Vision.seecs.edu.pk

J Med Syst (2012) 36:145–157DOI 10.1007/s10916-010-9454-7 Algorithms for the Automated Detection of DiabeticRetinopathy Using Digital Fundus Images: A Review Oliver Faust & Rajendra Acharya U. & E. Y. K. Ng &Kwan-Hoong Ng & Jasjit S. Suri Received: 23 January 2010 / Accepted: 28 February 2010 / Published online: 6 April 2010 # Springer Science+Business Media, LLC 2010 Abstract Diabetes is a chronic end organ disease that occurs therefore a generalization of individual results is difficult.
when the pancreas does not secrete enough insulin or the However, this review shows that the classification results body is unable to process it properly. Over time, diabetes improved has improved recently, and it is getting closer to affects the circulatory system, including that of the retina.
the classification capabilities of human ophthalmologists.
Diabetic retinopathy is a medical condition where the retinais damaged because fluid leaks from blood vessels into the Keywords Diabetic retinopathy . Fundus images .
retina. Ophthalmologists recognize diabetic retinopathy Automated detection . Blood vessel area . Exudes .
based on features, such as blood vessel area, exudes, Hemorrhages . Microaneurysms . Maculopathy hemorrhages, microaneurysms and texture. In this paper wereview algorithms used for the extraction of these featuresfrom digital fundus images. Furthermore, we discuss systems that use these features to classify individual fundusimages. The classifications efficiency of different DR The fast progression of diabetes is one of the main systems is discussed. Most of the reported systems are challenges of current health care. The number of people highly optimized with respect to the analyzed fundus images, afflicted with the disease continues to grow at an alarmingrate. The World Health Organization expects the number ofpeople with diabetics to increase from 130 million to 350 O. Faust (*) R. Acharya U.
million over the next 25 years ]. The situation is made Department of Electronics and Computer Engineering,Ngee Ann Polytechnic, worse by the fact that only one half of the patients are 535 Clementi Road, aware of the disease. And in the medical perspective, Singapore, Singapore 599489 diabetes leads to severe late complications. These compli- e-mail: [email protected] cations include macro and micro vascular changes which result in heart disease, renal problems and retinopathy. For School of Mechanical and Aerospace Engineering, example, studies in the United States show that diabetes is College of Engineering, Nanyang Technological University, the fifth-deadliest disease, and still there is no cure. In the 50 Nanyang Avenue, Singapore 639798 United States, the total annual economic cost of diabetes in 2002 was estimated to be $132 billion, this translates to one Department of Biomedical Imaging, University of Malaya, out of every 10 health care dollars spent [ Kuala Lumpur, Malaysia Diabetic retinopathy (DR) is a common complication of diabetes. Indeed, it is so common that it is the leading cause J. S. SuriBiomedical Technologies, of blindness in the working population of western countries The rate of diabetes is increasing, not only in developedcountries, but in underdeveloped countries as well. Unfortu- nately, most developing countries lack basic recoding of DR University of Idaho,Moscow, ID, USA cases []. It is estimated that 75% of people with diabetic J Med Syst (2012) 36:145–157 retinopathy live in developing countries []. The situation reported that type 2 diabetes is often part of a metabolic in developing countries is especially bad, because there is syndrome that includes obesity, elevated blood pressure, inadequate treatment. Regardless of the health care situation and high levels of blood lipids ].
in their country of origin, people with diabetes are 25 timesmore likely to develop blindness when compared with individuals who do not suffer from this disease DR isa silent disease, because it may only be recognized by the The recent increase in diabetes can be attributed to an aging patient when the changes in the retina have progressed to a population and increasing prevalence of obesity as well as level where treatment is complicated and nearly impossible.
sedentary life habits. Genetic inheritance plays a role in The prevalence of retinopathy varies with the age of onset of both, type 1 and type 2 diabetes. But it appears that type 1 diabetes and the duration of the disease.
diabetes is also triggered by some (mainly viral) infections.
So far, the most effective treatment for DR can be There is also a genetic element in individual susceptibility administered only in the first stages of the disease.
to some of these triggers which has been traced to particular Therefore, early detection through regular screening is of human leukocyte antigen genotypes. However, even in paramount importance. To lower the cost of such screen- those who have inherited the susceptibility, type 1 DM ings, digital image capturing technology must be used, seems to require an environmental trigger. Some evidence because this technology enables us to employ state-of-the- indicates that the B4 virus might be such a trigger.
art image processing techniques which automate thedetection of abnormalities in retinal images.
This paper reviews automated detection systems for DR.
This review is structured as follows: First we discuss the Diabetes affects the kidney, eyes, nerves and heart. In the underlying disease, i.e. diabetes, in terms of its causes and following sections, we have discussed these affects briefly.
effects on the human body. Following the goals of thispaper, we focus on the effects of diabetes on the eye. These Diabetic nephropathy effects lead to features, such as blood vessel area, exudes,hemorrhages, microaneurysms and textures []. These Diabetic nephropathy is the main cause of end-stage renal features are used for the automatic detection of DR. In the diseases. When the body digests protein it contaminates the automatic detection of DR stages section we reviewed blood with waste products. The kidneys filter out these different automated detection systems which have been waste products. A large number of small blood vessels reported in scientific literature. In the discussion section we (capillaries) are an essential component of this filter. After discussed the advantages and disadvantages of different 20-30 years, they start to leak and useful protein is lost in methods. The last section of this paper presents conclusions the urine ].
and outlines further work.
It was stated that interruption of the renin–angiotensin system slows the progression of renal diseases in patients with type 1 diabetes, but similar data are not available forpatients with type 2 ].
Diabetes mellitus (DM) is the name of a chronic, systemic,life-threatening disease. It occurs when the pancreas does Diabetic cardiomyopathy not secrete enough insulin or the body is unable to processit properly. This results in an abnormal increase in the Patients with both diabetes and ischemic heart disease seem glucose level in the blood. Over time this high level of to have an enhanced myocardial dysfunction leading to glucose causes damage to blood vessels. This damage accelerated heart failure (diabetic cardiomyopathy). Thus, affects both eyes and nervous system, as well as heart, patients with diabetes are prone to congestive heart failure kidneys and other organs [].
In general there are two types of diabetes. Diabetes type 1 results from a failure of the human body to produce Diabetic neuropathy insulin. Type 1 diabetes is less common than type 2diabetes. People with type 1 diabetes take insulin injections.
