Computational intelligence in eye disease diagnosis: a comparative study.

In recent years, eye disorders are an important health issue among older people. Generally, individuals with eye diseases are unaware of the gradual growth of symptoms. Therefore, routine eye examinations are required for early diagnosis. Usually, eye disorders are identified by an ophthalmologist via a slit-lamp investigation. Slit-lamp interpretations are inadequate due to the differences in the analytical skills of the ophthalmologist, inconsistency in eye disorder analysis, and record maintenance issues. Therefore, digital images of an eye and computational intelligence (CI)-based approaches are preferred as assistive methods for eye disease diagnosis. A comparative study of CI-based decision support models for eye disorder diagnosis is presented in this paper. The CI-based decision support systems used for eye abnormalities diagnosis were grouped as anterior and retinal eye abnormalities diagnostic systems, and numerous algorithms used for diagnosing the eye abnormalities were also briefed. Various eye imaging modalities, pre-processing methods such as reflection removal, contrast enhancement, region of interest segmentation methods, and public eye image databases used for CI-based eye disease diagnosis system development were also discussed in this paper. In this comparative study, the reliability of various CI-based systems used for anterior eye and retinal disorder diagnosis was compared based on the precision, sensitivity, and specificity in eye disease diagnosis. The outcomes of the comparative analysis indicate that the CI-based anterior and retinal disease diagnosis systems attained significant prediction accuracy. Hence, these CI-based diagnosis systems can be used in clinics to reduce the burden on physicians, minimize fatigue-related misdetection, and take precise clinical decisions.

Keywords: Human eye; Anterior eye; Retina; Ophthalmology; Eye diseases; Digital imaging; Segmentation; Computational intelligence; Classification

Copyright comment Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Graph

The eye is the most important organ of the human body, which is used to see the world by refracting the light that falls on objects. The characteristics of the human eye and visual system are similar to that of the camera. The cross-section of the human eye and visual system with anterior and posterior subdivisions are seen in Fig. 1. The human eye is composed of the retina, cornea, iris, pupil, and lens regions. The light emitted from the individual object focuses on the retina region through the transparent cornea, iris, and pupil structures. The eye contains light-sensitive cells along with optical nerves that allow the light into the human visual system. Typically, the iris region controls the light that reaches the human eye. After regulation, the light is obtained by the retinal photoreceptor cells that observe the information from the controlled light. This light information is transferred as nerve impulses and directed through the optical nerves to the brain. The human brain's response depends on the information captured by the eye. Hence, abnormalities existing in the human eye are viewed critically.

Graph: Fig. 1 Cross-section of the human eye (image source: The National Eye Institute, USA)

Eye disorders are the most significant health issue among the elderly population. According to a recent study [[1]], cataracts caused 45.5% of global blindness. Apart from cataracts, lifestyle-related diseases also cause anomalies in the human eye. Cholesterol and diabetes are the predominant lifestyle-related disorders that cause eye anomalies. The accumulation of cholesterol in the human eye is linked with multiple lifestyle-related disorders [[2]]. Cholesterol accumulation in the peripheral corneal stroma is known as corneal arcus (also known as arcus senilis). According to medical studies, corneal arcus is substantially linked with coronary heart disease [[3]], blood pressure, hypercholesterolemia [[4]], total cholesterol, low-density lipoprotein levels [[5]], and calcific atherosclerosis [[6]]. Another lifestyle-related disease diabetes also causes defects in the eye [[7]]. High blood sugar levels have weakened the retinal blood vessels (clinically referred to as diabetic retinopathy (DR)) and caused painless vision loss [[8]]. The DR is commonly known as non-proliferative DR (NPDR) and proliferative diabetic retinopathy (PDR) [[9]]. In the NPDR level, the disease can be mild, moderate to the extreme with various lesions, except for the development of new blood vessels. In the PDR stage, new blood vessels are grown along the retina region. Other than the DR, glaucoma and age-related macular degeneration (AMD) are also the major causes of blindness among persons aged 50 years and older [[1]]. Glaucoma is a condition of excessively high pressure in the eye that affects the optic nerve, which is essential for the human visual system. The AMD is a progressive eye abnormal condition in the macula region that can lead to permanent vision loss. Hence, the assessment of the retina and anterior (cornea) regions of an eye is essential to detect major abnormalities and prevent blindness.

Ocular disorders are commonly diagnosed by an ophthalmologist through the slit lamp assessment. A traditional slit lamp may also be used with a microscope to detect various abnormalities in the eye [[10]]. The slit lamp test helps to specifically examine the iris, lens, sclera, eyelid, and cornea parts. However, the interpretations of the slit lamp test may be influenced by the memory of the medical practitioner, inconsistency in grading, and the capability to keep records [[11]]. Hence, digital eye image processing and smart decision support systems are preferred as assistive methods in eye disease diagnosis. Corneal and retinal digital imaging techniques provide a non-invasive assessment of the optical nerves and vascular structure existing in the human eye. These methods are used in the early detection and diagnosis of ocular diseases. However, the evaluation of raw eye images is a time-consuming process due to the volume of patient data, and it also necessitates specialist skills. Therefore, process automation is a critical requirement in corneal and retinal eye image analysis.

