Algorithms for red lesion detection in Diabetic Retinopathy: A review

Biomedicine & Pharmacotherapy 107 (2018) 681–688

Contents lists available at ScienceDirect

Biomedicine & Pharmacotherapy journal homepage: www.elsevier.com/locate/biopha

Review

Algorithms for red lesion detection in Diabetic Retinopathy: A review R.S. Biyani, B.M. Patre

⁎

T

Department of Instrumentation Engineering, Shri Guru Gobind Singhji Institute of Engineering and Technology, Nanded 431606, India

A R T I C LE I N FO

A B S T R A C T

Keywords: Diabetic Retinopathy Microaneurysms Haemorrhages Computer aided diagnosis

Diabetic Retinopathy (DR) is an outcome of prolonged diabetes which directly or indirectly affect the human vision. DR is asymptomatic in its early stages and the late diagnosis lead to undeviating loss of vision. The computer aided diagnosis with the assistance of medical images helps in timely and accurate treatment. Microaneurysms (MA) mark the onset of DR, thus a vital point in screening of this disease. This review discuses various state of the art methods available till date for automated computer aided analysis of microaneurysms and haemorrhages. The paper also highlights qualitative and quantitative comparison of the existing literature with limitations for analysis of microaneurysms and haemorrhages. It is an attempt to systematize the available algorithms for an easy gathering and guidance to researchers working in this domain for future research.

1. Introduction Diabetes is a chronic disease which is the outcome of insufficient insulin production or its ineffective utilization in the human body. Uncontrolled and prolonged diabetes eventually brings several complications and disorders along with it. The most common amongst it, is the effect on retina of the eye which leads to Diabetic Retinopathy. Diabetic Retinopathy directly or indirectly affect the human vision and may also lead to irreparable loss. The pathogenesis of DR is ornately explained in the research study by Eshaq et al. [1]. DR is asymptomatic in its early stages, and the late diagnosis leads to undeviating loss of vision. Thus, the screening of Diabetic Retinopathy with the help of recent computer vision and image processing techniques helps in timely diagnosis and could assist ophthalmologist in its early treatment. Moreover, use of conventional methods during bulk screening programs is time-consuming, demand efforts and could be prone to error. The use of recent medical image processing algorithms could lessen the workload of ophthalmologists and results in an accurate diagnosis. Thus, it has been focused and developing research area in the past three decades for the screening of different stages of DR. DR is primarily divided into two stages viz. proliferative DR (PDR) and non-proliferative DR (NPDR). The pathologies involved in NPDR are microaneurysms (MA), dot and blot haemorrhages (HEM), exudates, cotton wool spots whereas PDR can lead to retinal detachment called as Neovascularisation (Fig. 1) [70]. MA are the prime indicators during the onset of Diabetic Retinopathy. These are tiny red dots and occur as a result of the focal dilations caused within thin vessels. These

⁎

microaneurysms rupture and lead to haemorrhages. These haemorrhages usually occur in the close vicinity of blood vessels. As MA and HEM mark the beginning of disease, the detection of these lesions is crucial for automated screening of DR. Thus, the paper reviews the existing literature available in automated screening of MA and HEM. The paper also discusses about the publicly available datasets, quantitative comparison, challenges and future research directions in computer aided screening of MA. The objectives of this paper are: (1) The detailed review of the existing literature for the detection of microaneurysms and haemorrhages in the screening of DR with the help of image processing techniques. (2) To enable a quantitative comparison, qualitative findings in the available methods and identify the existing gaps. (3) To find out the potential research areas and needed development to fill the gaps in the screening algorithms. Several approaches have been extended as a review to the existing algorithms in computer aided screening of Diabetic Retinopathy [2,3]. There has been an extensive investigation and research carried out for the extraction of MA and HEM from the retinal images. 2. Database There are various publicly available retinal datasets for carrying out research and scientific evaluations. The ultimate aim of all the available datasets is to provide an unambiguous platform to researchers to

Corresponding author. E-mail address: [email protected] (B.M. Patre).

https://doi.org/10.1016/j.biopha.2018.07.175 Received 17 May 2018; Received in revised form 31 July 2018; Accepted 31 July 2018 0753-3322/ © 2018 Elsevier Masson SAS. All rights reserved.

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

Fig. 1. Normal retina and retina with Diabetic Retinopathy. Source: Singapore Eye Research Institute

The annotated ground truth is not available for the database. e-Ophtha: e-Ophtha is a database of retinal images particularly designed for scientific research in DR [9]. The database consist two subsets of retinal images i.e. e-ophtha-MA (MicroAneurysms), and eophtha-EX (EXudates). Table 1 depicts the number of normal and abnormal images in various databases.

Table 1 Normal and abnormal images in databases. Dataset

ROC DIARETDB0 DIARETDB1 e-Ophtha-MA MESSIDOR

No. of images Normal

Abnormal

10 20 5 223 546

90 110 84 148 654

3. Performance metric Three extensively used performance metrics, namely Sensitivity (Se), Specificity (Sp) and Accuracy (Acc) are used for the validation of the algorithm. They are stated as follows:

validate their algorithms. These databases are acquired considering the practical environment and the methods run on similar kind of environment helps to make an easy comparison and thus justified inference. The commonly used datasets for MA detection are as follows: ROC: ROC database is a set of 100 retinal images selected from a huge dataset (150,000 images) gathered during a DR screening programme [4]. The images with three different resolutions are available in the database. The dataset was randomly divided into training and test set, each containing 50 images. The details of the image specification and database are illustrated in [5]. ROC (Retinopathy Online Challenge) [5] is also an unique MA detection competition which makes the dataset and evaluation methodology available on the same platform. DIARETDB: It is a publicly available database split into two groups, DIARETDB0 [6] and DIARETDB1 [7]. This database has been formed with a motive to unambiguously identify a testing protocol that can be a benchmark in the screening of DR. The comparison of different methods could be made under the same set of data. MESSIDOR: MESSIDOR is a database of 1200 images and provides the retinopathy grade and risk of macular edema for each image [8].