Diabetic neuropathy results in a gradual loss of nerve It is estimated that 90-95% of Americans, who are function which limits the amount of sensation on the plantar diagnosed with diabetes, have type 2 diabetes []. This aspects of the feet [This diminished sensation disables form of diabetes usually develops in adults age 40 and older individuals from being able to feel the onset or occurrence and is most common in the age group over age 55. About of a foot injury. As a result, patients with this disease are 80% of people with type 2 diabetes are overweight. It was more inclined to experience plantar ulceration [

J Med Syst (2012) 36:145–157 People with DM can develop nerve problems at any following list describes three subclasses of NPDR as well time, but the longer a person has diabetes, the greater the risk. Acharya et al. state that abnormal plantar pressures & Mild NPDR: at least one microaneurysm with or play a major role in the pathologies of neuropathic ulcers in without the presence of retinal haemorrhages, hard the diabetic foot [].
exudates, cotton wool spots or venous loops (Fig. Approximately 40% of people with diabetes have Diabetic retinopathy at least mild signs of diabetic retinopathy [ & Moderate NPDR: numerous microaneurysms and reti- Diabetes mellitus often results in diabetic retinopathy which nal haemorrhages are present. A limited amount and is caused by pathological changes of the blood vessels cotton wool spots of venous beading can also be seen which nourish the retina. DR is the main cause of new cases (Fig. 16% of the patients with moderate NPDR of blindness among adults aged 20–74 years. During the will develop PDR within 1 year [ first 20 years of the disease, nearly all patients with type 1 & Severe NPDR: is characterized by any one of the diabetes and >60% of patients with type 2 diabetes have following (4-2-1 rule) characteristics: (1) numerous retinopathy. In the Wisconsin Epidemiologic Study of DR, haemorrhages and microaneurysms in 4 quadrants of 3.6% of younger-onset patients (type 1 diabetes) and 1.6% the retina (2) venous beading in 2 or more quadrants (3) of older-onset patients (type 2 diabetes) were legally blind Intraretinal microvascular abnormalities in at least 1 ]. In the younger-onset group, 86% of blindness was quadrant (Fig. ). Severe NPDR carries a 50% attributable to DR. In the older-onset group, in which other chance of progression to PDR within 1 year [ eye diseases were common, one-third of the cases of legal & PDR: is the advanced stage; signals, sent by the retina for blindness were due to DR. Figure shows the different nourishment, trigger the growth of new blood vessels.
features of the typical DR image.
These blood vessels do not cause symptoms or vision loss.
DR occurs when the increased glucose level in the But, their walls are thin and fragile, this leads to a high risk blood damages the capillaries, which nourish the retina.
that they leak blood (Fig. ). This leaked blood As a result of this damage, the capillaries leak blood contaminates the vitreous gel and this causes severe and fluid on the retina [The visual effects of this vision loss and even blindness. About 3% of people, with leakage are features, such as microaneurysms, hemor- this condition, may experience severe visual loss ].
rhages, hard exudates, cotton wool spots or venous loops,of DR ].
Detection methods Types of diabetic retinopathy DR can be broadly classi-fied as nonproliferative DR (NPDR) and proliferative Early detection of DR is important, because treatment DR (PDR). Depending on the presence of specific DR methods can slow down the progression of the disease.
features, the stages can be identified , The Most treatment methods are based on laser technology.
Fig. 1 Different features in a

J Med Syst (2012) 36:145–157 Fig. 2 Typical fundus images: (a) normal (b) mild DR (c) moderate DR (d) severe DR (e) prolific DR Laser photocoagulation cauterizes ocular blood vessels, provides an excellent window to the health of a patient which effectively stops their leakage. The focal laser affected by DR. Figure shows an example of blood vessel treatment method reduces retinal thickening. This may detection from different types of DR The blood vessel prevent worsening of retinal swelling. To be specific, this structure was obtained by subjecting the green component treatment reduces the risk of vision loss by 50%. For a of the RGB fundus image to a number of image processing small number of cases, with total vision loss, improvement Blood vessels were detected using two-dimensional matched filters [Gray-level profile of cross section of blood vessel approximated by Gaussian shaped curve. Theconcept of matched filter detection of signals was used to Medical image analysis is a research area that currently detect piecewise linear segments of blood vessels after the attracts lots of interest from both scientists and physicians.
The objective of this field is to develop computational tools Vessel points in a cross section are found with a fuzzy C- which will assist quantification and visualization of means classifier ]. They have located and outlined blood interesting pathology and anatomical structures. These tools vessels in images by the use of a novel method to segment work with digital fundus images of the eye. The procedure blood vessels that compliments local vessel attributes with of taking fundus images starts by dilating the pupil with region-based attributes of the network structure.
pharmaceutical eye drops. After that the patient is asked to Hayashi et al. have developed a computer aided diagnosis stare at a fixation device in order to steady the eyes. While system to assist physicians in detecting abnormalities associ- taking the pictures, the patient will see a series of bright ated with fundus images of the retina []. Their proposed flashes. The entire process takes about five to ten minutes.
system can detect blood vessel intersections and it can To ensure that DR treatment is received on time, the eye identify abnormal widths in blood vessels.
fundus images of diabetic patients must be examined at Computerized system for both extraction and quantita- least once a year ].
tive description of the main vascular diagnostic signs from Feature extraction methods and analysis Image processing can do both reduce the workload ofscreeners and play a central role in quality assurance tasks.
Therefore, there has been an increase in the application ofdigital image processing techniques for automatic detectionof DR ]. For example, color features on Bayesianstatistical classifier were used to classify each pixel intolesion or non-lesion classes [ The following sections describe blood vessels, exudes, hemorrhages, microaneurysms and maculopathy detectiontechniques. These detection techniques yield most of thefeatures which are used in automated DR detection systems.
Digital fundus photography from the human eye gives clearimages of the blood vessels in the retina. This method Fig. 3 Results of blood vessel detection for normal and PDR

J Med Syst (2012) 36:145–157 fundus images in hypertensive retinopathy was presented labeled ground truth segmentation for five images and ]. The features they have taken into account are vessel achieved 84.37% sensitivity and 99.61% specificity.
tortuosity, generalized and focal vessel narrowing, presenceof Gunn or Salus signs.
