing regular screening procedures. Also, various feature extraction methods, classifcation techniques, and diferent algorithms are implemented for the early diagnosis of DR [5].

Related Works
Tere exists some of the signifcant studies which are reviewed in the following sections.

Review of CNN-Based
Techniques. CNN is one of the deep learning models which is mainly concentrated on image classifcation and object identifcation techniques. Te study [6] highlighted that deep learning approaches have a signifcant impact on the automated detection of diabetic retinopathy in comparison to traditional detection methods as it yielded better accuracy, specifcity, and sensitivity. Te study examined PDR and NPDR in terms of low, medium, and high levels using the Messidor-2 data source constituting fundus photographs. Te IDX-DR X2.1 software was used for implementing CNN (convolutional neural network)-based detection techniques so that microvascular diseases could be easily identifed. Te authors [7] explained the various severity levels of DR emphasizing DME (diabetic macular edema). A DCNN (deep convolutional neural network) technique was implemented with a clinical scaling system to classify the diferent stages in disease identifcation methods when applied to the Digi fundus image dataset from Finland. Te grading aspects of retinal images were explored and a measure of the disease's seriousness was identifed with sensitivity, specifcity, and AUC (area under the curve). Te study [8] implemented a hybrid-deep learning method for easy recognition of DR with fundus pictures using the EyePACS dataset from the Kaggle repository. Also, linear-SVM techniques based on convolutional neural networks were used to classify the images, which led to enhanced performance.
Te authors in [9] presented a deep learning framework that implemented a CNN model consisting of three diferent layers to identify the diferent retinal layers that helped to predict diseases such as DR, drusen, CNV (choroidal neovascularization), and DME. Various preprocessing techniques were applied to get enhanced image quality, and optimization techniques were further implemented to remove unnecessary noise from OCT (optical coherence tomography) images, ensuring that accurate results are generated. Te authors in [10] mentioned diabetic retinopathy as being one of the most challenging issues. To detect the level of severity, deep visual features are extracted from the images that are D-color-SIFT (color dense in scale- invariant), and the GLOH (gradient location-orientation histogram technique) was implemented. Te selected features from the three image datasets, namely, DIRETDB1 and MESSIDOR-2, are fed into the semisupervised multilayer method. Te results are validated using the AUC, sensitivity, and specifcity metrics. Similarly, the study [11] presented ultrawide-feld fundus photographs that provided detailed imaging of the surface of the retina. Te study helped to automatically detect diabetic retinopathy using deep learning algorithms such as the ResNet-34 architecture, consisting of 34 layers that enabled classifcation. Also, the image segmentation method ETDRS-7SF (early treatment of diabetic retinopathy study-7 standard feld) was proposed to achieve enhanced results when validating the output using various metrics. Te authors of [12] proposed a pretrained image classifcation technique using CNN models, namely, AlexNet, VGG-16, and SqueezeNet, for classifying diabetic retinopathy images. Te customized 5-layer CNN model was a fully connected neural network classifer, which helped in the easy training of the image datasets, and the ReLU activation function was used to achieve better performance. Te existing contributions to various diabetic retinopathy datasets with deep learning approaches are shown in Table 1.

Review of DCNN-Based Techniques.
Te study in [13] highlighted that a convolutional neural network is a major contributor to image prediction. Te study proposed a multistage transfer learning approach for various labeling when applied to relevant datasets. Moreover, an automatic identifcation of diabetic retinopathy using fundus photographs generated Chen's quadratic kappa score, sensitivity, and specifcity. Te study in [14] focused on the analysis of microaneurysms at the early stages of DR and DME (diabetic macular edema) identifcation. A DCNN (deep convolutional neural network) model with a lesion detection algorithm was applied to fundus images. A semantic segmentation method categorized the pictures where ophthalmologists could easily detect the severity levels of proliferative and nonproliferative diabetic retinopathy, ensuring enhanced accuracy and efciency. Te authors in [15] aligned with the similar facts that diabetic retinopathy could be identifed using deep learning approaches and they can predict various risk factors in patients. As part of the study, the authors proposed a one-feld and three-feldinception-V3 architecture wherein validation was performed considering data from one eye selected randomly from patient data of two datasets. Finally, an internal validation dataset and an external validation dataset of 3678 and 2345 eyes, respectively, were generated. Te model's results were checked with AUC, which produced better results than traditional systems. In [16], the authors presented an entropy-based enhancement technique for refning the features of the images on the imbalanced dataset. Tis method resulted from an enhanced classifcation using a hybrid neural network and low computational cost. Te deep visual features of DR images are enhanced using graph convolutional networks (GCNs) in association with relationawarechannel-spatial attention (RACSA). In addition, a modifed deer hunting optimization algorithm is employed for extracting optimal features for enhanced classifcation accuracy [17].