In recent years, numerous intelligent eye disease diagnosis systems have been developed by utilizing computer technology advancements. These systems can reduce the stress related to the clinical workload of an ophthalmologist, fatigue-related false positive detection, and even help to make accurate clinical decisions [[12]]. Medical studies also indicate the importance of early diagnosis. A clinical study shows that early detection of DR can decrease the risk of serious vision loss by up to 90% [[13]]. The conventional method of DR diagnosis is manual, costly, and involves skilled ophthalmologists [[14]–[17]]. Early detection and precise recognition of ocular anomalies are vital for the effective diagnosis of eye abnormalities to ensure intensive care and further surgical procedures. Therefore, there is a need for reliable and cost-effective intelligent screening systems to identify retinal abnormalities at an early stage [[18]]. Although there are significant works on automated retinal image analysis, research on corneal image processing-based intelligent screening systems has lagged for several reasons such as the deficiency in corneal image analysis approaches and the existence of multiple abnormalities in the cornea. Hence, there is a massive need for intelligent corneal image analysis methods compared with retinal image analysis methods.

In this paper, we have discussed the comparative analysis of CI-based corneal and retinal eye disease diagnosis systems. The major objective of this paper is to provide a broad research view about the growing CI-based eye disease diagnosis field, which is emerging by the increasing data availability and computational power. This paper highlights the need for automated analysis of digital format human eye images and the potential barriers considered to its advancement. Furthermore, we give a performance comparison of various pre-processing, segmentation, and CI-based diagnosis methods used for eye disease diagnosis. This article will be extremely useful for computer scientists and medical practitioners, as multiple research problems are useful for them to assist in finding effective solutions to the automated eye disease diagnosis research domain. This paper is structured as follows: The background of eye imaging modalities and an overview of CI systems are discussed in Sect. 2. The state-of-the-art eye image pre-processing, segmentation, and CI-based diagnosis methods are briefed in Sect. 3. Section 4 compares the performance of the above methods. Finally, Sect. 5 concludes the paper.

Medical imaging is used in ophthalmology over the past 100 years, and it plays a vital role in clinical eye care and disease management. The first research work on retinal image analysis [[19]] published in 1973 was focused on retinal vascular segmentation. In the early 1980s, large-scale investigations and the development of medical imaging systems were initiated. In 1984, Baudoin et al. [[20]] proposed a diagnosis method to spot the DR lesions in the retina using image analysis. Over the past 30 years, several automated diagnosis systems were developed for detecting eye disorders. These systems can detect different types of eye diseases such as DR [[21]], glaucoma [[22]], AMD [[23]], and cataract [[24]] using several image analysis techniques. Moreover, automated diagnostic methods have the potential to be employed in large-scale screening programs, saving the clinician's time and diagnosis cost, also eliminating observer bias. In recent years, a large set of medical image analysis-based automated diagnostic approaches has emerged in clinical ophthalmology to offer remarkable insights into eye illnesses diagnosis based on the analysis of datasets with millions of data points. A large set of advanced medical image analysis-based techniques were used to observe high-resolution information related to the anatomic and functional changes of the eye.

Various imaging modalities were used in automated eye disease diagnostic systems. The most important imaging modalities for different eye disease diagnoses are listed in Table 1. Slit-lamp imaging is a conventional method for observing the pathological changes in the cornea, iris, and lens of the eye. However, conventional slit-lamp systems are not portable, and patients need to visit an ophthalmologist for their eye examination. Nowadays, smartphone-based slit-lamp devices [[25]] are used to overcome the limitations of conventional slit-lamp systems. Next to slit-lamp devices, digital fundus photograph (DFP) is the most preferred imaging modality due to the reason that the information acquired from the eye fundus can be valuable in diagnosing the conditions such as hypertension, heart disease, peripheral vascular disorders, stroke, and DR [[26]]. The Heidelberg retinal tomography (HRT) is another well-known imaging modality used for the clinical examination of the optic nerve [[27]]. The HRT technology is preferred for its accurate and repeatable analysis of the posterior fundus region as well as the optic nerve, which is useful for glaucoma diagnosis. This imaging technology added a new dimension to the medical practitioner's posterior fundus assessment [[28]]. The optical coherence tomography (OCT) is another commonly used medical imaging method in eye disease diagnosis. When compared with the DFP, the most significant advantage of OCT is that it offers quantitative depth information, which allows for a 3D scan of the target eye region. As a result, in OCT, diseases with topological abnormalities can be detected in vivo [[29]]. In recent times, visible wavelength (VW) images of an eye were also used to diagnose the eye disorders such as cataracts and corneal arcus automatically [[11], [30]]. A small digital camera-based VW imaging modalities have the potential to improve eye healthcare access in middle and low-income countries also.

Table 1 Medical imaging modalities used in eye disease diagnosis

The CI is a term that refers to the usage of computers to learn under new conditions, which leads to the perception that possesses one or more properties of reasoning such as generalization, identification, association, and abstraction. These intelligence approaches are largely used in medical diagnosis and prognosis [[31]]. The CI systems are effectively used in eye disease diagnosis also. The general block diagram of the CI-based eye disease diagnosis system is shown in Fig. 2. These systems used the digital image of an eye as input, and the eye abnormalities are diagnosed using the features extracted from the input images. These features are extracted from the ROI of an eye image with the support of image segmentation algorithms. After system development, the diagnosis results of the CI-based decision support systems are compared with the ophthalmologist's decision to evaluate the diagnosis accuracy.

Graph: Fig. 2 General block diagram of CI-based eye disease diagnosis system

The CI-based methods have increasing popularity among researchers due to the ability to handle massive clinical data and information even under uncertain conditions. There are three major types of learning in CI-based systems: supervised machine learning (ML), unsupervised ML, and deep learning (DL) [[32]]. Neural networks, genetic algorithms, and fuzzy systems are the key components of the CI-based diagnosis systems. The neural network is the conventional supervised approach in CI, used to approximate the functions or data classification [[33]]. The genetic algorithm [[34]] is the search algorithm, based on the biological genetic operator's cross-over and mutation. This algorithm optimizes the fitness function-based classification problems using a random-guided method. The fuzzy logic system is the state-of-the-art method, which uses fuzzy set theory [[35]] to solve classification problems. The DL is the advanced method in CI, which can learn from the data samples by updating the weights of the neural layers over the optimization function. The DL is highly self-learning and useful in real-time clinical diagnosis because it can process unstructured data also. Moreover, the DL-based diagnosis systems can classify the eye images according to the disease category without any manual feature extraction.