A new system is proposed for the automatic extraction of the vascular structure in retinal images, based on a Exudates are accumulations of lipid and protein in the sparse tracking technique was proposed . Blood vessel retina. Typically they are bright, reflective, white or cream points in a cross section are found by means of a fuzzy c- colored lesions seen on the retina. They indicate increased means classifier. After tracking the vessels, identified vessel permeability and an associated risk of retinal edema.
segments were connected using greedy connection algo- Although, not sight threatening in themselves, they are a rithm. Finally bifurcations and crossings were identified marker of fluid accumulation in the retina. However, if they analyzing vessel end points with respect to the vessel appear close to the macula center they are considered sight threatening lesions. Most of the time they are seen together Blood vessel tracker algorithm was developed to with microaneurysms. These microaneurysms indicate determine the retinal vascular network captured using a themselves increased leakage, therefore the classical lesion digital camera [The tracker algorithm detects optic is a circular ring of exudates with several microaneurysms disk, bright lesions such as cotton wools spots, and dark at its center. Figure shows an example exudates detection lesions such as haemorrhages. This algorithm identifies from different types of DR In the result pictures, black arteries and veins with an accuracy of 78.4% and 66.5% indicates no exudates and white indicates the area where exudates were detected. An important step in the extraction Vallabha et al. have proposed a method for automated process is removing prominent structures of the retina, such detection and classification of vascular abnormalities in as blood vessel tree and optic disc. After these structures diabetic retinopathy ]. They detected vascular abnor- have been removed, the exudates were detected using a malities using scale and orientation selective Gabor filter sequence of image processing algorithms banks. The proposed method classifies retinal images as A novel approach which combines brightness adjustment either mild or severe cases based on the Gabor filter procedure with statistical classification method and local- window-based verification strategy was proposed [ The microaneurysms in retinal fluorescien angiograms Their results indicate that they were able to achieve 100% was identified by first locating the fovea by sub-sampling accuracy in terms of identifying all the retinal images with image by factor of four in each dimension ]. Subse- exudates while maintaining a 70% accuracy in correctly quently, the image was subjected to median filtering with a classifying the truly normal retinal images as normal.
5 by 5 mask to reduce high-frequency components. Then Hunter et al. have studied neural network based exudates the image was correlated with a two-dimensional circularly detection []. They introduced a hierarchical feature symmetric triangular function with modelled gross shading selection algorithm, based on sensitivity analysis to of the macula.
Blood-vessel detection algorithm based on the regional recursive hierarchical decomposition using quadtrees andpost-filtration of edges to extract blood vessels was studied]. This method was able to reduce false dismissals ofpredominately significant edges and faster in comparison tothe existing approach with reduced storage requirements forthe edge map.
Li et al. have used the arteriolar-to-venular diameter ratio of retinal blood vessels as an indicator of disease relatedchanges in the retinal blood vessel tree [Theirexperimental results indicate a 97.1% success rate in theidentification of vessel starting points, and a 99.2% successrate in the tracking of retinal vessels.
A new method of texture based vessel segmentation to overcome this problem was proposed ]. The Fuzzy C-Means (FCM) clustering algorithm was used to classify thefeature vectors into vessel or non-vessel based on thetexture properties. They compared their method with hand- Fig. 4 Results of exudates detection for normal, PDR

J Med Syst (2012) 36:145–157 distinguish the most relevant features. The final architecture be useful for detecting clinically important bright lesions, achieved 91% lesion-based performance using a relatively enhancing early diagnosis, and reducing visual loss in small number of images.
patients with diabetes.
A new approach to automatically extract the main A set of optimally adjusted morphological operators features in color fundus images was proposed ]. Optic were used for the detection of exudate in diabetic disk was localized by the principal component analysis retinopathy patients' non-dilated pupil and low-contrast (PCA) and its shape was detected by a modified active images ]. They used these operators to design an exudes shape model (ASM). Exudates were extracted by the detection system. This system achieved sensitivity and combined region growing and edge detection. Their results specificity of 80% and 99.5%, respectively.
show 99%, 94%, and 100% for disk localization, diskboundary detection, and fovea localization respectively.
Microaneurysms detection The sensitivity and specificity for exudate detection were100% and 71%.
Microaneurysms detection is very important, because these Osareh et al. have presented results for fundus image structures constitute the earliest recognizable feature of DR.
based exudes classification Their method evaluated The first reports which link these structures to DR date different learning algorithms, such as neural network and back to 1879 []. More recently, Jalli et al. have analyzed support vector machine. The neural network based approach the appearance and disappearance of microaneurysms in performs marginally better than the support vector machine different phases of fluorescein angiography [. In a based approach, the latter is more flexible given boundary similar study both formation rate and disappearance of conditions such as control of sensitivity and specificity rates.
microaneurysms in early DR were analyzed []. The The neural network results were: accuracy = 93.4%, sensitiv- microaneurysms turnover were computed reliabibly from ity = 93.0%, specificity = 94.1%.
color fundus images They used a new method called Exudates are found using their high grey level variation, MA-tracker to count microaneurysms. They showed that and their contours were determined by means of morpho- the microaneurysms remain stable over time, but only 29% logical reconstruction techniques ]. The optic disc was remain at the same place.
detected by means of morphological filtering techniques Figure shows the results of microaneurysms detection and the watershed transformation. Their results show a for normal and PDR []. In example the green component, mean sensitivity of 92.8% and a mean predictive value of of the RGB fundus image, was chosen to obtain the microaneurysms. Similar to the exudates detection algo- Local contrast enhancement fuzzy C-means and support rithm, first the prominent structures within retina images, vector machine was used to detect and classify bright such as blood vessel tree and optic disc are to be removed.
lesions ]. Their classification results are as follows: After that a sophisticated sequence of image processing algorithms was used to determine the areas within the Classification between bright lesions and bright non- fundus images to get microaneurysms [].
lesion: sensitivity = 97%, specificity = 96%.
& Classification between exudates and cotton wool spots: sensitivity = 88%, specificity = 84%.
Fuzzy C-means clustering and morphological reconstruc-tion was used to detect exudates detection on low-contrastimages taken from non-dilated pupils ]. The sensitivityand specificity for the exudates detection are 86% and 99%respectively.
Flemming et al. have used multi-scale morphological algorithms to obtain what they call candidate exudates ].
The final classification was done by determining thebackground (drusen) of the candidates. Exudates weredetected with sensitivity 95.0% and specificity 84.6% in atest set of 13219 images of which 300 contained exudates.
Automated system capable of detecting exudates and cotton-wool spots and differentiating them from drusen incolor images obtained in community based diabetic patientshas been developed ]. The machine learning can befurther improved with additional training data sets, and can Fig. 5 Results of microaneurysms detection for normal, PDR

J Med Syst (2012) 36:145–157 The automated identification of diabetic retinopathy exudates detection were 88.5% and 99.7%, respectively and based on the presence of microaneurysms was studied algorithm achieved a sensitivity of 77.5% and specificity of ]. The optometrists achieved 97 per cent sensitivity at 88 88.7% for detection of HMA.
per cent specificity and the automated retinopathy detector Larsen et al. have used image processing for the achieved 85 per cent sensitivity at 90 per cent specificity.
detection of both hemorrhages and microaneurysms [Their algorithm demonstrated a specificity of 71.4% and a sensitivity of 96.7%.