Se =

TP TP + FN

(1)

Sp =

TN TN + FP

(2)

Acc =

TP + TN TP + TN + FP + FN

(3)

where TP = rightly classified lesion, FP = non-lesion misclassified as lesion, TN = rightly classified non-lesion regions, FN = true lesion misclassified as non-lesion. The algorithm is statistically analysed with the help of Receiver Operating Characteristic (ROC) curve. It plots Sensitivity against False Positive Rate (1-Specificity). Free-response ROC (FROC) [10] is a curve that plots the sensitivity against the average number of false positive detection per image (FPI). FROC curve is summarised in a single number for a smoother comparison between the different algorithms. The single number is the average per lesion sensitivity at the FPI values ϵ [1/8, 1/4, 1/2, 1, 2, 4, 8], also known as competition metric (CPM). 682

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

Table 2 Performance measure of MA detection methods (on Private Dataset). Author

Dataset size

Highlights

Performance measure

Spencer et al. [13]

NA

Sensitivity-82%, Specificity-86%

Cree et al. [14]

20

Hansgen et al. [15]

NA

Frame et al. [16]

20

Hipwell et al. [17] Sinthanayothin et al. [18] Usher et al. [19] Fleming et al. [20] Streeter and Cree [21] Walter et al. [22] Niemeijer et al. [27] Quellec et al. [45] Saleh [60] Tavakoli et al. [62]

3783 30 773 1441 20 94 50 2739 98 170

Morphological Image Processing, Matched Filtering, Region Growing Morphological Image Processing, Matched Filtering, Region Growing Matched Filtering, Region Growing, Image Compression by DWT and JPEG Matched Filtering, Region Growing, LDA, SVQ-NN, RBS Classifier Intensity and Shape features, Red-free photographs Moat operator and RRGS Moat operator and RRGS Contrast Normalisation, Watershed RRGS Top-hat transform, matched filter, LDA Morphological Image Processing, Kernel Density Estimation Pixel classification using k-NN Optimal Filter Framework h-Maxima Transform, Multilevel Threholding Radon Transform, Multi-overlapping Windows

Sensitivity-82%, Specificity-84% Sensitivity-95.30%(DWT), Sensitivity-93.60%(JPEG) Sensitivity-84%, Specificity-85% Sensitivity-81%, Specificity-93% Sensitivity-77.50%, Specificity-88.70% Sensitivity-95.10%, Specificity-46.30% Sensitivity-85.40%, Specificity-83.10% Sensitivity 56% Sensitivity-88.5% Sensitivity-100%, Specificity-87% AUC: 0.927 Sensitivity 84.31%, Specificity 93.63% Sensitivity 94%, Specificity 75%(Database-I) Sensitivity 100%, Specificity 70%(Database-II)

Table 3 Performance measure of MA detection methods. Author

Dataset

Highlights

Performance measure

Akram et al. [34] Roychowdhury et al. [38]

DIARETDB, MESSIDOR DIARETDB1, MESSIDOR

Hybrid Classifier (GMM and m-Mediods) GMM, k-NN, SVM, AdaBoost Classifiers,

Seoud et al. [40]

Dynamic Shape Features, RF Classifier

Srivastava et al. [41]

DIARETDB1, CARA143, MESSIDOR DIARETDB1, MESSIDOR

Sensitivity 97.83%, Specificity 98.36% Sensitivity 80%, Specificity 85%(DIARETDB1) Sensitivity 100%, Specificity 53.16%(MESSIDOR) NA

Ren et al. [42]

E-Ophtha

Rocha et al. [49]

DIARETDB, MESSIDOR

Orlando et al. [52] Dai et al. [54] Agurto et al. [57]

MESSIDOR, DIARETDB1, E-Ophtha DIARETDB1 MESSIDOR

Figueiredo et al. [63]

DIARETDB, MESSIDOR

Kar et al. [65]

DIARETDB1, MESSIDOR

Multiple kernel learning, Support Vector Machines (SVM) Imbalanced data analysis, Extreme learning machine, Ensemble learning Visual Word Dictionaries, SURF features, SVM Classifier CNN, Ensemble Deep learning

ROC = 0.97

Multi-Sieving-CNN, image-to-text mapping Amplitude Modulation (AM) - Frequency Modulation (FM) features Anisotropic wavelet bands, Cartoontexture decomposition, Curvelet, Differential Evolution algorithm

Precision 99.7%, Recall 87.8% Sensitivity 92% Specificity 54%.

Sensitivity 96.1%, Specificity 82.1%, Sensitivity 90%, Specificity 60% Sensitivity 91.09% (MESSIDOR)

Sensitivity 72.48%, Specificity 80%(DB0) Sensitivity 77.38%, Specificity 80%(DB1) Sensitivity 75.11%, Specificity 91.81%(MESSIDOR) Sensitivity 95.23%, Specificity 95.12%,

Table 4 Performance measure of HEM detection methods. Author

Dataset

Highlights

Performance measure

Gardner et al. [23] Zhang and Chutatape [24] Hatanaka et al. [25]

NA Private Dataset (30)

ANN using back propagation 2-D PCA, SVM

Sensitivity-73.80% True Positive Ratio - 89.10%

Private Dataset (125)

Sensitivity- 80%, Specificity-80%

Tang et al. [37]

MESSIDOR

Srivastava et al. [41] Grinsven et al. [53] Saleh et al. [60]

DIARETDB1, MESSIDOR KAGGLE, MESSIDOR Private Dataset (98)

Density Analysis, Mahalanobis distance classifiers Splat based feature classification, k-NN classifier, Multiple kernel learning

Figueiredo et al. [63]

DIARETDB, MESSIDOR

Kar et al. [65]

DIARETDB1, MESSIDOR

Fast CNN, Selective Data Sampling h-maxima transform, multilevel threholding Anisotropic wavelet bands, Cartoontexture decomposition, Curvelet, Differential Evolution algorithm