In this comparative study, the authenticity of the electronic sources and the language is the major criterion for the selection of articles for review. The research articles written in the English language are mainly included in this comparative study because English is the globally recognized language of science as well as biomedical research domains. The year of publishing of the articles included in this study is up to 2022. To conduct this comparative study, we used the research article resource platforms such as IEEE Xplore, PubMed, Web of Science, and Science Direct. The research articles reviewed in this study are categorized based on the processing methods of the CI-based eye disease diagnosis systems. This categorization includes the following types: (a) eye image pre-processing methods, (b) ROI segmentation methods, (c) CI methods for anterior eye abnormalities diagnosis, and (d) CI methods for retinal abnormalities diagnosis. The processing methods related to the above four major categories are illustrated in Fig. 3.

Graph: Fig. 3 Overview of CI-based eye disease diagnosis

VW imaging is a simple and cost-effective method used to diagnose anterior part eye abnormalities [[11], [30]]. The specular reflection (SR) is a key challenge in VW anterior eye images. The SRs are the peak intensity points in the VW eye images. The main cause of SR reflection is the light source used for eye image acquisition and the reflective characteristics of the cornea. The SRs are a common problem for both biometric and biomedical eye imaging systems. These reflections affect the segmentation of the iris region and lead to wrong diagnosis. Hence, there is a need to minimize the SR. Few pre-processing techniques are available for the elimination of SRs in eye images. Z. He et al. [[36]] developed a method using bilinear interpolation combined with an adaptive threshold for removing the SR regions in the eye images obtained in a constrained setting. The key purpose of this approach is to support rapid segmentation of the iris. Therefore, this approach performed the SR removal with fewer processing steps.

Walid Aydi et al. [[37]] suggested a new approach for eliminating SRs using the linear interpolation technique. This method yields relatively better outcomes than the bilinear interpolation-based SR removal approach. The algorithm developed by Tsai et al. [[38]] minimized and isolated the SRs utilizing a top-hat filter. To reduce the difficulty in the elimination of SR, Radman et al. [[39]] developed a simple method based on image complement. In this technique, the input anterior eye image has complemented, the holes in the background image are filled, and the resulting image is complemented again to build a reflection-removed image. Wang et al. [[40]] developed an in-painting process-based method using the Navier–Stokes equations to eliminate the reflections by filling the gaps of SR regions. Jamaludin et al. [[41]] proposed a bottom-hat filtering-based solution to remove SRs in eye images. To study the different conditions of SRs, a simulation eye model was developed by Kinsman and Pelz [[42]]. The above SR removal methods are mainly developed by using the VW eye image datasets for biometric applications and have not attained the significant structural similarity index measure (SSIM) and low mean square error (MSE) values. Few recently proposed SR removal studies [[43]] attained significant SSIM and low MSE values. Hence, these methods can be suitable for anterior eye disease diagnosis systems. Figure 4 shows the anterior segment eye image collected from the public database UBIRIS.v1 [[45]] and the reflection removed image using the recently developed morphological filtering-based SR removal method [[43]].

Graph: Fig. 4 SR removal in anterior eye images. a Eye image with SR. b SR-removed image

Fundus image analysis is a commonly used method for retinal disease diagnosis. The major objective of pre-processing the fundus eye image is to provide an enhanced resulting image, which is appropriate for further image analysis. The major challenge of retinal image enhancement is the processing of the contrast variation inside the fundus images. Hence, there is a need for a pre-processing method intended for image normalization and contrast enhancement to perform accurate fundus analysis. Recently, few studies [[46]] used Contrast Limited Adaptive Histogram Equalization (CLAHE) technique for enhancing the fundus eye images. The CLAHE is a modification of the Adaptive Histogram Equalization (AHE) technique with modified enhancement computation by imposing the user-specific level on the local histogram height. Hence, the enhancement is controlled in the uniform areas of the fundus image, which avoids the enhancement of noise and decreases the edge shadowing limitation of the conventional unlimited AHE. The performance evaluation results reported in [[48]] evidenced that the CLAHE-based methods attained better enhancement outcomes compared to other contrast enhancement techniques. The normal fundus image from the MESSIDOR database [[49]] and the contrast-enhanced outcome of the same image using CLAHE are shown in Fig. 5.

Graph: Fig. 5 Contrast enhancement of fundus eye images using CLAHE. a Normal fundus image. b Contrast-enhanced fundus image

Cholesterol-associated eye abnormalities usually occur in the iris circle of the anterior eye region. Due to this reason, in automated anterior eye disease diagnosis systems, the iris circle is segmented from other parts of the eye. Numerous techniques are available for iris segmentation. Daugman [[50]] developed a state-of-the-art method of iris segmentation that is preferred in most of the real environment iris recognition systems. This approach locates the inner pupil border and outer sclera boundary of the iris using the integrodifferential operator. Wildes [[51]] established an iris segmentation system using the mathematical approach Circular Hough Transform (CHT). In this system, the inner and outer boundaries of the iris were computed using the gradient-based binary edge map followed by the CHT. Wilde's approach is more robust for noise fluctuations also. Camus and Wildes [[52]] developed a method for identifying the iris and pupil boundary of an eye. This method uses a multi-resolution technique to quickly and reliably recognize the boundary contours of the eye image. Moreover, this approach fits well even in cases of eye images with inadequate contrast, SRs, and distorted perspectives.