The robust detection of red lesions in digital color As the degree of DR advances retinal hemorrhages become fundus photographs is a critical step in the development of evident. They indicate an increased ischemia (loss of oxygen) automated screening systems for diabetic retinopathy [ retina. As their numbers increase the retinal vessels become Their method achieved a sensitivity of 100% at a specificity more damaged and leaky this leads to exudation of fluid, lipid of 87% in detecting the red lesions.
and proteins. Figure shows the result of hemorrhage Bottom-up and top-down strategies were applied to cope detection ]. The white patches indicates the hemorrhages with difficulties in lesions detection, such as inhomoge- in the image. There are two parts in haemorrhages detection: neous illumination [After the application of appropriatestrategy, they used local contrast enhancement, fuzzy C- i) Detection of blood vessels; means and hierarchical support vector machine to classify ii) Detection of blood vessels with haemorrhages.
bright non-lesion areas, exudates and cotton wool spots.
The image with blood vessel alone was subtracted fromimage with blood vessel and haemorrhages to get the image Distance of exudates from fovea with haemorrhages [].
Ege et al. have developed a tool which provides In diabetic maculopathy, fluid rich in fat and cholesterol, automatic analysis of digital fundus images ]. In their leaks out of damaged blood vessels. If fluid and cholesterol study, a Bayesian, a Mahalanobis, and a k nearest neighbor accumulates near the center of the retina (the macula) it can classifier were used on 134 retinal images. The Mahalano- cause distortion and permanent loss of central vision. There bis classifier showed the best results: microaneurysms, are the two types of maculopathy eye disease: haemorrhages, exudates, and cotton wool spots were & Non-Clinically Significant Macular Edema (NCSME).
detected with a sensitivity of 69%, 83%, 99%, and 80%, Figure shows a fundus image of NCSME, in this stage the patient will not realize that he is affected, Fully automated computer algorithms were able to detect because there are no visible symptoms. Exudates start hard exudates and haemorrhages and microaneurysms to leak, and the retina becomes boggy like a sponge.
(HMA) using of a new technique, termed a 'Moat Operator', But, the patient's vision is not seriously affected, was proposed , The sensitivity and specificity for because the locations of the exudates are far awayfrom the fovea.
& In the Clinically Significant Macular Edema (CSME) stage, most of the retinal blood vessels are damaged andthe leakage area becomes bigger. The exudates leak outand this liquid concentrates very close to the fovea. Thevisibility is greatly affected, because the detected imagecannot be focused on the macula properly. Figure shows a CSME fundus image.
Fig. 6 Results of hemorrhages detection for normal, PDR [] Fig. 7 Fundus images: (a) NCSME (b) CSME J Med Syst (2012) 36:145–157 Philips et al. have studied diabetic maculopathy and detection aneurysms, hard exudates, and cotton wool spots, was of exudates on fundus images [, ].
studied []. The method was able to identify the NPDR Nayak et al. have present a computer- based system for stage correctly with an accuracy of 81.7%.
the identification of CSME, non-CSME and normal fundus Exudates, haemorrhages, and microaneurysms were used eye images ]. Features are extracted from raw fundus for screening of DR subjects [The sensitivity and images which are then fed to an artificial neural network specificity of their software was 74.8% and 82.7%, classifier. They demonstrated a sensitivity of more than respectively in differentiating DR and normal subjects 95% for these classifiers with a specificity of 100%.
Early detection of DR (presence of microaneurysms) was proposed based on decision support system by Kahai etal. []. Bayes optimality criteria was used to detect Texture is a measure of properties, such as smoothness, microaneurysms. Their method was able to identify the coarseness, and regularity of pixels, in an image ]. One early stage of DR with a sensitivity of 100% and specificity way to define texture is: a mutual relationship among intensity values of neighboring pixels repeated over an area Normal, mild, moderate, severe and prolific DR stages larger than the size of the relationship Conventional were automatically classified using both area and perimeter texture recognition systems can be grouped into three of the RGB components of the blood vessels together with classes: structural, statistical and spectral , Textures a feedforward neural network []. System average can be defied using statistical approaches, this yields classification efficiency was 84% and sensitivity, specificity characterizations such as smooth, coarse, grainy and so were 90% and 100% respectively. Nayak et al. have used on. Statistical algorithms are based on the relationship exudates and blood vessel area along with texture param- between intensity values of pixels; measures include eters coupled with neural network to classify fundus images entropy, contrast, and correlation based on the gray level into normal, NPDR and PDR [. They obtained a co-occurrence matrix ]. Different texture parameters can detection accuracy of 93%, sensitivity and specificity of be used for the detection of DR stages [ 90% and 100% respectively. Recently, bispectral invariantfeatures were used as features for the support vectormachine classifier to classify the fundus image in to Automatic detection of DR stages normal, mild, moderate, severe and prolific DR classes byAcharya et al. []. They have demonstrated an average Over the last two decades there was a rapid development of accuracy of 82% and sensitivity, specificity of 82% and Computer-aided diagnosis (CAD) ]. The idea of using 88% respectively. Normal, mild, moderate, severe and computers to help in medical image diagnosis is in more prolific classes of DR were classified automatically based practice. However, the quality of these CAD systems on haemorrhages, microaneurysms, exudates and blood increased with more accurate sensor data, more processing vessel areas with a support vector machine classifier power and better understanding of the underlying disease.
The system was able to identify the unknown class Recently, Lee et al. have concluded that the performance of accurately with an efficiency of more than 85% and a their computer vision system in diagnosing early retinal sensitivity of more than 82% and a specificity of 86%.
lesions is comparable with that of human experts [In Nicolai et al. have designed an automated lesion system, the next section we have reviewed different classification which identified 90.1% of patients with DR and 81.3% of patients without DR, when applied in a screening popula-tion comprising of patients with untreated DR [The Classification methods automated system demonstrated a sensitivity of 93.1% anda specificity of 71.6%.