Review of Deep Learning-Based Diabetic Retinopathy
Screening in Various Areas. Te initial survey is conducted highlighting that most deep learning applications use image processing techniques for DR recognition. Te number of individuals afected by the disease is expected to grow rapidly in the next 22 years, especially in developing countries. Te WHAGAP (World Health Assembly Global Action Plan) has been intended to diminish this everincreasing number since its inception in 1990 and is still ongoing. Te objective is to identify and report all visionthreatening problems and fnd ways to eliminate them. Te only way to reduce the occurrence of the disease is to initiate early and frequent screening of the disease at an early stage among diabetic patients. Te traditional techniques fail to serve the purpose, and ophthalmologists remain unable to detect the severity level. Te use of machine learning detects the disease at varying severity levels, reduces the complexities involved in traditional detection systems, and thus helps to save the lives of potential patients having a higher chance of getting afected by the disease [18,19]. Te comparison between the normal eye and the diabetic retinopathy-afected eye is specifed in Figure 2.
Te present study highlights various information pertinent to diferent deep-learning approaches utilized for the screening of diabetic retinopathy on both publicly available and real-time datasets. Te unique oferings of studies are the following: (i) Te basic information was relevant to the diabetic retinopathy disease, its types, the cause, and its presence globally which kindles the need to review the diferent techniques involved in it (ii)