In VW eye image processing, H. Proenca developed a fuzzy k-means clustering-based algorithm [[53]] to segment the iris circle in eye images acquired at the VW spectrum. This method attained a significant accuracy in iris circle segmentation. However, the computational complexity of fuzzy k-means clustering is more. Due to this reason, this method is not suitable for rapid iris circle detection. The recently proposed robust iterative scheme for Hough's transform [[54]] attained significant accuracy in iris circle segmentation. The above studies show that the CHT is a rapid and precise method for iris segmentation. As a result of rapidness and precision, the CHT is preferred in most of the anterior eye abnormalities detection systems [[11], [30], [55]]. Usually, the SR-removed image is used as the input for the CHT. Figure 6 shows the stage-by-stage outcome of iris circle detection using a CHT-based technique [[11]]. The detected iris circle is used as the ROI for anterior eye abnormalities detection.

Graph: Fig. 6 ROI segmentation process in anterior eye images using CHT. a Anterior eye image. b Canny edge map of the anterior eye image. c Segmented ROI using CHT

Eye diseases such as glaucoma, DR, and macular degeneration (MD) can be rapidly diagnosed by quantitatively analyzing the retinal fundus images. Segmentation of retinal blood vessels is an essential processing step to extract the clinically relevant features such as the blood vessel density, length, and tortuosity from the fundus images. Conventionally, the blood vessels in the retinal fundus images can be manually segmented by ophthalmologists. The manual segmentation of retinal blood vessels is a time-consuming and difficult task. Therefore, automated retinal blood vessel segmentation methods are used to overcome the limitations of manual segmentation methods. The retinal blood vessels can be automatically isolated using an average filter as shown in Fig. 7. The average filter can isolate the retinal blood vessels rapidly. However, this method is not suitable to isolate the minor regions in the retinal blood vessel structure. Hence, there is a vital need for accurate computer-aided retinal vessel segmentation methods.

Graph: Fig. 7 Blood vessel segmentation using an average filter. a Contrast-enhanced fundus image. b Segmented blood vessels

Recently, there is immense progress in the development of new computer-aided methods for retinal blood vessel segmentation. Various segmentation approaches [[56]–[63]] have been developed due to the importance of automated vessel segmentation in retinal disease diagnosis. The retinal blood vessel segmentation methods can be split into two categories: supervised approaches and unsupervised approaches. Supervised approaches segmented the retinal blood vessels by classifying each pixel in the retinal image, and they comprise two fundamental phases. In the first phase, a feature vector is extracted from the each pixel of the fundus image. The gray-level features are commonly extracted from the fundus image. However, other features like the moment invariant feature [[64]], morphological gradient feature, and line strength features [[65]] are also been used in several studies. After feature extraction, classifiers such as neural networks [[56]], support vector machines (SVM) [[65]], and a Bayesian classifier [[66]] are used to identify whether a pixel belongs to the vessel structure or background region. Supervised methods attained significant performance in retinal vessel segmentation. However, supervised methods require several ground truth images, prior knowledge, and more model training time.

The unsupervised methods generally attain a faster blood vessel segmentation and instinctive also. However, commonly the blood vessel segmentation accuracy of unsupervised methods is lower than the supervised methods. Deep neural networks are also employed in retinal vessel segmentation to eliminate the requirement of complex feature engineering, which is essential to process the fundus images with vessel structure complexity. A deep neural network-based retinal segmentation approach developed in [[67]] attained a significant area under the receiver operating characteristic (ROC) curve value (> 0.99) and good segmentation accuracy (> 0.97). However, this method is computationally more expensive than the traditional blood vessel segmentation algorithms due to several deep neural layers. The recently proposed retinal blood vessel segmentation approach [[68]] performed the retinal blood vessel segmentation using the weighted line detector and the hidden Markov model. This method is robust and attained a significant segmentation accuracy (0.9475). The recent retinal vessel segmentation methods attained significant segmentation accuracy. However, accurate retinal vessel segmentation is still a computationally challenging task due to the structural complexity of vessel networks with numerous biological tissues, which causes the pathological regions difficult to segment.

Cataracts and corneal arcus are the major abnormalities existing in the anterior eye. Few optical imaging-based intelligent systems are available for the diagnosis of anterior eye abnormalities. Ramlee and Ranjit [[30]] developed an iris recognition and Otsu's algorithm-based automatic system for detecting the presence of corneal arcus in VW eye images. For the quantification of corneal arcus deposition, Nasution and Cahya [[69]] developed a digital image acquisition system. This imaging device acquires an eye image by a web camera with a USB-powered LED illumination. The quantification method proposed by Nasution et al. [[70]] identified the arcus stages using arc length and thickness features derived from the anterior eye images. However, this quantification approach is not focused on classifying healthy eye images from unhealthy eye images. For the diagnosis of cataracts in eye images, Jagadish Nayak [[71]] developed a SVM-based classification system. In this system, the input images were categorized as healthy, cataract, and post-cataract stages using optical features of an eye.

Acharya et al. [[72]] established a multiclass classification framework to classify healthy, cataract, and post-cataract stage images using fuzzy k-means clustering along with a backpropagation algorithm (BPA). The method developed by Supriyanti et al. [[73]] recognized the presence of a cataract in anterior eye images using a SR test. The main drawback of the above methods is a single eye abnormality diagnosis limitation. Hence, methods with multiple eye abnormalities diagnosis ability were preferred in recent times. The radial basis function-based system model [[74]] compacts with the diagnosis of multiple types of eye diseases such as cataracts, corneal arcus, haze, and iridocyclitis. This model employs a fuzzy k-means clustering algorithm for eye disease diagnosis. However, the classification unit of the above system was limited with single-eye abnormality classification. To overcome the binary class and single-eye abnormality classification limitations, a ML-based approach [[11]] was developed using VW eye image analysis. This method focused on the diagnosis of three stages of eye disease such as healthy, one eye abnormality (arcus), and more than one eye abnormality (arcus and cataract) in the anterior region. This method attained significant classification accuracy in multiple eye abnormalities diagnoses. However, still, there is a need for a higher-order intelligent classification system which performs similarly to an ophthalmologist in the detection of multiple eye abnormalities.