Colour features were used on Bayesian statistical classifier Usher et al. have designed a support system for DR classify each pixel into lesion or non-lesion classes ].
screenings ]. Their system showed a maximum sensi- They have achieved 100% accuracy in identifying all the tivity for the detection of any retinopathy on a per patient retinal images with exudates, and 70% accuracy in basis of 95.1%, accompanied by a specificity of 46.3%. The classifying normal retinal images as normal.
specificity could be increased as far as 78.9%, but this DR and normal retina were classified automatically increase was accompanied by a fall in sensitivity to 70.8%.
using image processing and multilayer perceptron neural At a setting with 94.8% sensitivity and 52.8% specificity, network ] The system yielded a sensitivity of 80.21% no cases of sight threatening retinopathy were missed.
and a specificity of 70.66%. Automated diagnosis of Neubauer et al. have investigated both photography and NPDR, based on three lesions: hemorrhages and micro- optic disc topography mode of the retinal thickness J Med Syst (2012) 36:145–157 analyzer [The system yielded a mean 93% sensitivity Li et al. have proposed a method for screening DR and for PDR together with 100% specificity for DR cases.
distinguishing PDR from NPDR automatically using color A software to grade the severity of 3 types of early retinal images [Their method showed a sensitivity lesions, hemorrhages and microaneurysms, hard exudates 80.5%, positive predictive value 90.8%, true positive ratio and cotton wool spots of DR was proposed to classify 95.8%, false positive ratio 16.7% in detecting PDR and NPDR [They were able to identify 82.6%, 82.6%, and NPDR accurately.
88.3% using the 430 images and 85.3%, 87.5%, and 93.1% Abràmoff et al. have evaluated the performance of a using the 361 images, respectively, for hemorrhages and system for automated detection of DR in digital retinal microaneurysms, hard exudates, and cotton wool spots.
fundus images []. The system was constructed entirely Philip et al. have assessed the efficiency of automated from published algorithms and it was tested in a large, "disease/no disease" grading for DR within a systematic representative, screening population. They achieved a screening programme [Detection of retinopathy was sensitivity of 84% and a specificity of 64%.
achieved by automated grading with 90.5% sensitivity and Higher order spectra features were used as input to a 67.4% specificity.
support vector machine classifier in order to classify fundus A system, designed by Estabridis et al., has detected images into normal, mild DR, moderate DR, severe DR and features such as fovea, blood vessel network, optic disk, PDR classes with an accuracy of 82% bright and dark lesions, which are associated with DR Vujosevic et al. have determined single lesions to grade successfully ]. It has achieved a classification accuracy clinical levels of DR and diabetic macular edema using both 1 and 3 nonmydriatic digital color retinal images [ Table 1 Comparison of different classification methods Minimum distance discriminant classifier Sinthanayothin et Usher et al. 2003 Singalavanija et al.
Blood vessels, exudates, haemorrhages, microaneurysms Lee et al. 2005 [] Hemorrhages, microaneurysms, hard exudates, cotton wool spots Neubauer et al.
Retinal thickness analyzer Kahai et al. 2006 Decision support system (DSS) Philip et al. 2007 Fovea, blood vessel network, optic disk, bright and dark lesions Bright lesions, retinal vessel patterns Abràmoff et al.
Optic disc, retinal vessels, hemorrhages, microaneurysms, vascular, abnormalities, exudates, cotton wool spots, drusen Area of blood vessel Nayak et al. 2008 Blood vessels, exudates and texture Acharya et al. 2008 Higher order spectra Acharya et al. 2009 Blood vessel, exudates, microaneurysms, haemorrhages Vujosevic et al.
J Med Syst (2012) 36:145–157 Sensitivity and specificity for detecting referable levels of different texture parameters along with other DR features DR were 82% and 92%, respectively.
can improve the classification efficiency.
Table summarizes the results of the 15 automated DR The accurate detection of macula, optic disc, micro- classification systems. The table entries are chronologically aneurysm and haemorrhages is challenging. But, we feel ordered and the percentage values for accuracy of classifi- that, with recent advances in the medical imaging and data cation, sensitivity and specificity are rounded to the nearest mining techniques as well as novel algorithms for the detection of these features, it may be possible.
Also, we feel that, the early detection of the DR (mild DR) by detecting the microaneurysm can save the progres- sion of the disease and hence can save the loss of visionand improve the quality of life.
The prolonged diabetes leads to the formation of micro-aneurysms and subsequently it leads to exudates as well ashaemorrhages. These are the features of DR and they may lead to severe vision loss or even blindness. In order toavoid these complications, it is very important to detect DR Prolonged diabetes leads to DR, where the retina is early. This can be done by an accurate detection of damaged due to fluid leaking from the blood vessels.
Usually, the stage of DR is judged based on blood vessels, It is very difficult to detect the exudates clearly, because exudes, hemorrhages, microaneurysms and texture. In this they are tiny spots on the retina. Also, the detection of paper, we have discussed different methods for features haemorrhages is very challenging. The texture of haemor- extraction and automatic DR stage detection. An ophthal- rhages and macula is almost the same. So, we need to have mologist uses an ophthalmoscope to visualize the blood robust algorithms which detect these features.
vessels and his or her brain to detect the DR stages.
In the previous section we reviewed and compared 15 Recently digital imaging became available as a tool for DR automated DR detection systems. The results were obtained screening. It provides high quality permanent records of the by optimizing the algorithms for a specific set of fundus retinal appearance, which can be used for monitoring of progression or response to treatment, and which can be In the earlier part of the research, authors have classified reviewed by an ophthalmologist, digital images have the into two classes using fundus images based on two or three potential to be processed by automatic analysis systems. A features. Then subsequently, more features were introduced combination of both accurate and early diagnosis as well as to improve the classification efficiency. Also, the classifi- correct application of treatment can prevent blindness cation efficiency was improved further by using non-linear caused by DR in more than 50% of all cases. Therefore, methods like higher order spectra The algorithms regular screenings for DR of patients with diabetes is involving four features namely, area of blood vessel, important. The grading of the resultant fundus images is an exudates, haemorrhages and microaneurysms coupled with important cost factor. Automated DR detection can reduce support vector machine were used to classify fundus images the grading cost and thereby make the whole screening into five classes (normal, mild DR, moderate DR, severe process less expensive. Some of the algorithms and systems DR and prolific DR) with an efficiency of 86%, sensitivity reviewed in this paper are close to achieve DR identifica- and specificity of 82% and 86% respectively [].
tion in clinical practice.
Most of algorithms, discussed in the earlier section, have used only a few features like blood vessels, haemorrhages,exudates and microaneurysms etc. We predict that an algorithm involving all features namely, blood vessels,exudates, haemorrhages, microaneurysms, distance between 1. Aboderin, I., Kalache, A., Ben-Shlomo, Y., Lynch, J. W., Yajnik, exudates and macula, and texture parameters will be more C. S., Kuh, D., and Yach, D., Life course perspective on coronary robust. However, for this forecast to hold it is of paramount heart disease: key issues and implications for policy and research.