Diabetic Retinopathy and Related DL/ML Techniques.
DR is an eye retinal disease that is one of the most common chronic diseases all over the world. It usually appears when the person has had diabetes for a long period, which the disease may cause. Te major issue is that the disease does not reveal diabetic retinopathy's early signs and symptoms. Te disease was classifed as low, medium, or rigorous to determine its severity. Numerous researchers have worked to prevent this type of disease complication using various machine learning and deep learning techniques. Deep learning applications incorporating medical imaging techniques have been spotlighted due to their ability to generate efective results. Tese techniques help in the classifcation and prediction of diseases without much efort. Also, the CNN architecture enables achieving optimized and enhanced quality outcomes in comparison to traditional disease prediction techniques. According to the authors of [20], DR is one of the most complicated diseases that results in vision problems all over the world. Also, this study concentrated on image quality improvement using a contrastconstrained adaptive histogram equalization model for image segmentation. Te image was optimized using a Bayesian optimization technique, and the hyperparameter tuning inception-v4 model improved the evaluation results. Some of the applications of deep learning in various disciplines are specifed in Figure 3.
Te authors [21] study explained an image classifcation approach using CLAHE (contrast limited adaptive histogram equalization) and HE (Histogram Equalization) techniques. Te study implemented the image screening method using CNN on the MESSIDOR dataset. In [22], the study employed the easy identifcation of retinal disease using deep learning methods. Te image preprocessing procedures reduced the image's dimensions using principal component analysis (PCA) and also eliminated noisy data to produce better results. Te fre-fy algorithm and standard scalar strategies were used to normalize the data in the study. Te study in [23] proposed a Siamese-like structure with a binocular network on CNN for taking fundus images of the left and right eyes. Tose features adapt the transfer learning method for classifcation and prediction. Te study computed the quadratic kappa score as one of the metrics to fnd accuracy and similarity between histogram matrix prediction and actual images. Te study in [24] proposed a convolutional neural network with novel AlexNet, VggNet, GoogleNet, and ResNet models for image classifcation to automatically recognize a set of fundus images. To enhance the performance of image datasets, hyperparameters and transfer learning techniques are considered. Data standardization and data argumentation methods were implemented to improve the quality of images.
Te advancement in CNN architecture addressed the CNN pooling layer's low accuracy issue when specifying image view characteristics. Matrix multiplication, dynamic routing, and squashing functions are shown in CapsNet architecture features on the MESSIDOR dataset for enhanced results [25]. In [26], the study proposed three different neural network models and preprocessing techniques  to describe how to classify diabetic retinopathy fundus images. Te Fuzzy C-means algorithm trains fundus images without sacrifcing image quality and the labels cannot be modifed. Te authors in [27] focused on retinal image lesions to identify the severity level of the illness. Te study explained a lesion localization model and a patch-based technique on a deep neural network for various layers of patches applied, wherein the localization of the red lesions was conducted. Te results revealed improvement in AUC level, sensitivity, and specifcity metrics. Te authors in [28] employed a convolutional neural network-based system that works on an entropy image of a green component for classifcation. Along with that, the grayscale unsharpmasking (UM) technique of image extraction used greyscale images to improve the performance. In [29], the authors proposed a bag of words model for detecting the level of severity in diabetic retinopathy images on the basic applications of support vector machine and random forest.
Various preprocessing methods were included as multiclass classifers were used to classify various severity levels of the disease. Te authors in [30] developed a novel automatedfeature approach that worked on a large collection of fundus image datasets. Te proposed approach helped in the easy identifcation of diabetic retinopathy using a computeraided model integrated with a visual heat map. Te study in [31] proposed a new dataset named DDR that was used for fundus image collection. Tose images for lesion recognition and segmentation at each level recognized the four stages of lesion detection annotations. Te study in [32] presented a deep learning interpretable classifer to organize retinal images into various levels concerning the condition. In this approach, a distribution score was obtained in the last layer of image pixel classifcation, and the pixelwise score propagation model identifed the image visual maps. Te DNN (deep neural network) concentrates on a layered approach for feature extraction methods to compute implicit results. Similarly [33], the authors surveyed diferent computer vision-related approaches for classifcation, preprocessing, and feature extraction methods recognizing the irregularities in retinal blood vessels. Te study in [34] presented the training and test datasets, and DCNN worked to classify fundus photographs. Tese outcomes indicated DR and the presence of DME (diabetic macular edema). Te authors in [35] proposed to fnd the severity level of the disease using fundus image classifcation techniques and an image normalization approach to detect the size and shape of the images. On the basis of the processed images, an inception-v4 algorithm was used to enhance the quality of an image. Te study in [36] proposed machine learning algorithms such as SVM, KNN (K-nearest neighbor), and bagged trees applied to diabetic patients' physical health records to easily identify diabetic retinopathy. Te model generated a superior level of accuracy using the bagged trees prediction method. Te authors in [37] proposed a collective intelligence approach (human + eye) for enhanced results in comparison to the traditional artifcial intelligence techniques to fnd the various categories of diabetic retinopathy. Te proposed method used EfcientNetB3, EfcientNetB4, and Ef-cientNetB5 to improve the accuracy of images that had already been trained. Tis approach restricts a smaller training dataset due to the variance and amount of data distribution. Te study in [38] presented a preprocessing approach and a contrast-limited histogram method to refne the images on a dataset and resulting in high-quality images. A multiclass classifer transfer learning method trained with GoogleNet architecture was used to extract the image features to generate efcient results. Te study in [39] proposed automatic diagnosis approaches implemented on retinal images to identify the severity of the disease. In this proposed machine learning algorithm, the inception network architecture is trained and tested on the EYEPACS dataset for classifying the fundus images. Tese approaches yielded enhanced accuracy results in comparison to the traditional approaches. Tese deep learning and machine learning models also help to detect and prevent various health-related issues, for example, mental disabilities in humans [40]. Te CNN model in [41] is used to identify the patient's condition based on image emotion exposure. Moreover, deep learning methods were used to process biomedical signal information. Te prognosis of dermatological diseases and related CNN implementations are discussed in [42]. A summary of the various studies reviewed and discussed in this section is presented in Table 2.

Performance Evaluation of the Reviewed DL/ML
Implementations. Te various DL/ML techniques used in DR diagnosis are critically reviewed. Each of the implementations were evaluated using commonly used metrics such as accuracy, specifcity, sensitivity, precision, recall, and ROC. Tis section presents a comparative analysis of the performance of these techniques as published in the reviewed study. Te section also identifes the limitations of each of the implementations which enable researchers to understand the associated challenges thereby providing ideas for model optimization. Te consolidated parametric evaluation for performance analysis is presented in Table 3.