Recently, smartphone applications also developed to detect anterior eye abnormalities. The digital eye images acquired by smartphone devices are simple and useful solution for ophthalmologists, particularly in rural areas with inadequate healthcare facilities. In [[75]], a mobile phone application was developed to assist the people for eye image acquisition intended for the identification of eye diseases. Moreover, this application is also aimed to support the clinicians through telemedicine-based eye examinations. A study on smartphone eye imaging [[76]] proved that a detailed background information about eye disease is an important aspect to improve the trustworthiness of any artificial intelligence (AI)-based eye disease diagnosis. Moreover, the "Eye Guide" mobile application developed in the above study can be used in AI-based eye disease diagnosis systems and telemedicine-based eye examinations. Latest advancements in smartphone cameras and mHealth applications transform a smartphone into a medical diagnosis system.

The DR is the major abnormality existing in the retina region of the human eye. To detect the DR stages, optical image processing of the retina is commonly used. The DR quantification and classification using eye features such as blood vessels and exudates have been performed since 1982, and various diagnosis systems [[77]–[81]] were also developed using fundus data analysis to classify the DR stages. To identify the quality of fundus images, Yu et al. [[82]] developed a method that mechanically examines the input retinal fundus image, if the image is suitable for computer-aided DR detection. This method uses various image characteristics such as histogram, texture, and vessel density for quality evaluation. The initial level problems involved with blindness and DR can be detected by identifying microaneurysms. For this purpose, Agurto et al. [[83]] developed a three-stage system for early DR diagnosis. Most of the fundus imaging systems used the technique of microaneurysm spotting to identify the initial stage of DR. However, microaneurysm detection can be useful to identify the mild stages of DR only. Hence, researchers analyzed the other symptoms in the retina.

The method proposed by Yu and Chen [[84]] addressed the use of Doppler's imaging in vivo angiography for 3D analysis of an eye image. This method uses the transition of blood circulation velocity to graph the retina region and choroid vessels in addition to the typical optical Doppler tomography images. Retina walls are weakened, cracked, and also bleed due to DR. Hence, retinal blood vessels are considered as an important ROI for DR detection. N. Singh and Tripathi [[85]] proposed an image analysis method, which is intended for retinal blood vessel-based DR detection. In this method, the structural modifications in multiple elements of the human retina were studied by the blood vessel analysis method. Computer vision–based methods are also used to build various DR diagnosis systems. Sundhar and Archana [[86]] proposed a machine vision approach to detect DR using spatial features and artificial neural network (ANN). The various DR diagnosis systems developed in [[87]–[90]] achieved good sensitivity in DR detection; however, the accuracy of these methods is less compared to their sensitivity. To improve the accuracy of DR diagnosis and resolve the bagging and boosting issues in DR lesion classification, in recent times, advanced ML-based approaches are also preferred in DR classification systems [[91]].

Numerous CI systems for DR diagnosis were developed using ML approaches such as the SVM [[93]], Naïve Bayes classifier [[94]], decision tree [[95]], k-nearest neighbor (KNN) classifier [[96]], ANN [[98]], and ensemble classifiers [[99]]. The SVM is used in the DR classification systems [[100]] to overcome the training and classification-related complexities. However, the conventional SVM is not suitable for large dataset analysis and multi-stage DR classification. Jelinek et al. [[102]] developed a multi-stage ML-based classification system to identify the different stages of DR, which used the outcome of the fusion of visual dictionary-based detectors created by the set of ROIs. The random forest-based approach [[103]] also supports multiclass classification of the DR. However, this approach requires complex decision rules. A genetic programming (GP)–based DR detection method [[104]] was focused on the diagnosis of small retinal lesions also. However, the GP-based method is not suitable for long-term DR prediction. Recently, long-term predictive models [[105]] were developed for the diagnosis of DR using the medical features acquired from the data warehouse and ML techniques.

ML-based approaches eliminate bagging and boosting issues in DR classification systems. However, ML-based DR detection methods have shown meager generalization through the feature extraction process and classification by the deployment of small datasets. Moreover, ML-based DR detection methods required more training time, which causes inaccuracy in DR prediction when using large datasets [[106]]. Therefore, DL, a new field of ML, is used in recently developed DR detection systems. DL-based DR diagnosis systems can handle the DR classification using small datasets, with the help of accurate data processing approaches. Convolutional neural network (CNN) [[107]] is largely used in DL-based DR detection systems because of its unsupervised feature extraction ability. The DL-based DR diagnosis systems [[108]] developed using CNN and achieved good specificity, but the precision of these systems is less. Moreover, CNN-based DL methods were incompetent to attain better outcomes in terms of feature extraction and optimizing the classifier. To overcome the feature extraction and optimization-related issues, N. Gundluru et al. [[110]] proposed an enhanced DL model, which used the principal component analysis (PCA) method for dimensionality reduction and the Harris hawks optimization algorithm to optimize the feature extraction process.