World Health Organization, Geneva, 2002.
importance that the individual parameter extraction algo- 2. Abràmoff, D. M., Niemeijer, M., Suttorp-Schulten, S. A. M., rithms are also as robust as possible.
Viergever, A. M., Russell, R. S., and van Ginneken, B., The design of good classifiers will increase the auto- Evaluation of a system for automatic detection of diabetic matic detection rate. Huge diverse training data will retinopathy from color fundus photographs in a large popula-tion of patients with diabetes. Diabetes Care 31(2):193–198, significantly improve the classification efficiency. Also, fundus images taken under uniform good lighting con- 3. Acharya, U. R., Chua, K. C., Ng, E. Y. K., Wei, W., and Chee, C., ditions will improve the detection of DR. Furthermore, Application of higher order spectra for the identification of J Med Syst (2012) 36:145–157 diabetes retinopathy stages. J. Med. Syst., USA 32(6):431–488, diabetic retinopathy screening. Phys. Med. Biol. 52(24):7385– 4. Acharya, U. R., Lim, C. M., Ng, E. Y. K., Chee, C., and Tamura, 22. Fong, D. S., Aiello, L., Gardner, T. W., King, G. L., Blankenship, T., Computer based detection of diabetes retinopathy stages using G., Cavallerano, J. D., Ferris, F. L., and Klein, R., Diabetic digital fundus images. J. Eng. Med. 223(H5):545–553, 2009.
retinopathy. Diabetes Care 26(1):226–229, 2003.
5. Acharya, U. R., Lim, C. M., Ng, E. Y. K., Chee, C., and Tamura, 23. Forracchia, M., Grisan, M. E., and Ruggeri, A., Extraction and T., Computer-based detection of diabetes retinopathy stages using quantitative description of vessel features in hypertensive retinop- digital fundus images. Proc Inst Mech Eng H. 223(5):545–553.
athy fundus images, Presented at CAFIA2001, 2001.
6. Acharya, U. R., Ng, E. Y. K., and Suri, J. S., Image modelling of 24. Frank, R. N., Diabetic retinopathy. Prog. Retin. Eye Res. 14 human eye. Artech House, MA, 2008.
7. Acharya, U. R., Tan, P. H., Subramaniam, T., Tamura, T., Chua, 25. Fujita, H., Uchiyama, Y., Nakagawa, T., Fukuoka, D., Hatanaka, K. C., Goh, S. C., Lim, C. M., Goh, S. Y., Chung, K. R., and Law, Y., Hara, T., Lee, G. N., Hayashi, Y., Ikedo, Y., Gao, X., and C., Automated identification of diabetic type 2 subjects with and Zhou, X., Computer-aided diagnosis: the emerging of three CAD without neuropathy using wavelet transform on pedobarograph. J.
systems induced by Japanese health care needs. Comput. Methods Med. Syst. 32(1):21–29, 2008.
Programs Biomed. 92(3):238–248, 2008.
8. Alberti, K. G., and Zimmet, P. Z., Definition, diagnosis and 26. Galloway, M. M., Texture classification using gray level run classification of diabetes mellitus and its complications, part 1: length. Comput. Graph. Image Process. 4:172–179, 1975.
diagnosis and classification of diabetes mellitus provisional report 27. Gonzalez, R. C., and Woods, R. E., Digital image processing, 2nd of a WHO consultation. Diabet. Med. 15(7):539–553, 1998.
edition. Prentice Hall, New Jersey, 2001.
9. Bernardes, R., Nunes, S., Pereira, I., Torrent, T., Rosa, A., Coelho, 28. Grisan, I. E., Pesce, A., Giani, A., Foracchia, M., and Ruggeri, A., D., and Cunha-Vaz, J., Computer-assisted microaneurysm turn- A new tracking system for the robust extraction of retinal vessel over in the early stages of diabetic retinopathy. Ophthalmologica structure, 26th Annual International Conference of the IEEE EMBS San Francisco, USA, pp. 1620-1623, 2004.
10. Bhuiyan, A., Nath, B., Chua, J., and Kotagiri, R., Blood vessel 29. Hayashi, J., Kunieda, T., Cole, J., Soga, R., Hatanaka, Y., Lu, M., segmentation from color retinal images using unsupervised texture Hara, T., and Fujita, F., A development of computer-aided classification. IEEE Int. Conf. Image Processing, ICIP 5:521–524, diagnosis system using fundus images, Proceeding of the 7th International Conference on Virtual Systems and MultiMedia 11. Microaneurysms in diabetic retinopathy. Br. Med. J. 3(5774):548– (VSMM 2001), pp. 429-438, 2001.
30. Hellstedt, T., and Immonen, I., Disappearance and formation rates 12. Brenner, M. B., Cooper, E. M., de Zeeuw, D., Keane, F. W., of microaneurysms in early diabetic retinopathy. Br. J. Ophthal- Mitch, E. W., Parving, H. H., Remuzzi, G., Snapinn, M. S., mol. 80(2):135–139, 1996.
Zhang, Z., and Shahinfar, S., Effects of Losartan on renal and 31. Hoover, A. D., Kouzanetsova, V., and Goldbaum, M., Locating cardiovascular outcomes in patients with type 2 diabetes and blood vessels in retinal images by piecewise threshold probing of nephropathy. NEJM 345(12):861–869, 2001.
a matched filter response. IEEE Trans. Med. Imag. 19(3):203– 13. Chaudhuri, S., Chatterjee, S., Katz, N., Nelson, M., and Goldbaum, M., Detection of blood vessels in retinal images 32. Hunter, A., Lowell, J., Owens, J., and Kennedy, L, Quantification using two-dimensional matched filters. IEEE Trans. Med. Imag.
of diabetic retinopathy using neural networks and sensitivity analysis, In Proceedings of Artificial Neural Networks in Medicine 14. Cigna healthcare coverage position- A Report, 2007. Retrieved from: and Biology, pp. 81-86, 2000.
33. International Council of Ophthalmology. International standards: international clinical diabetic retinopathy disease severity scale, Last accessed on 5th December 2007.
detailed table. Retrived from: 15. Cree, J. M., Leandro, J. J. G., Soares, J. V. B., Cesar, R. M. Jr., Last accessed on 17th January 2009.
Jelinek, H. F., and Cornforth, D., Comparison of various methods 34. Jalli, P. Y., Hellstedt, T. J., and Immonen, I. J., Early versus late to delineate blood vessels in retinal images, Proceedings of the staining of microaneurysms in fluorescein angiography. Retina 17 16th Australian Institute of Physics Congress, Canberra, 2005.