Significance of Deep Learning Applications Using Medical Imaging Techniques
Applications of deep learning have had a signifcant impact [43] on medical image analysis over the past few years. Tese applications have helped ophthalmologists easily detect various eye diseases and identify the severity levels of patient conditions relevant to the disease. Te signifcant applications of deep learning for medical image processing are specifed in Figure 4. Te study in [44] stated that multiple people were identifed as sufering from any eye disease at its initial stages by the ophthalmologists using various types of classifcation rules and prediction techniques to generate the best results. Te study in [45], explained that the layered approach of CNN helped to improve the accuracy of predictions concerned with fundus photographs. Te authors in [46], addressed a convolutional neural network model, which was implemented as an efective way to classify and predict the images in diferent layers leading to optimized performance. It provided real-time classifcation techniques to employ deep network features to achieve  8 Journal of Healthcare Engineering   [47] discussed medical imaging processes that were used as an efective way to grade the severity levels using a neural network classifer for fundus images. A multistage classifcation deep neural network was used in association with a shallow dense layer, being applied to the diabetic retinopathy dataset that generated grading of the disease with enhanced accuracy. Te study also implemented a graded equalization method to improve the effectiveness and transparency of the pictures for microaneurysm detection.

Classifcation Techniques.
In image processing, the fundus images are classifed using a CAD (computer-aided diagnosis) system by implementing various imaging techniques to achieve early diagnosis and clarity regarding patient health conditions. Te previous studies explain several preprocessing procedures for DR-related images and classify the presence of DR in a patient. Studies have implemented various classifers incorporating CNN layers to predict the disease. Also, multicast lesion detection procedures and modifed-CNN applications are implemented on fundus images to classify normal PDR and severe NPDR, thereby generating a PDR histogram, threshold edge values, accuracy, and specifcity [48,49]. Te authors in [50] used an automatic screening tool for reducing manual interruption and ensuring the diagnosis of the disease in a shorter time frame. Tese screening methods are implemented in different datasets to classify and detect the DR grade levels thereby implementing diferent levels of image-enhancing rules such as normalization and extension of the quality of lesions. Tese lesions can be classifed using image-level descriptors to monitor the patient's condition and grade the level of severity [51]. Te study in [52] surveyed various traditional works that were performed using several screening methods for early detection of DR. Tese screening methods are used for image cropping, rotating, and resizing aspects to fnd mild variations in retinal vessels that generate optimized outcomes. Additional approaches for evaluating DR levels at the initial stage include using the visual graphic (VG) method for the classifcation of images and then the performance is calculated using error-correcting output codes (ECOC) classifer [53]. Te study in [54,55], explained the use of pretrained CNN models with a transfer learning approach for the refnement of images. Tese models classify images using a support vector machine (SVM) considering publicly available databases. DR detections and lesions identifcation problems can also be resolved using deep MIL (multiple instance learning) wherein handling of the improper DR lesions is performed using a well-trained framework to achieve enhanced quality Te earliest sign of retina disease include the identifcation of exudates, namely, soft and hard exudates. Tese exudates were identifed in terms of yellow color cotton wool spots using morphological methods involving multiple numbers used for the fltering of the retina lesions. Te authors in [56,57] proposed a training method that efciently identifed the existence of exudates using CNN-based framework. Various studies show how normalizing pixel values for training and testing images can be performed with the help of image-enhancing techniques. In [58], the authors proposed a multilayer deep CNN and SVM classifer method that performed DR segmentation and detection. Tese retinal images were preprocessed using the gaussian fltering method. Tis method also explained about ways to fnd the age-related retinal abnormalities along with DR classifcation. Te authors in [59] presented an optical coherence tomography (OCT) approach using 3D-Markov Gibbs random feld (MGRF) for extracting higher-order image features. Tese image features are applied to each layer of the artifcial neural network (ANN) classifer to detect DR disease more accurately.