683

ROC curve 0.96 ROC = 0.92 ROC = 0.894, ROC = 0.972. Sensitivity 87.53%, Specificity 95.08% Sensitivity 86%, Specificity 90%(MESSIDOR) Sensitivity 74.36%, Specificity 85% (DB0) Sensitivity 73.58%, Specificity 100% (DB1) Sensitivity 97.67%, Specificity 97.74%

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

Table 5 Performance measure of MA detection methods on ROC database. Author

FROC score

Salient features

Sinthanayothin et al. [18] Waikato [26]

0.36 0.206

Niemeijer [27] Giancardo et al. [28] Antal and Hajdu [32] Lazar and Hajdu [35]

0.395 0.375 0.434 0.424

Seoud et al. [40] Wu et al. [43]

0.420 0.376

Quellec et al. [45]

0.38

Zhang et al. [47]

0.357

Chudzik et al. [55] Mizutani et al. [29] Sanchez et al. [56]

0.193 0.31 0.32

Ram [58] Zhou et al. [64] Kar and Maity [65]

0.264 0.275 0.222

Wang et al. [66]

0.464

Moat operator and RRGS Morphological Image Processing, Matched Filtering, Region Growing Pixel classification using k-NN Radon Transform, PCA,SVM. Ensemble-based system. Directional cross-section profile features. Dynamic Shape Features, RF Classifier. Local and Profile features, k-NN classifier Template matching, Optimal Wavelet transform Template matching, multi scale correlation filtering CNN, Batch Normalisation layers Double ring filter, RBA, ANN Mixture model-based clustering, Logistic regression classifier. Clutter-rejection, SVM classifier, Sparse PCA, unsupervised classification Curvelet, Differential Evolution algorithm Intensity profile, Singular Spectrum Analysis

Fig. 3. Figure shows the frequency of distribution to different techniques.

MA detection performance is evaluated by plotting FROC curve in Retinopathy Online Challenge (ROC). 4. Review of screening methods Earlier, fluorescein angiograms were used for computer aided retinal lesion detection using various image processing techniques. The MA candidates are usually extracted after the removal of optic disk, blood vessels, fovea, macula in various state-of-the art investigations. The brief explanation and systematic inference of the existing literature till date for the computer aided diagnosis of red lesions is presented in the following section. The coarse sectioning according to the technique involved is made for an easy inference and flow of the screening algorithms. The five sections according to the methodology involved is elaborated as follows.

Table 6 Distribution of references according to methodology.

4.1. Morphological image processing

A. Morphological image processing Lay et al. [11], Spencer et al. [13], Cree et al. [14], Hangsen et al. [15], Hipwell et al. [17], Sinthayothin et al. [18], Usher et al. [19], Fleming et al. [20], Walter et al. [22].

The section mainly includes the techniques which involve morphological image processing and region growing algorithms as their dominant attribute. The computer aided screening of MA began during 1980s. The literature reveals the first method proposed by Lay and Baudoin [11] based on mathematical morphology in 1984. Thereafter, MA were approximated to the 2-D Gaussian matched filter using multiple window sizes and sigma to match the various sizes and shapes of the lesions in the method proposed by Spencer et al. [12]. Spencer et al. [13] have proposed the segmentation of MA candidates using bilinear top-hat transformation and matched filtering, followed by thresholding. The final segmentation was obtained through a region growing algorithm. The technique proved to be of great help to further researchers to propose the computer aided methods for more accurate screening of DR. The method by Cree et al. [14] was an extension to the technique proposed by Spencer [13]. The region growing segmentation (RGS) was designed to increase the specificity and the classification was better in the approach as eight intensity based features were extracted on each candidate in spite of the four in [13]. Hangsen et al. [15] have proposed an approach which focussed on image compression using wavelets. The same MA detection technique as of given by Spencer [13] and Frame [16] was used. The above techniques were conducted with the help of fluorescein angiograms. Angiography is an invasive and not a suitable method for the routine medical screening of DR, thus red free photographs were used in the approach proposed by Hipwell et al. [17]. The method [17] was analysed on 3783 digital red free fundus images instead of fluorescein angiograms [14] to detect MA. Sinthanayothin et al. [18] have introduced a new technique ‘moat operator’ used to sharpen the edges of lesion followed by Recursive RGS and thresholding. The method was applied to the database of 30 images, acquired promising results in detecting the lesions of NPDR. Usher et al. [19] have focussed on grading of Diabetic Retinopathy lesions by further using Artificial Neural Network (ANN) classifier as an extension to the technique proposed by Sinthanayothin [18]. The progress in automated MA detection methods with the use of

B. Supervised classification Frame et al. [16], Gardner et al. [23], Zhang and Chutatape [24], Hatanaka et al. [25], Streeter and Cree [21], Cree [26], Niemeijer et al. [27], Giancardo et al. [28], Mizutani et al. [29], Antal and Hajdu [32], Akram et al. [34], Lazar and Hazdu [35], Sopharak et al. [36], Tang et al. [37], Roychowdhury et al. [38], Seoud et al. [40], Srivastava et al. [41], Ren et al. [42], Wu et al. [43], C. Template matching and dictionary learning Quellec et al. [45], Zhang et al. [47], Ding et al. [48], Rocha et al. [49], Zhang et al. [50], Javidi et al. [51]. D. Deep learning Orlando et al. [52], Grinsven et al. [53], Dai et al. [54], Chudzik et al. [55]. E. miscellaneous Sanchez et al. [56], Agurto et al. [57], Ram et al. [58], Quellec et al. [59], Saleh et al. [60], Kose et al. [61], Tavakoli et al. [62], Figueiredo et al. [63], Zhou et al. [64], Kar and Maity [65], Wang et al. [66]

Fig. 2. Figure shows the frequency of distribution in different years.