Smartphone fundus photography [[111]] is the latest advancement in retinal image analysis, which is used to examine the optic nerve and retinal blood vessels of an eye. The major advantage of this technique is the wide accessibility of smartphones, which enables to observe and document the optic nerve, and macula region changes in diverse settings, which were previously impractical. Numerous CI systems were developed to detect retinal abnormalities by using smartphone fundus images. In [[113]], neural network-based CI system was developed to classify the abnormal and healthy retinal images. Rajalakshmi et al. [[114]] developed a smartphone image-based CI system to grade the DR. This method attained high sensitivity for identifying the DR. Recently, a KNN-based mobile CI system [[115]] was developed to analyze the eye images taken by the smartphones for the diagnosis of retinal diseases. To validate the efficiency of a smartphone imaging-based CI system, B. Sosale et al. [[116]] conducted a study, which enrolled 922 subjects with diabetes mellitus condition. Retinal images were captured from each subject using the fundus-on-phone (FOP) camera. These images were used as the input for an AI-based diagnosis system, and the diagnosis result was recorded (normal or DR). After this, the diagnosis outcome of the AI-based system was compared with the manual diagnosis results of five retinal specialists. The comparison results show that for any type of DR, the sensitivity and specificity of the AI-based diagnosis system were 83.3% and 95.5%. The above experimental results evidenced that the smartphone image-based CI systems can be suitable for clinical applications also.

The CI-based eye disease diagnosis systems needs a valid dataset with a sufficient scale of information. Commonly, hospitals maintained the clinical eye image datasets with a scale that varies from hundred thousand to millions of scans. However, datasets maintained by the hospitals are inaccessible to the researchers, because of the barriers related to access and usability. The access-related barriers include governance barriers, cost barriers, and time barriers. Barriers related to usability include dataset format, quality, and image labeling issues. To overcome these barriers, researchers preferred publicly available eye image datasets. The UBIRIS.v1 [[45]] is a freely accessible database, which includes colored photographs of the anterior part of the human eye. The eye images in this database were obtained using visible light (400–700 nm) illumination. The noise factors particularly related to reflections, luminosity, and contrast were minimized in this database during the image acquisition process. Few anterior eye disease diagnosis systems [[30], [55]] used this database for model development and validation. However, this freely accessible database does not include medically categorized eye images. Hence, recent research works [[11], [117]] established a medically categorized dataset for anterior eye abnormalities diagnosis system development and validation.

In view of retinal disease diagnosis system development, numerous fundus image databases were publicly available for research purposes. A publicly available eye image database [[118]] contains the retinal fundus and the OCT images, which were used for vessel-based registration. The DR HAGIS database [[119]] is another standard publicly available fundus image database, which contains eye images with DR, hypertension, AMD, and glaucoma disease conditions. This database was mainly created to assist the development of retinal vessel segmentation algorithms required for retinal disease screening programs. The DRIVE database [[120]] is the benchmark database used to perform comparative studies on retinal blood vessel segmentation in fundus images. The eye images in this dataset were collected from the DR screening program conducted in the Netherlands. The REVIEW database [[121]] created by the researchers of the University of Lincoln, UK, was also used as a ground truth dataset for numerous retinal vessel segmentation approaches. For the purpose of rapid retinal vessel segmentation, the STARE database [[122]] was developed by scanning and digitization of the retinal photographs. The quality of eye images in this dataset is comparatively lower than the other standard public eye image datasets due to low resolution.

The MESSIDOR [[49]] is a golden standard database used in many CI-based DR diagnosis systems. This database includes 1200 fundus images of four DR classes collected by three different departments of ophthalmology. The MESSIDOR-2 database [[123]] is another golden standard collection of DR eye examination results, individually consisting of two macula-cantered fundus eye images (one per eye). The DIARETDB0 (calibration level 0) [[124]] and DIARETDB1 (calibration level 1) [[125]] are the standard fundus image database with numerous healthy and diabetic eye images. These public datasets were used as a benchmarking testing protocol for DR detection methods. For the purpose of specific disease diagnosis, the retinopathy online challenge database [[126]] was created to find the best specific algorithm for microaneurysms detection. A non-proprietary telemedicine system "EyePacs" [[127]] was created with a large dataset (35,126 retinal eye images) collected from various diabetic screening programs conducted in California. This system was used for storing and managing retinal eye images with different disease conditions. The Indian DR image dataset (IDRiD) [[128]] is a recently developed retinal fundus image dataset for the purpose of evaluating automatic DR detection and grading methods. This is the first eye image dataset which represents Indian clinical data. This dataset contains typical DR lesions and healthy retinal regions with pixel-level annotation. Recently, a smartphone-based retinal eye image montaging database [[129]] was also created for evaluating smartphone-based DR diagnosis systems. The technical details of the publicly available standard eye image datasets are provided in Table 2.

Table 2 Technical details of publicly available standard eye image datasets

Commonly, the performance of anterior eye image pre-processing methods was evaluated using the SSIM. The SSIM is a complete comparison metric intended to address the limitations of standard peak signal-to-noise ratio and mean squared error methods. The SSIM is used to determine the similarities between the original input image and the reflection-removed eye image. The SSIM value of zero means there is no similarity between the original and the reflection-removed eye images. If the SSIM is 1, it means that the reflection removed image is 100% identical to the original image, and the reflection removal algorithm does not change the pattern of the original image. Apart from the pre-processing methods, the accuracy of ROI segmentation in eye images was also validated by human observation. The ROI segmented by the automated segmentation algorithm is matched with the manually segmented ground truth image to compute the segmentation accuracy. The segmentation accuracy is an important measure for accurate classification of eye diseases.

The confusion matrix-based classification performance measures were broadly used to evaluate the CI-based eye disease diagnosis systems. The confusion matrix matches the classifier's prediction with an ophthalmologist's actual diagnosis (ground truth) to evaluate the classifier's performance. The confusion matrix for two-class classification is represented in Fig. 8.