16. Diabetic Retinopathy. Retrieved from: 35. Jelinek, H. J., Cree, M. J., Worsley, D., Luckie, A., and Nixon, P., Last accessed on 17th January 2009.
An automated microaneurysm detector as a tool for identification 17. Early Treatment Diabetic Retinopathy Study Research Group, of diabetic retinopathy in rural optometric practice. Clin. Exp.
Grading diabetic retinopathy from stereoscopic color fundus photo- Optom. 89(5):299–305, 2006.
graphs: an extension of the modified Airlie House classification, 36. Kahai, P., Namuduri, K. R., and Thompson, H., A decision ETDRS report number 10. Ophthalmology 98:786–806, 1991.
support framework for automated screening of diabetic retinopa- 18. Ege, B. M., Hejlesen, O. K., Larsen, O. V., Møller, K., thy. Int. J. Biomed. Imag. 2006:1–8, 2006.
Jennings, B., Kerr, D., and Cavan, D. A., Screening for 37. Kandiraju, N., Dua, S., and Thompson, H. W., Design and diabetic retinopathy using computer based image analysis and implementation of a unique blood vessel detection algorithm statistical classification. Comput. Methods Programs Biomed.
towards early diagnosis of diabetic retinopathy. Proceedings of the International Conference on Information Technology: Cod- 19. Englmeier, K. H., Schmid, K., Hildebrand, C., Bichler, S., Porta, ing and Computing (ITCC'05) IEEE Computer Society, pp. 26- M., Maurino, M., and Bek, T., Early detection of diabetes retinopathy by new algorithms for automatic recognition of 38. Klein, R., Klein, B. E. K., Moss, S. E., Davis, M. D., and DeMets, D.
vascular changes. Eur. J. Med. Res. 9(10):473–488, 2004.
L., The Wisconsin Epidemiologic Study of Diabetic Retinopathy III, 20. Estabridis K, de Figueiredo RJP, Automatic detection and prevalence and risk of diabetic retinopathy when age at diagnosis is diagnosis of diabetic retinopathy. IEEE Int. Conf. Image Process- 30 or more years. Arch. Ophthalmol. 102(4):527–532, 1984.
ing, ICIP 2007.
39. Kulakarni, D. A., Artificial neural networks for image under- 21. Fleming, D. A., Philip, S., Goatman, A. K., Williams, J. G., standing. Van Nostrand Reinhold, New York, 1993. ISBN:0-442- Olson, A. J., and Sharp, F. P., Automated detection of exudates for J Med Syst (2012) 36:145–157 40. Kumar, A., Diabetic blindness in India: the emerging scenario.
59. Phillips, R., Spencer, T., Ross, P., Sharp, P., and Forrester, J., Indian J. Ophthalmol. 46(2):65–66, 1998.
Quantification of diabetic maculopathy by digital imaging of the 41. Larsen, M., Godt, J., Larsen, N., Lund-Andersen, H., Sjolie, A.
fundus. Eye 5(1):130–137, 1991.
K., Agardh, E., Kalm, H., Grunkin, M., and Owens, D. R., 60. Ramana, K. V., and Ramamoorthy, B., Statistical methods to Automated detection of fundus photographic red lesions in compare the texture features of machined surfaces. Pattern diabetic retinopathy. Invest. Ophthalmol. Vis. Sci. 44(2):761– Recogn. 29:1447–1459, 1996.
61. Reaven, G. M., Role of insulin resistance in human disease.
42. Lee, S. C., Lee, E. T., Kingsley, R. M., Wang, Y., Russell, D., Diabetes 37:1595–1607, 1988.
Klein, R., and Warn, A., Comparison of diagnosis of early retinal 62. Scott, M., Grundy, C., Benjamin, I. J., Burke, G. L., Chait, A., lesions of diabetic retinopathy between a computer system and Eckel, R. H., Howard, B. V., Mitch, W., Smith, S. C., and Sowers, human experts. Arch. Ophthalmol. 119(4):509–515, 2001.
J. R., Diabetes and cardiovascular disease. A statement for 43. Lee, S. C., Lee, E. T., Wang, Y., Klein, R., Kingsley, R. M., and Healthcare Professionals From the American Heart Association.
Warn, A., Computer classification of nonproliferative diabetic Circulation 100:1134–1146, 1999.
retinopathy. Arch. Ophthalmol. 123(6):759–764, 2005.
63. Screening for Diabetic Retinopathy in Europe 15 years after the St.
44. Li, H., and Chutatape, O., Fundus image feature extraction.
Vincent Declaration. The Liverpool Declaration 2005. Retrieved Proceedings 22nd Annual EMBS International Conference, Chicago, pp. 3071-3073, 2000.
Last accessed on 20th December 2007.
45. Li, H., Hsu, W., Lee, M. L., and Wong, T. Y., Automated grading 64. Shahidi, M., Ogura, Y., Blair, N. P., and Zeimer, R., Retinal of retinal vessel caliber. IEEE Trans. Biomed. Eng. 52:1352–1355, thickness change after focal laser treatment of diabetic macular oedema. Br J Ophthalmol. 78(11):827–830, 1994.
46. Li, Q., Jin, X.-M., Gao, Q., You, J., and Bhattacharya, P., 65. Sinthanayothin, C., Boyce, J. F., Williamson, T. H., and Cook, H.
Screening diabetic retinopathy through color retinal images.
L., Automated detection of diabetic retinopathy on digital fundus Medical Biometrics 4901:176–183, 2008.
image. Diabet. Med. 19(2):105–112, 2002.
47. Mirmehdi, M., Xian, X., and Suri, J. S., Hand book of texture 66. Sinthanayothin, C., Kongbunkiat, V., Phoojaruenchanachai, S., analysis. Imperial College Press, UK, 2008.
and Singalavanija, A., Automated screening system for diabetic 48. Nayak, J., Bhat, P. S., Acharya, U. R., Lim, C. M., and Kagathi, retinopathy, 3rd international Symposium on Image and Signal M., Automated identification of different stages of diabetic Processing and Analysis 44(2):767-771, 2003.
retinopathy using digital fundus images. J. Med. Syst., USA, 32 67. Sopharak, A., and Uyyanonvara, B., Automatic exudates detection from diabetic retinopathy retinal image using fuzzy C-means and 49. Nayak, J., Bhat, P. S., and Acharya, U. R., Automatic identifica- morphological methods, Proceedings of the third IASTED tion of diabetic maculopathy stages using fundus images. J. Med.
international conference Advances in Computer Science and Eng. Technol. 33(2):119–129, 2009.