684

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

method was evaluated on ROC database. Combinations of several pre-processing and candidate extraction were employed in this novel method given by Antal et al. [30] instead of one. An annealing based stochastic search algorithm helped to find the optimal combination and obtained promising results. The extension to the approaches presented in [30,31], was proposed by Antal in [32] as an ensemble system. The method was run on ROC and DIARETDB1 database and have attained leading results in the ROC competition. The method performed well under difficult circumstances but the detection of false positives at the instances where serious DR occurs, affected the performance of proposed system. A hybrid classifier approach including the Gaussian mixture model (GMM), support vector machine (SVM) and m-Mediods based modelling was presented in this method given by Akram et al. [33]. The approach in [33] is extended to improve the efficiency of classification in [34]. Hybrid Classifier was an extension of m-Mediods based modelling approach in combination with GMM. The system performs well in the grading of DR. MA's were analysed by rotating intensity profile centered at candidate pixel and the feature set was fed to Naive Bayes Classifier in [35] in this approach proposed by Lazar et al. The method needed analysis of local maxima of pre-processed image for lesion identification. The method distinguished amongst the vessel bifurcations and crossings from MA. The method attained competitive results in ROC. The method did not include optic disk detection, thus found false positives in the region of optic disc in some images. The coarse segmentation was employed using mathematical morphology followed by Naive Bayes classifier for the fine segmentation of MA in the work proposed by Sopharak et al. in [36]. The dependence of this algorithm on vessel segmentation led to some incorrect MA classification. Haemorrhages largely vary in shape and size. Thus, a splat based feature classification was proposed in [37] by Tang et al. to detect haemorrhages. Optimal splat features were selected with filter and wrapper approach and k-NN classifier was used for final stage classification. The segments of blood vessels which lied in connection with haemorrhages resulted in false positives, which reduced the sensitivity of the system. The author Roychowdhary et al. have analysed several classifiers as like k-NN, GMM, SVM and AdaBoost for the lesion classification in Diabetic Retinopathy in [38]. The paper dealt with several issues related in computer aided detection as like unbalanced dataset, lengthy no. of features and computation time requirement. Dynamic shape features that allow differentiation between lesion and vessel segments were proposed in the approach by Seoud et al. [39] and extended in [40], followed by Random Forest Classifier. The proposed approach did not need any prior segmentation of lesions and proved efficient in detection of blot HEM. The method ranked fourth in ROC analysis. However, the method faced difficulty in flame HEM detection. Also HEM linked to blood vessels were missed, which led to false negatives. The method proposed by Srivastava et al. [41] focused on the false detection of small vessel segments which interfere during the candidate extraction. Various grid sizes in filtering followed by multiple kernel learning outperformed the existing literature in MA detection. The novel filters dealt with the false positives due to small vessel segments. The method by Ren et al. [42] have addressed the issue of class-imbalance and proposed ensemble adaptive over-sampling algorithm for improved MA detection. An ensemble framework comprising of bagging, boosting and random subspaces have been used in the approach. The method performed well with respect to class-imbalance problem and false positive reduction. 27 local and profile features were extracted and fed as input to k-NN classifier in the approach by Wu et al. [43]. The profile features mentioned in [35] along with new additions were illustrated. The method was evaluated on ROC and e-Ophtha database. The supervised classification methods led to the enhanced and more accurate screening of red lesions. The essence of these classification algorithms lies in the optimal selection of features and

local contrast enhancement and vessel detection has been described in this work by Fleming [20]. Region growing algorithm using watershed is a distinct feature of this algorithm. Candidate microaneurysms were extracted with the help of top-hat transform and optimal matched filtering in [21] proposed by Cree and Streeter. After region growing technique, Linear Discriminant Analysis (LDA) was used as the final stage classifier. The limitation of the algorithm was, it could only detect lesion greater than 10 pixels in size due to some high frequency noises captured during the image acquisition. A one more morphological image processing approach was proposed by Walter et al. [22]. The method did not include region growing segmentation and used kernel density estimation along with Gaussian and k-NN (k-Nearest Neighbour) classifier to measure classification. The major limitation in the region growing and morphological approach is the detection of large no. of false positives which affect the overall specificity of the algorithm. 4.2. Supervised classification The feature extraction followed by the use of classifier is most widely studied domain in the screening of red lesions. The optimal selection of features and classifier is mandate requirement for the effective screening algorithm. Artificial Neural Network (ANN) using back propagation was employed for the first time to detect the lesions of fundus image in [23]. The technique which was trained using 179 images and evaluated on 300 images, proved to be a potential tool in the screening of DR. A further addition in the classification technique was an extension to the work of Spencer [13] and Cree [14], proposed by Frame et al. [16]. Thirteen features were extracted from each candidate and were given as input to three individual classifiers viz. Rule Based Analysis (RBA), Linear Discriminant Analysis (LDA) and Linear Vector Quantisation (LVQ) to classify MA. Rule based analysis gave superior performance at the cost of computation time required. The features were extracted with the help of 2-D Principle Component Analysis for Haemorrhage detection in [24] proposed by Zhang and Chutatape. Support Vector Machines (SVM) was used for classification of the HEM. The method attained an improved accuracy by the use of rotation and illuminance invariance to obtain virtual support vectors. HEM detection method was proposed by Hatanaka et al. [25] using density analysis. Mahalanobis distance classifiers and rule based analysis were used in particular to remove false positives. An automated MA Detector named ‘Waikato Microaneurysms Detector’ was designed using approaches mentioned in Spencer [13], Cree [14], Hipwell [17] for retinopathy online screening challenge in [26]. NaiveBayes classifier was used for classification. A novel hybrid approach in combination with the earlier works of Spencer [13] and Cree [14] proposed by Niemeijer et al. [27] outperformed all the existing approaches. Pixel classification based candidate detection and the use of k-Nearest Neighbour (k-NN) classifier were introduced for lesion detection. Though the performance was encouraging, the more advanced classifiers could optimise the performance and reduce the computation time. ‘Retinopathy Online Challenge’ (ROC) is the first international MA Detection competition organized by Niemeijer and his team [5]. The purpose is to compare the various published methods on the same set of data and evaluate the individual performance. The set of data associated with ROC are publicly available and submissions are accepted for every year challenge. It is a great platform to identify the best performing methods and the required room for improvement. Radon space based features were analysed in this work proposed by Giancardo et al. [28]. Principal Component Analysis (PCA) and SVM assisted in an easy training and classification of MA with a superior performance. Blood vessel segmentation was not required in this approach. A double ring filter was used by Mizutani et al. [29] to extract the MA candidates. After removal of blood vessels, rule based analysis and a back propagation ANN was used for the classification of MA. The 685

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

4.5. Miscellaneous techniques

relevant classifier for the technique.