Detection Techniques.
Detection means identifying an object or specifc part of important data in images. Tis mechanism mainly worked on the medical imaging process, identifying the defects, and security of the surveillance system. Tese approaches are bound with computer vision techniques that computed the human vision system complications. In this scenario, artifcial intelligence techniques computed enhanced results in image recognition. When it comes to multiimage recognition methods, machine learning and deep learning mechanisms are efectively performed. Te study in [60] proposed the detection of medical images using deep learning applications. Tis study implemented blood vasculature method to detect the retina blood vessels' position in the afected area on images. Tese vascular results are used by ophthalmologists to give treatment to patients at early stages. Te study in [61], explained the SVM method to detect NPDR. Te SVM classifer estimated the disease's severity using around 400 test images from the CHASE dataset. Apart from that, a confusion matrix was used to evaluate the results. Another study in [62] developed a hybrid machine learning method as a detection tool to identify if a patient has retinopathy or not using retinal images. Tis technique used multiple instances of learning strategies and feature extraction techniques for the results generated. Te study in [63], surveyed the results of numerous studies on diabetic retinopathy disease detection and classifcation issues. It also reviewed microaneurysms, exudates, hemorrhages, and optic disc disorders and their prognosis details. Tese were used to create awareness on potential future research studies. Te authors in [64] developed a DTL (deep transfer learning) approach to detect the various layers of CNN architecture using the inception-v3 algorithm. Tese methods divided the network into depths and widths to reduce noisy data for enhanced computation results such as sensitivity and specifcity.
Te study in [65], proposed an image processing method instead of using lesion segmentation and implemented an image classifcation method. Tese approaches helped to identify the retinal lesion patches and classifed viable retinal vascular along with encoding the texture of the image with local binary patterns. Te study in [66] demonstrated an image recognition system for the neurovascular retinal detection method that used a matching flter and a fuzzy entropy-based technique. Tis technique employs a gaussian flter to precisely recognize small retinal blood vessels in datasets such as MESSIDOR and DRIVE with precise measurements. Te study in [67] proposed toboggan segmentation and the multiagent process that worked on the two-sided segmentation of damaged retinal vessels. Tese were assessed to enhance the quality of the image using flters such as the gaussian and modifed kirsch. Te authors in [68] explained retinal images based on color characteristics extraction with an ensemble machine learning approach. In this approach, the image analysis, image cropping, resizing, and removal of irrelevant image data were performed using the data source available from the Kaggle repository. Finally, the improved classifcation metrics were achieved under the ROC curve. Te study in [69] presented a novel transfer learning approach for detecting the exudates in retinal images using ResNet-50, inception-v3, and VGG-19 techniques. Tese were trained using a CNN model that had already been trained, and then they were applied to one-Ophtma and DIRETDB1 data sources to get accurate classifcation results.

Segmentation Techniques.
Segmentation is a technique for partitioning images into subgroups or segments to reduce complexities in subsequent processing and analysis. Te purpose of segmentation is always to change the orientation of an image so as to evaluate it more easily. Also, it could be used to partition the image into sets of pixels, and each pixel gives a clear contrast. Tis technique yields a high degree of accuracy and enhanced results. In [70], the study proposed a segmentation method that worked on several tissues in the human brain, facial parts, ears, and eyes easily segmented on the visual Chinese human head (VCH) model. Deep learning-based techniques were used on fve separate MRI imaging datasets to identify the disease's consequences. It results achieved enhanced accuracy in comparison to the manual segmentation processes. Tis technique had limitations and is not suited to very big MRI datasets. In [71], Te study described how to predict leaf disease and segment it for easy identifcation, as well as how to use color transforms to fnd the afected leaf. Tis study used a kmeans clustering algorithm to categorize disease symptoms and separate them as clusters at various stages. Te study in [72], proposed a new wavelet-based image segmentation method that involves transforming input images into different orientations while maintaining the standard pixels needed to detect image alterations. To state a clear representation of the specifc item, we used morphological approaches to retain specifc image segment frontiers.
Te authors in [73] proposed a deep-learning architecture with 2D and 3D images for cardiac function segmentation. In this work, 2D modifed U-Net architecture was implemented in CNN models segmented on short-axis MR images, resulting in enhanced performance. Te authors in [74] stated that segmentation applications play an important role in medical images and Francis' investigation of iris, biometric, and vein segmentation for real-world problems. Deep learning models used encoder-decoder architecture and pretrained classifers such as VGG-Net, DenseNet, and ResNet to get information that could be used on a large scale. Te study in [75] explained MRI brain tumor images and observed a typical tissue to remove noise. Te proposed method concentrated on decreasing the noise from a tumor cell and used image segmentation models such as FSSN (Fernandez-Steel Skew Normal), GMM (Gaussian Mixture Model), and Fuzzy C-means.