Graph: Fig. 8 Confusion matrix used for the evaluation of CI-based eye disease diagnosis systems

From the confusion matrix shown in Fig. 8, four types of potential outcomes can be obtained. The possible outcomes are mentioned as follows:

• TP: If the predicted diagnosis is unhealthy, the result is TP for an unhealthy eye image.
• FN: For an unhealthy eye image, if the predicted diagnosis is healthy, the result is FN.
• TN: For a healthy eye image, if the predicted diagnosis is healthy, the result is TN.
• FP: If the predicted diagnosis is unhealthy, the result is FP for a healthy eye image.

Using the TP, FP, TN, and FN values, the following performance measures can be computed using Eqs. (1)–(3).

1 $$\mathrm{Sensitivity}(\mathrm{Se})=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

Graph

2 $$\mathrm{Specificity}(\mathrm{Sp})=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$

Graph

3 $$\mathrm{Accuracy}(\mathrm{Acc})=\frac{\mathrm{TP}+\mathrm{TN}}{\mathbf{T}\mathrm{P}+\mathbf{F}\mathbf{P}+\mathrm{FN}+\mathrm{TN}}$$

Graph

The receiver operating characteristic (ROC) curve is an additional performance evaluation metric used for validating the CI-based DR detection systems. The region under the ROC curve encircles an area under the curve (AUC), which is a sophisticated approach for measuring the diagnosis accuracy of the DR predictive model. The AUC performance measure was used for establishing substantial comparisons, if a robust imbalance among DR classes occurred. According to the report [[130]], a diagnostic system with an AUC of 0.5 represents that there is no perception in classification. The AUC between 0.7 and 0.8 represents that the classification result is an acceptable level. Similarly, if the diagnosis system shows the AUC ranges between 0.8 and 0.9, then it indicates an excellent classification result, and if the AUC is larger than 0.9 means it can be categorized as outstanding classification performance.

The comparison of state-of-the-art eye image pre-processing methods is shown in Table 3. The eye image pre-processing methods were compared based on the performance measure SSIM and the technique used for pre-processing. The nearest neighbor algorithm is commonly used in SR removal pre-processing step of eye image analysis. The interpolation of the nearest neighbor-based SR removal method is not the smoothest one. Hence, the bilinear interpolation technique along with the adaptive threshold [[36]] was used in SR removal to attain smoother interpolation. This method attained a significant SSIM of 0.9660. However, visible artifacts were created in the eye images after interpolation process. To reduce the processing steps of interpolation, a linear interpolation technique [[37]] is used in the SR removal. This method attained better SSIM (0.9768) compared with the bilinear interpolation-based SR removal method [[36]].

Table 3 Comparison of eye image pre-processing methods

The top-hat [[38]] and bottom-hat [[41]] filters were used in the SR removal process to remove the reflections caused by the sources other than the VW light illumination. However, the low SSIM value (0.8170) shows that the top-hat filter-based reflection removal process significantly affects the shape of the input eye image. Compared with the top-hat filter, the bottom-hat filter affects the shape of the image lesser (SSIM of 0.9298). Due to this shape-related limitation, the top-hat and bottom-hat filter-based methods are not suitable for medically categorized eye images. The pre-processing method proposed by Radman et al. [[39]] attained a significant SSIM of 0.9515. However, this method was not tested using medically categorized eye images. The recently proposed morphological filtering-based method [[43]] attained the SSIM of 0.9962, which is comparatively better than the other state-of-the-art SR removal pre-processing methods. Moreover, this method was tested using medically categorized VW eye images also.

The comparison of recent studies in eye image ROI segmentation is provided in Table 4. The eye image ROI segmentation methods were compared based on the segmentation accuracy and the technique used for segmentation. In the early stage anterior eye image-based systems [[55]], the integro–differential operator [[50]] was used as a rapid ROI segmentation technique. However, this technique attained less segmentation accuracy (88.23%) for VW eye images. Hence, the CHT is preferred as a ROI segmentation approach in recent anterior eye disease diagnosis systems [[11], [30], [55]]. Table 4 shows that the CHT-based segmentation technique attained the significant segmentation accuracy of 97.97%, which is comparatively better than the other state-of-the-art ROI segmentation techniques. However, the CHT-based methods attained the good segmentation accuracy by using more iterations for iris boundary detection.

Table 4 Comparison of ROI segmentation methods

In retinal disease diagnosis systems, the blood vessels are considered as the ROI for feature extraction. The line detectors [[57]] were commonly used to segment the retinal blood vessels. The line detectors can segment the tiny blood vessels also. However, false blood vessel detection due to noises is the major issue in a line detector. To eliminate noise and false blood vessel detection, line detectors with multiple masks [[58]] were used in retinal blood vessel segmentation. The ANN and random-walk-based segmentation approaches [[56], [59]] attained significant segmentation accuracy (> 94%). However, these methods require extensive training for the classification of blood vessel pixels. Morphological operators were also used in retinal blood vessel segmentation [[61]]. The segmentation accuracy of the morphological operator-based method [[61]] is significant (> 95%). However, this method requires a guided thresholding process to segment the retinal blood vessels. Recently, DL-based approaches [[60], [62], [67]] were also used in retinal blood vessel segmentation. Table 4 shows that a deep neural network-based blood vessel segmentation approach [[67]] attained the segmentation accuracy of 97%, which is comparatively better than the other state-of-the-art segmentation approaches. Moreover, the feature extraction of DL-based methods is an automated one. However, the DL-based methods require extensive training for classifying the blood vessel pixels.