Though various techniques have been put forward by researchers, the removal of false positives and fast processing is still the grave need for lesion detection. Moreover, the subtle MA and the one very close to blood vessels still are very difficult to be segmented in many computer aided algorithms. Thus, few miscellaneous techniques have been proposed for the extraction of red lesions. A statistical mixture-model based clustering was proposed for candidate identification in this method given by Sanchez et al. [56]. Logistic Regression (LR) classifier was used for classification of the candidates as MA or non MA. The limitation of the approach was larger no. of false positives and missed MA adjacent to blood vessels. The retinal structures were characterised based on their instantaneous amplitude and frequency in this method proposed by Agurto et al. [57]. Amplitude Modulation and Frequency Modulation features were used to differentiate lesions in retinal image. MA detection was formulated as finding target from clutter by successive rejection of clutter responses in [58] by Ram et al. After rejection of clutters, a score was assigned to the each remaining candidate by finding its degree of similarity to true MA candidate. The method faced difficulty in candidate selection due to variability amongst datasets. The optimal set of filters were generated by Quellec et al. [59] to detect lesions almost in less than a second. The feature space was automatically derived from a set of reference image samples including target lesions. A decision support system was proposed in [60] involved in automated screening of MA and HEM. H-maxima transform was used to extract the candidate lesions. An inverse segmentation based technique to detect various lesions and use of Naive Bayes classifier was applied in this approach given by Kose et al. [61]. The algorithm was tested on various datasets and attained acceptable results. The approach proposed by Tavakoli et al. [62] used Radon transform for segmentation of lesions followed by thresholding. The algorithm was run on fluorescein angiography fundus images. The approach detected higher no. of false negatives in the areas where MA were adjacent to blood vessels or larger in size. The features were extracted by using several wavelet bands and combination of Hessian multiscale analysis followed by binary classifiers in [63] by Figueirdo et al. The method performed well when considered in numbers of sensitivity and specificity. An unsupervised classification algorithm based on Sparse PCA was introduced for effective MA detection in [64] by Zhou et al. The statistic measure was defined which restricts the non-MA samples, and the approach took care of class-imbalance issue. The algorithm was run on ROC database and led to competitive results. The approach found some false positives due to missing of MA during feature extraction. Kar and Maity [65] have used curvelet to extract lesions and Differential Evolution (DE) algorithm was used in optimizing the thresholding operation. The mutual information of matched filter response and LoG filter response were also optimised by DE algorithm. The system surpassed the existing approaches but needed a high run time. The intensity profiles of MA were processed with the help of Singular spectrum analysis in this work proposed by Wang et al. [66]. The features of the profile were input to the k-NN classifier. The method performed well for large datasets and also in the presence of artefacts. The proposed approached experienced small no. of false negatives due to missing of subtle or low contrast MA. They also missed few MA during candidate extraction. The summary of MA detection methods is presented in Tables 2 and 3. The summary of HEM detection methods is briefed in Table 4. FROC is a measure of statistic quantitative comparison for the algorithms evaluated on ROC database. The summary of MA detection methods validated on ROC database is presented in Table 5. The approach proposed by Wang et al. [66] outperforms all the approaches with a FROC score of 0.464.

4.3. Template matching and dictionary learning The kernel matching the size of MA lesions is selected as core element in the template matching algorithms. The template matching using wavelet transform for MA detection was introduced by Quellec et al. [44]. The same work have been extended in [45] by use of optimal wavelet transform and optimization using genetic algorithm. The template of sub-bands of transformed image was matched to the lesion template. The processing time required was very low, also it did not require extracting candidates and classification. The proposed algorithm missed to extract MA adjacent to blood vessels. The another efficient template matching approach using multi-scale correlation filtering was proposed by Zhang et al. [46]. The approach was revised and again presented in [47]. The high correlation co-efficient was calculated for each pixel using sliding window multi-scale Gaussian kernel. The requirement of high correlation co-efficient depends on the value of sigma. The approach was implemented on ROC database and received promising results. The limitation of the approach was dependence on the scale of Gaussian kernel and the images with low contrast resulted in detection of false positives. A template matching algorithm was also proposed by Ding et al. [48] termed as Dynamic multi-parameter template (DMPT) matching for MA detection. The candidates extracted through the DMPT algorithm were scored using adaptive weighing and summing technique. The dictionaries were learned using the samples extracted from content of training images and those were used for classifying the lesion candidates in Dictionary Learning algorithms. The use of visual words and dictionary in lesion identification was proposed for the first time in this paper by Rocha et al. [49]. Region of Interest (RoI) containing lesions were identified and designated as words of the visual dictionary. A specific dictionary for each lesion was formed and SVM was used as final stage classifier. The approach worked satisfactorily independent of color space or image resolution. A dictionary learning approach using sparse representation classifier was proposed in the paper by Zhang et al. [50]. Two dictionaries, each for MA and non MA were formed using the candidates extracted by multi scale Gaussian Filtering. Two discriminative dictionaries were learned and used to classify the extracted candidates obtained by applying the Morlet Wavelet in [51] by Javidi et al. The method attained promising results in sparse representation domain. 4.4. Deep learning Deep Learning is an emerging computer vision application in medical image processing. Few methods have been proposed and several other researches are ongoing in this domain. Features were learned using Convolutional Neural Networks (CNN) and ensemble vector was thereafter classified using Random Forest Classifier in the method proposed by Orlando et al. [52]. The limitation of CNN network is the requirement of larger training time. Selective data sampling was used to reduce the time required to train a CNN network for detection of haemorrhages in [53]. The technique dynamically selected the informative samples during training stage, and higher weights were assigned to them. The final trained CNN network classifies each pixel as lesion or not. Multi-sieving convolutional neural network was used to detect the potential MA regions via image to text mapping in feature space in [54]. Results were evaluated on clinical datasets and are encouraging. A full convolutional neural network was employed in this method proposed by Chudzik et al. [55]. Each Convolutional layer was followed by batch normalisation layer and the dice loss function was used to detect MA. The algorithm was run on three standard databases and attained competitive results. 686