Technology, Thailand, pp. 359-364, 2007.
50. Neubauer, A. S., Chryssafis, C., Thiel, M., Priglinger, S., Welge- 68. Sopharak, A., Uyyanonvara, B., Barman, S., and Williamson, H.
Lussen, U., and Kampik, A., Screening for diabetic retinopathy T., Automatic detection of diabetic retinopathy exudates from and optic disc topography with the retinal thickness analyzer.
non-dilated retinal images using mathematical morphology meth- Ophthalmologe 102(3):251–258, 2005.
ods. Comput. Med. Imaging Graph. 32(8):720–727, 2008.
51. Nicolai, L., Jannik, G., Michael, G., Henrik, L. A., and Michael, 69. Tan, J. H., Ng E. Y. K., and Acharya, U. R., Study of normal L., Automated detection of diabetic retinopathy in a fundus ocular thermogram using textural parameters. Infrared Phys.
photographic screening population. Invest. Ophthalmol. Vis. Sci.
Technol. 53(2):120–126, 2009.
70. Vallabha, D., Dorairaj, R., Namuduri, K., and Thompson, H., 52. Niemeijer, M., van Ginneken, B., Russell, R. S., Suttorp-Schulten, Automated detection and classification of vascular abnormalities S. A. M., and Abramoff, D. M., Automated detection and in diabetic retinopathy, Proceedings of 13th IEEE Signals, differentiation of drusen, exudates, and cotton-wool spots in Systems and Computers 2:1625-1629, 2004.
digital color fundus photographs for diabetic retinopathy diagno- 71. Vujosevic, S., Benetti, E., Massignan, F., Pilotto, E., Varano, M., sis. Invest. Ophthalmol. Vis. Sci. 48(5):2260–2267, 2007.
Cavarzeran, F., Avogaro, A., and Midena, E., Screening for diabetic 53. Niemeijer, M., van Ginneken, B., Staal, J., Suttorp-Schulten, M., and retinopathy: 1 and 3 nonmydriatic 45-degree digital fundus photo- Abramoff, M., Automatic detection of red lesions in digital color graphs vs 7 standard early treatment diabetic retinopathy study fields.
fundus photographs. IEEE Trans. Med. Imag. 24(5):584–592, 2005.
Am. J. Ophthalmol. 148(1):111–118, 2009.
54. Ong, G. L., Ripley, L. G., Newsom, R. S., Cooper, M., and 72. Walter, T., Massin, P., Erginay, A., Ordonez, R., Jeulin, C., and Casswell, A. G., Screening for sight-threatening diabetic retinop- Klein, J. C., Automatic detection of microaneurysms in color athy: comparison of fundus photography with automated color fundus images. Med. Image Anal. 11(6):555–566, 2007.
contrast threshold test. Am. J. Ophthalmol. 137(3):445–452, 2004.
73. Wang, H., Hsu, W., Goh, K. G., and Lee, M., An effective 55. Orbis. Retrieved from: Last accessed approach to detect lesions in colour retinal images, In Proceedings December 2009.
of the IEEE Conference on Computer Vision and Pattern 56. Osareh, A., Mirmehdi, M., Thomas, B., and Markham, R., Compar- Recognition, 181-187, 2000.
ative exudate classification using support vector machines and neural 74. Watkins, J. P., ABC of diabetes retinopathy. British Medical networks, The 5th International Conf. on Medical Image Computing Journal 326:924–926, 2003.
and Computer-Assisted Intervention, pp. 413-420, 2002.
75. Wong, L. Y., Acharya, U. R., Venkatesh, Y. V., Chee, C., Lim, C.
57. Philip, S., Fleming, A. D., Goatman, K. A., Fonseca, S., M., and Ng, E. Y. K., Identification of different stages of diabetic Mcnamee, P., Scotland, G. S., Prescott, G. J., Sharp, P. F., and retinopathy using retinal optical images. Information Sciences 178 Olson, J. A., The efficacy of automated "disease/no disease" grading for diabetic retinopathy in a systematic screening 76. World Diabetes, A newsletter from the World Health Organiza- programme. Br. J. Ophthalmol. 91(11):1512–1517, 2007.
tion, 4, 1998.
58. Phillips, R., Forrester, J., and Sharp, P., Automated detection and 77. Zhang, X., and Chutatape, O., Detection and classification of quantification of retinal exudates. Graefes Arch. Clin. Exp.
bright lesions in colour fundus images. Int. Conf. on Image Ophthalmol. 231(2):90–94, 1993.
Processing 1:139–142, 2004.
J Med Syst (2012) 36:145–157 78. Zhang, X., and Chutatape, O., Top-down and bottom-up strategies 81. Singalavanija, A., Supokavej, J., Bamroongsuk, P., Sinthanayothin, in lesion detection of background diabetic retinopathy. IEEE C., Phoojaruenchanachai, S., and Kongbunkiat, V., Feasibility Computer Society Conference on Computer Vision and Pattern study on computer-aided screening for diabetic retinopathy. Jpn. J.
Recognition 2:422–428, 2005.
Ophthalmol. 50(4):361–366, 2006.
79. Parving, H. H., Brenner, B. M., Cooper, M. E., de Zeeuw, D., 82. Usher, D., Dumskyj, M., Himaga, M., Williamson, T. H., Nussey, Keane, W. F., Mitch, W. E., Remuzzi, G., Snapinn, S. M., Zhang, S., and Boyce, J., Automated detection of diabetic retinopathy in Z., and Shahinfar, S., Effect of losartan on renal and cardiovascular digital retinal images: a tool for diabetic retinopathy screening.
complications of patients with type 2 diabetes and nephropathy.
Diabet. Med. 21(1):84–90, 2004.
Ugeskr. Laeger 163(40):5514–5519, 2001.
83. The American Orthopaedic Foot and Ankle Society, 1999 web 80. Samuel, C. L., Elisa, T. L., Yiming, W., Ronald, K., Ronald, M.
page: (Last accessed 21.01.2010).
K., and Ann, W., Computer classification of a nonproliferative 84. Acharya, U. R., Ng, E. Y. K., and Suri, J. S., Image modeling of diabetic retinopathy. Arch. Ophthalmol. 123:759–764, 2005.
human eye. Artech House, MA, 2008.

Source: http://vision.seecs.edu.pk/wp-content/uploads/10_springer_Algorithms-for-the-Automated-Detection-of-Diabetic-Retinopathy-Using-Digital-Fundus-Images-A-Review.pdf