The comparison of CI-based anterior eye disease diagnosis systems is provided in Table 5. Diagnosis systems for anterior eye abnormalities such as corneal arcus, cataract, and corneal haze were focused in this comparative analysis. These systems were compared based on the performance measures classification accuracy (Acc), sensitivity (Se), specificity (Sp), and the computational techniques used for anterior eye disease diagnosis. The diagnosis systems developed in [[70]], and [[73]] attained significant classification accuracy (> 90%). However, the classification outcome of these systems was limited with single anterior eye abnormality diagnosis. The fuzzy K-means clustering-based CI systems [[72], [74]] deal with multiple stages of eye diseases and attained 100% specificity. However, the classification outcome of these systems is binary class (normal/abnormal) and restricted with the diagnosis of various stages of cataract abnormality only. Moreover, the overall classification accuracy of these methods is less compared with the ML-based classification system [[11]]. The ML-based classification system [[11]] focused on the diagnosis of more than one abnormality existing in the anterior eye region and attained 96.96% classification accuracy.

Table 5 Comparison of CI-based anterior eye disease diagnosis systems

The comparison of retinal eye disease diagnosis systems is shown in Table 6. The BP-ANN-based system [[98]] attained significant classification accuracy, sensitivity, and specificity (> 95%). However, this system was focused on the binary classification (normal/abnormal category) only. The fuzzy random forest-based method [[99]] classified the DR classes by using dominance and rough set-based balanced rules. This method attained the classification accuracy of 84%, which is comparatively lower than the accuracy of few recent DR diagnosis systems. The SVM, KNN, and binary trees-based intelligent classification framework [[100]] attained 100% specificity in DR diagnosis using red lesion and bright lesion features. However, this method attained less sensitivity (83.67%). The SVM-based method [[101]] attained significant classification accuracy (95.1%) in multiple retinal lesion detection. However, this method is not suitable to analyze large-scale DR data due to the limitation with texture analysis. The random forest classifier-based diagnosis system [[103]] attained good DR classification accuracy (99.83%) using Haralick, multi-resolution features extracted from the fundus images. This classification accuracy is significantly better than the other state-of-the-art DR diagnosis systems. However, this method required more processing steps due to decision trees.

Table 6 Comparison of CI-based retinal eye disease diagnosis systems

The GP-based diagnosis system [[104]] attained significant classification accuracy, sensitivity, and specificity (> 95%). However, this method is focused only on microaneurysms detection and is not suitable for effective DR grading. The ML-based study [[105]] focused on the vision‑threatening DR using the analysis of medical features from the data warehouse. Nevertheless, this study was not focused on the funduscopic findings. Recently developed DL and CNN-based intelligent diagnosis systems [[108]]-[[110]] attained significant classification accuracy (88–96.7%). However, the classification framework of the CNN-based CI systems is slower due to the several layers of deep-CNN. The AI-based smartphone eye disease diagnosis systems [[114]]-[[116]] attained significant specificity (up to 95.5%) in DR detection. These diagnosis systems make eye tests more practical and cheaper. However, smartphone fundus photography has few limitations such as the adjustment of filming distance, glare from the condensing lens and SRs caused by the ceiling lights. These mobile fundus photography-related limitations are the major research challenges in the current AI-based smartphone diagnosis system development field.

Eye diseases are the most important health problem among the older people. The conventional slit-lamp eye examination is inadequate due to the ophthalmologist's memory, inconsistency in diagnosis, and record management issues. Digital pictures of an eye and CI-based diagnosis systems can be considered as assistive methods of a traditional eye examination. The various CI-based systems used for the diagnosis of anterior and retinal eye diseases were discussed in this paper. We have discussed the anterior and retinal eye imaging modalities, pre-processing methods, ROI segmentation approaches, soft computing-based eye disease diagnosis systems, multi-scale-based diagnosis systems, ML-based CI diagnosis systems, and current trends in DL and smartphone-based CI eye disease diagnosis systems. A comprehensive detail about the benchmark eye image datasets and performance evaluation criteria of the referred research works were also discussed in detail. In the comparative analysis, the morphological operation-based pre-processing method has shown an effective performance than the other pre-processing methods. In ROI segmentation of anterior eye images, CHT-based methods attained better segmentation accuracy compared with other ROI segmentation techniques. In retinal blood vessel segmentation, deep neural networks have shown an effective performance than the other segmentation methods. However, the training time and computational complexity of deep learning-based segmentation is more compared with traditional segmentation approaches. The comparative analysis results show that the ML-based CI systems showed an more effective performance than the other anterior eye disease diagnosis systems. Furthermore, the ML-based CI systems have shown an effective performance in retinal disease diagnosis also. However, the feature extraction algorithm in the ML-based diagnosis system needs to be manually programmed and varied according to the disease category. Hence, the ML-based diagnosis is not appropriate for the unknown eye disease category. The comparative analysis results also illustrate that the DL-based models show significant classification performance in anterior and retinal eye disease diagnosis. The usage of DL-based CI systems can enhance the diagnosis efficiency, when the subjects are suffering from multiple types of eye diseases.

Future research works can be based on the following ways. It is observed from the comparative analysis that many researchers used their own eye image database for the implementation and evaluation of their anterior eye disease diagnosis systems. Hence, there is a need to develop a large-scale publicly available anterior eye image database, which can be used for the development and evaluation of future anterior eye disease diagnosis systems. Moreover, the pathological symptoms of eye diseases varied according to the ethnic group of participants. Hence, the acquisition of data from multiple ethnic groups is also an essential task to make the CI systems as global systems. The usage of DL algorithms with large scale of eye image benchmarks is also a probable future research direction in the CI-based eye disease diagnosis field. Model development based on different ethnic data and DL-based models can simplify the complexities in traditional ROI-based learning for eye disease diagnosis.

This article does not contain any studies involving human participants or animals performed by any of the authors.

None.

The authors declare no competing interests.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.