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

councils in collaboration with funding agencies are taking initiatives for creation of database with ground truth. There are several algorithms in the literature which can aid the ophthalmologist in an easy computer aided screening of DR. The method is rated as best when it is quick, cost-effective and accurate. All these constraints altogether is the need of time. It is indeed difficult for the researchers to identify the best algorithms and get the efficient throughputs. Young researchers have immense emerging areas like deep learning but with minimal training time, dictionary learning for specific lesions, finding best features and robust classifiers to attain the highest FROC scores and accuracy.

The distribution of all the existing algorithms according to the methodology involved is mentioned in Table 6. 5. Discussion Some 425 million people are estimated to be suffering with diabetes globally [67]. Diabetes leads amongst the worldwide health emergencies of the century. This chronic disease affect several organs of the human body and Diabetic Retinopathy is an outcome of it is effect on retina. The diagnosis of DR is asymptotic in its onset stages and the late findings may have severe loss of vision. Currently, about 93 million people suffer from some sort of eye damage due to diabetes. The ratio of a person having DR if one is suffering with diabetes, is 1:3 [68]. Thus, it's a grave need to get it screened in initial stages and have confined treatment before an irreparable loss. Several novel techniques have led to the effective results and the accuracy of the algorithms is increasing each upcoming day. The efficiency of overall system is the need and each step in it should ensure high sensitivity and specificity of the proposed algorithm. Although considerable achievements have been made in retinal image analysis, there is still room for finding the best algorithm which surpasses all in terms of accuracy. Moreover, the proposed method should bear minimum false negatives and help the ophthalmologists to focus on the lesions in fundus images. The literature showcases various techniques proposed by several researchers and also refinement made by few others to increase accuracy of the algorithm. The bar chart of frequency of distribution of existing methods as per their publication year and methodology involved is depicted in Fig. 2 and Fig. 3 respectively. Few initial research papers [13,14,16,69,19] made emphasis on mathematical morphology and region growing approaches. It was a good initiative with basic image processing methods but there were issues like size of structuring element which varies with the size of lesion, larger no. of false positives. The another approach is a typical process in computer vision which includes candidate extraction, feature selection and classification. This led to several advanced and helpful CAD methods. The use of ANN [23],k-NN [27,43], SVM [41,49] and ensembled classifiers [31,33,38] yielded high performance metrics. The template matching strategies [45,47] also provided a distinct and useful perspective to detect lesions. Sparse representation and dictionary learning techniques [50,51] have also evolved competitive results. Few researchers combined mathematical morphology and supervised classification which added to the performance of the existing approaches. The miscellaneous techniques [58,59,65,66] used by the researchers have gained great accuracy in the screening of red lesions. The major issue in CAD of MA detection is the failure of an algorithm to detect the MA's very close to blood vessels or very subtle MA. The algorithm must also be robust to the varying imaging conditions. Eventually, the performance of the system needs to be evaluated with the best features, diverse images and robust classifiers. Deep learning is an emerging area and proving to be of great help to mankind in machine learning. Medical imaging is also turning towards few CNN techniques. Few algorithms have been developed using CNN [52,54,55] techniques and the results are fruitful, the only limitation is CNN training is time-consuming and still challenging.