Registration Techniques.
Image registration is an important aspect of the imaging system because it aids in the completion of certain tasks in medical image processing. In feature and intensity-based classifcations, image attributes such as image alignment and diferent measurements are required to acquire results from the target image. Te authors in [76] proposed various image registration frameworks for enhanced outcomes based on picture longitudinal values. Tis study focused on removing noisy data from retinal vessels for enhanced outcomes using quadratic, lower order, and elastic transformation framework models that eliminate errors. Te study in [77] proposed fuorescein angiography (FA) and optical coherence tomography (OCT) methods that were identifed as scanning laser ophthalmoscope images for future extraction purposes. Te resultant extracted features were highly successful in generating global registration results of the complex blood vessel networks pertinent to retinal images. Also, deformable and intensity-based transformations were implemented to improve the motion magnitude of FA and OCT images.
Te study in [78] proposed a supervised multispectral fundus image registration (MSI-R-NET) method by analyzing various levels of blood vessel structures in the retina. It also observed the eye movement of retinal vessels for a short period and worked on image irregularity positions. Tese multispectral images were forecasted without classifcation and improved outcomes in the training and testing parts. Te retinal vessel complications lead to complete vision loss in adults. Te authors in [79] presented a novel volume of interest (VOI) for various levels of OT image registration, which ofered superior results in B-spline transformations, utilizing a stochastic gradient descent optimizer. Also, it used the Jacobian determinant transformations expansion on retinal vessels and monitored precise values. Te study in [80] mentioned the fact that complete image rotations, intensity, and scaling factors are all key challenges in image registrations. To achieve appropriate image feature augmentation, the feature descriptor statistical properties (FiSP) models were employed to ensure minimal scalability. Te resized parameters were regulated efciently, and the resultant outcomes were superior in comparison to traditional approaches. Te authors in [81] focused on enhancing the accuracy of retinal images using the SIFT (scale-invariant feature transform) registration method. Tis method helped in depicting an image of a higher pixel value without using noise data. Te automatic image scalability and image alterations were applied to the FIRE dataset to obtain robust values to enable refned scaling and brightness of images, ensuring enhanced performance.

Deep Learning Algorithms for Diabetic
Retinopathy Diagnosis: Case Studies

Use Case 1: Remote Image Analysis Was Used to Diagnose
Diabetic Retinopathy in Mexican Patients. Te problem of diabetic retinopathy occurs in young-and middle-aged people. Te study in [82] suggested the development of an assistive index framework for identifying diseases that are unlikely to be controlled. Tis study responded to the immediate demand for AI-based DL methods combined with radiographic imaging technology in Mexico to ensure the rapid expansion of artifcial intelligence (AI) methods in diagnosis, prognosis, and the provision of quality medical services. ARIA (automated retinal image analysis) is a basic premise of a web-based resource platform consisting of an image analysis module for referable and nonreferable DR classifcation. Te hospital administration provided highquality services to eye technicians to check the patient's condition and determine the severity of the illness. Both the ophthalmologists' and the ARIA system's results were used to determine the patient's condition. It is an extreme environment to recognize chronic eye diseases such as glaucoma and the DR diagnosis. It also required more ophthalmologists to meet the extremely high standards. To accomplish these methods one should be in online mode to save time which is necessary to raise the number of persons registered to a large extent. In [87], the authors discussed fractional-order flters for edge identifcation in biomedical imaging technologies. Te study employed left-sided and right-sidedfractional-order mask flters to fne-tune the image with whole edges detected without noisy data and the performance evaluation of fundus images on the STARE dataset. Tis process provides complete noise-free images for disease diagnosis.

Use
In the study in [88], the authors proposed a case study to prevent risk factors in diabetic retinopathy in diabetic patients at Tikur Anbessa hospital. A person sufering from diabetes and related complications fail to maintain the required glucose level in the body. On the contrary, such individuals tend to eat healthy diet which have contrareactions of increased blood pressure level and consequential neovascularization. Tis study suggested frequent health check-ups and glucose levels were used to identify diabetic retinopathy problems in the early years. Te study in [89], described the signifcant factors for the occurrences of diabetic retinopathy in those who had diabetes recognized in Debre Markos Referral hospital in Northwest Ethiopia. Various patients from this hospital maintained retinal vascular images examined using binary and multiple logistic regression approaches. Te primary aim of this study is to prevent the occurrences of this disease and conduct a systematic review of screening methods to maintain a healthy glycaemic level.