References [1] R.S. Eshaq, A.M. Aldalati, J.S. Alexander, N.R. Harris, Diabetic retinopathy: breaking the barrier, Pathophysiology 24 (4) (2017) 229–241. [2] M.R.K. Mookiah, U.R. Acharya, C.K. Chua, C.M. Lim, E. Ng, A. Laude, Computeraided diagnosis of diabetic retinopathy: a review, Comput. Biol. Med. 43 (12) (2013) 2136–2155. [3] R. Winder, P. Morrow, I. McRitchie, J. Bailie, P. Hart, Algorithms for digital image processing in diabetic retinopathy, Comput. Med. Imaging Graphics 33 (8) (2009) 608–622. [4] M.D. Abramoff, M.S. Suttorp-Schulten, Web-based screening for diabetic retinopathy in a primary care population: the eyecheck project, Telemed. e-Health 11 (6) (2005) 668–674. [5] M. Niemeijer, B. van Ginneken, M.J. Cree, A. Mizutani, G. Quellec, C.I. Sanchez, B. Zhang, R. Hornero, M. Lamard, C. Muramatsu, X. Wu, G. Cazuguel, J. You, A. Mayo, Q. Li, Y. Hatanaka, B. Cochener, C. Roux, F. Karray, M. Garcia, H. Fujita, M.D. Abramoff, Retinopathy online challenge: automatic detection of microaneurysms in digital color fundus photographs, IEEE Trans. Med. Imaging 29 (1) (2010) 185–195. [6] T. Kauppi, V. Kalesnykiene, J.-K. Kamarainen, L. Lensu, I. Sorri, H. Uusitalo, H. Kälviä inen,, J. Pietilä, Diaretdb0: Evaluation Database and Methodology for Diabetic Retinopathy Algorithms, (2007). [7] T. Kauppi, V. Kalesnykiene, J.-K. Kamarainen, L. Lensu, I. Sorri, A. Raninen, R. Voutilainen, H. Uusitalo, H. Kälviäinen, J. Pietilä, Diaretdb1 diabetic retinopathy database and evaluation protocol, Proc. Medical Image Understanding and Analysis (MIUA), vol. 2007 (2007). [8] E. Decencière, X. Zhang, G. Cazuguel, B. Lay, B. Cochener, C. Trone, P. Gain, R. Ordonez, P. Massin, A. Erginay, B. Charton, J.-C. Klein, Feedback on a publicly distributed image database: the MESSIDOR Database, Image Anal. Stereol. 33 (3) (2014). [9] E. Decencière, G. Cazuguel, X. Zhang, G. Thibault, J.-C. Klein, F. Meyer, B. Marcotegui, G. Quellec, M. Lamard, R. Danno, D. Elie, P. Massin, Z. Viktor, A. Erginay, B. Laÿ, A. Chabouis, Teleophta: machine learning and image processing methods for teleophthalmology, IRBM 34 (2) (2013) 196–203 (special issue: ANR TECSAN: Technologies for Health and Autonomy). [10] P.C. Bunch, J.F. Hamilton, G.K. Sanderson, A.H. Simmons, A free response approach to the measurement and characterization of radiographic observer performance, Proc. SPIE 0127 (1977) 0127-0127-12. [11] B. Lay, C. Baudoin, J.-C. Klein, Automatic detection of microaneurysms in retinopathy fluoro-angiogram, Proc. SPIE 0432 (1984) 0432-0432-9. [12] T. Spencer, R. Philips, P. Sharp, J. Forrester, Automated detection and quantification of microaneurysms in fluorescein angiograms, Graefe's Arch. Clin. Exp. Ophthalmol. 230 (1992) 230–236. [13] T. Spencer, J. Olson, K.C. McHardy, P. Sharp, J. Forrester, An image-processing strategy for the segmentation and quantification of microaneurysms in fluorescein angiograms of the ocular fundus, Comput. Biomed. Res. 29 (4) (1996) 284–302. [14] M.J. Cree, J.A. Olson, K.C. McHardy, P.F. Sharp, J.V. Forrester, A fully automated comparative microaneurysm digital detection system, Eye 11 (1997) 622–628. [15] P. Hansgen, P. Undrill, M. Cree, The application of wavelets to retinal image compression and its effect on automatic microaneurysm analysis, Comput. Methods Progr. Biomed. 56 (1) (1998) 1–10. [16] A.J. Frame, P.E. Undrill, M.J. Cree, J.A. Olson, K.C. McHardy, P.F. Sharp, J.V. Forrester, A comparison of computer based classification methods applied to the detection of microaneurysms in ophthalmic fluorescein angiograms, Comput. Biol. Med. 28 (3) (1998) 225–238. [17] J.H. Hipwell, F. Strachan, J.A. Olson, K.C. McHardy, P.F. Sharp, J.V. Forrester, Automated detection of microaneurysms in digital red-free photographs: a diabetic retinopathy screening tool, Diabet. Med. 17 (8) (2000) 588–594. [18] C. Sinthanayothin, J.F. Boyce, T.H. Williamson, H.L. Cook, E. Mensah, S. Lal, D. Usher, Automated detection of diabetic retinopathy on digital fundus images, Diabet. Med. 19 (2) (2002) 105–112. [19] D. Usher, M. Dumskyj, M. Himaga, T.H. Williamson, S. Nussey, J. Boyce, Automated detection of diabetic retinopathy in digital retinal images: a tool for diabetic retinopathy screening, Diabet. Med. 21 (1) (2004) 84–90. [20] A.D. Fleming, S. Philip, K.A. Goatman, J.A. Olson, P.F. Sharp, Automated microaneurysm detection using local contrast normalization and local vessel detection, IEEE Trans. Med. Imaging 25 (9) (2006) 1223–1232. [21] L.V. Streeter, M.J. Cree, Microaneurysm detection in colour fundus images, Image Vision Comput. (2018).

6. Conclusion Recent trends in medical image processing are assisting in an easy and automated screening of the disease. We discussed various algorithms proposed in literature for the automated screening of MA and HEM. The sufficiently high sensitivity and specificity are mandate requirements for any screening method. The British Diabetic Association proposed that any screening programme for Diabetic Retinopathy should have at least 80% sensitivity and 95% specificity. Moreover the identification and formation of ground truth is one of the fundamental issue in validation of automated screening of DR. Several research 687