Various Issues, Challenges, and
Lessons Learned 6.1. Lessons. Te processing of medical images in a deeplearning environment ensures accurate results. In this study, various methods stated the performance level in 2D images.
It is still an ongoing process to be found in deep learning previous work problems, as shown below: (i) Most of the cases in traditional works lacked the high resolution of quality images in larger datasets. As a result, creating a proper dataset that collects images of the required quality is necessary. Tose images came from several sources, including public and private, and are combined to yield positive results. (ii) Improved image enhancement and contrast are needed to detect damaged retinal vessels. Because retinal images are small and noisy, it is difcult to identify the disease in the early stage. (iii) Te lack of standardization in data collecting is one of the most signifcant issues in medical image analysis. It is vital to remember that as the amount of data available grows, so does the necessity for big datasets to ensure that deep learning models produce realistic results. In medical image processing, transfer learning approaches were used to identify object detection and classifcation to achieve enhanced accuracy. But there are scenarios where these approaches failed to deliver optimal classifcation accuracy. In such cases, it becomes necessary to retrain the model to achieve enhanced accuracy.

Te Process of Incorporating DL Applications and
Telemedicine. In rural areas, neither medical nor hospitality resources are available to diagnose health issues, especially eye diseases, for which there are no routine screenings. Te integration of telemedicine with cloud data sources enables the use of AI techniques to upload images of the retina.

Pool-Based Data Sampling.
In this article, traditional image datasets are usually classifed as either deep learning or machine learning models. Te main concern is that it uses a large volume of records for DR classifcation and uses both a large amount of processing capacity and memory.

Low Computational Power and Network
Size. Two major factors contributed to the deep learning model's overftting needs: computational power and network size.
An object-based screening model is identifed as more comfortable in comparison to the image-based screening method popular in medical image processing that examines hemorrhages and other ocular lesions.
. Conclusion and Future Directions 7.1. Conclusion. Deep learning is regarded as an efective technique for ofering technical solutions in disease prediction and classifcation. Tis study discusses various machine and deep learning techniques for the early diagnosis of diabetic retinopathy. Tese studies were conducted in various geographical locations wherein various DL applications for medical image processing were implemented in the last century. DL has been used to achieve DR identifcation solutions employing medical imaging techniques such as classifcation, detection, segmentation, and registration. But, all these various applications of DL incorporating medical image processing methods require enhancement in accuracy, performance, and reduction in computational cost. Tis study presents an exhaustive review of the various deep learning and image processing-based techniques and relevant applications of the same in DR detection. It also enlists the challenges identifed in the earlier studies, namely, relevant to the availability of realtime data, computational power, network size, and various others. Finally, the study mentions certain recommendations for future scope in research that would enable diabetic retinopathy detection at an early stage. Tese research detections involved facilitating efective hybridization techniques and the implementation of advanced hyperparameter tuning methods to overcome some of the prominent challenges identifed in DR detection using DL techniques.

Future Directions.
Deep learning algorithms have gained immense momentum in improving diagnosis in medical imaging systems. Te traditional studies are concentrated on CNN models and deep-layered architectures to detect diabetic retinopathy. Furthermore, it specifes the use of the massive amount of data presented on image datasets, which was combined with retinal lesions to design a robust model for efective implementation. Due to the limited number of publicly available datasets, DR diagnosis remains a challenge. Te recent advancements in DL have produced promising classifcation results despite having difculty identifying the damaged lesions. For example, the study in [90,91] identifed segmented intraretinal variations. Images were classifed according to their severity level using the multiclass classifcation approach. Tis system performed well on various fundus image datasets and identifed the patient's condition. Tese techniques have promising potential in the detection of diferent eye disorders, namely, glaucoma and other intraretinal abnormalities. Another future enhancement is identifed DR at very large image datasets such as Messidor and EyePACS. To classify the severity level, this system completely relies on the quality of the images that are presented easily. Tose images are retrained using the transfer learning method to detect the lesions easily [92]. To develop a system with minimal confguration, overftting and computational cost need to be reduced.
To overcome the data argumentation issues mentioned in [93], use of more features on the large dataset need to be analyzed to improve accuracy. Both data preprocessing methods and feature enhancement techniques are combined to design a robust hybrid deep learning model for detecting the DR classifcation and detection. Tese methods were used to evaluate the information about diferent patients. It also provides the best outcomes for DR and other eye-related diseases in further proceedings.

Data Availability
Te datasets and material used to support the fndings of the study can be obtained from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.