Biomedicine & Pharmacotherapy 107 (2018) 681–688

R.S. Biyani, B.M. Patre

[22] T. Walter, P. Massin, A. Erginay, R. Ordonez, C. Jeulin, J.-C. Klein, Automatic detection of microaneurysms in color fundus images, Med. Image Anal. 11 (6) (2007) 555–566. [23] G. Gardner, D. Keating, T.H. Williamson, A.T. Elliott, Automatic detection of diabetic retinopathy using an artificial neural network: a screening tool, Br. J. Ophthalmol. 80 (1996) 940–944. [24] X. Zhang, O. Chutatape, A SVM approach for detection of hemorrhages in background diabetic retinopathy, Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005, vol. 4 (2005) 2435–2440. [25] Y. Hatanaka, T. Nakagawa, Y. Hayashi, M. Kakogawa, A. Sawada, K. Kawase, T. Hara, H. Fujita, Improvement of automatic hemorrhage detection methods using brightness correction on fundus images, 2008 International Conference of the IEEE Engineering in Medicine and Biology Society (2008) 5429–5432. [26] M. Cree, The Waikato Microaneurysm Detector, The University of Waikato, 2018. [27] M. Niemeijer, B. van Ginneken, J. Staal, M.S.A. Suttorp-Schulten, M.D. Abramoff, Automatic detection of red lesions in digital color fundus photographs, IEEE Trans. Med. Imaging 24 (5) (2005) 584–592. [28] L. Giancardo, F. Meriaudeau, T.P. Karnowski, Y. Li, K.W. Tobin, E. Chaum, Microaneurysm detection with radon transform-based classification on retina images, 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (2011) 5939–5942. [29] A. Mizutani, C. Muramatsu, Y. Hatanaka, S. Suemori, T. Hara, H. Fujita, Automated microaneurysm detection method based on double ring filter in retinal fundus images, Proc. SPIE 7260 (2009) 7260-7260-8. [30] B. Antal, I. Lazar, A. Hajdu, Z. Torok, A. Csutak, T. Peto, A multi-level ensemblebased system for detecting microaneurysms in fundus images, 4th International Workshop on Soft Computing Applications (2010) 137–142. [31] B. Antal, A. Hajdu, Improving microaneurysm detection using an optimally selected subset of candidate extractors and preprocessing methods, Pattern Recognit. 45 (1) (2012) 264–270. [32] B. Antal, A. Hajdu, An ensemble-based system for microaneurysm detection and diabetic retinopathy grading, IEEE Trans. Biomed. Eng. 59 (6) (2012) 1720–1726. [33] M.U. Akram, S. Khalid, S.A. Khan, Identification and classification of microaneurysms for early detection of diabetic retinopathy, Pattern Recognit. 46 (1) (2013) 107–116. [34] M.U. Akram, S. Khalid, A. Tariq, S.A. Khan, F. Azam, Detection and classification of retinal lesions for grading of diabetic retinopathy, Comput. Biol. Med. 45 (2014) 161–171. [35] I. Lazar, A. Hajdu, Retinal microaneurysm detection through local rotating crosssection profile analysis, IEEE Trans. Med. Imaging 32 (2) (2013) 400–407. [36] A. Sopharak, B. Uyyanonvara, S. Barman, Simple hybrid method for fine microaneurysm detection from non-dilated diabetic retinopathy retinal images, Comput. Med. Imaging Graphics 37 (5) (2013) 394–402. [37] L. Tang, M. Niemeijer, J.M. Reinhardt, M.K. Garvin, M.D. Abramoff, Splat feature classification with application to retinal hemorrhage detection in fundus images, IEEE Trans. Med. Imaging 32 (2) (2013) 364–375. [38] S. Roychowdhury, D.D. Koozekanani, K.K. Parhi, Dream:, Diabetic retinopathy analysis using machine learning, IEEE J. Biomed. Health Inform. 18 (5) (2014) 1717–1728. [39] L. Seoud, T. Faucon, T. Hurtut, J. Chelbi, F. Cheriet, J.M.P. Langlois, Automatic detection of microaneurysms and haemorrhages in fundus images using dynamic shape features, 2014 IEEE 11th International Symposium on Biomedical Imaging (ISBI) (2014) 101–104. [40] L. Seoud, T. Hurtut, J. Chelbi, F. Cheriet, J.M.P. Langlois, Red lesion detection using dynamic shape features for diabetic retinopathy screening, IEEE Trans. Med. Imaging 35 (4) (2016) 1116–1126. [41] R. Srivastava, L. Duan, D.W. Wong, J. Liu, T.Y. Wong, Detecting retinal microaneurysms and hemorrhages with robustness to the presence of blood vessels, Comput. Methods Progr. Biomed. 138 (2017) 83–91. [42] F. Ren, P. Cao, W. Li, D. Zhao, O. Zaiane, Ensemble based adaptive over-sampling method for imbalanced data learning in computer aided detection of microaneurysm, Comput. Med. Imaging Graphics 55 (2017) 54–67. [43] B. Wu, W. Zhu, F. Shi, S. Zhu, X. Chen, Automatic detection of microaneurysms in retinal fundus images, Comput. Med. Imaging Graphics 55 (2017) 106–112. [44] G. Quellec, M. Lamard, P.M. Josselin, G. Cazuguel, B. Cochener, C. Roux, Detection of lesions in retina photographs based on the wavelet transform, 2006 International Conference of the IEEE Engineering in Medicine and Biology Society (2006) 2618–2621. [45] G. Quellec, M. Lamard, P.M. Josselin, G. Cazuguel, B. Cochener, C. Roux, Optimal wavelet transform for the detection of microaneurysms in retina photographs, IEEE Trans. Med. Imaging 27 (9) (2008) 1230–1241. [46] B. Zhang, X. Wu, J. You, Q. Li, F. Karray, Hierarchical detection of red lesions in

[47] [48]

[49]

[50] [51]

[52]

[53]

[54]

[55]

[56]

[57]

[58]

[59]

[60]

[61]

[62]

[63] [64]

[65] [66]

[67] [68]

[69]

[70]

688

retinal images by multiscale correlation filtering, Proc. SPIE 7260 (2009) 72607260-12. B. Zhang, X. Wu, J. You, Q. Li, F. Karray, Detection of microaneurysms using multiscale correlation coefficients, Pattern Recognit. 43 (6) (2010) 2237–2248. S. Ding, W. Ma, An accurate approach for microaneurysm detection in digital fundus images, 2014 22nd International Conference on Pattern Recognition (2014) 1846–1851. A. Rocha, T. Carvalho, H.F. Jelinek, S. Goldenstein, J. Wainer, Points of interest and visual dictionaries for automatic retinal lesion detection, IEEE Trans. Biomed. Eng. 59 (8) (2012) 2244–2253. B. Zhang, F. Karray, Q. Li, L. Zhang, Sparse representation classifier for microaneurysm detection and retinal blood vessel extraction, Inf. Sci. 200 (2012) 78–90. M. Javidi, H.-R. Pourreza, A. Harati, Vessel segmentation and microaneurysm detection using discriminative dictionary learning and sparse representation, Comput. Methods Progr. Biomed. 139 (2017) 93–108. J.I. Orlando, E. Prokofyeva, M. del Fresno, M.B. Blaschko, An ensemble deep learning based approach for red lesion detection in fundus images, Comput. Methods Progr. Biomed. 153 (2018) 115–127. M.J.J.P. van Grinsven, B. van Ginneken, C.B. Hoyng, T. Theelen, C.I. Sanchez, Fast convolutional neural network training using selective data sampling: application to hemorrhage detection in color fundus images, IEEE Trans. Med. Imaging 35 (5) (2016) 1273–1284. L. Dai, R. Fang, H. Li, X. Hou, B. Sheng, Q. Wu, W. Jia, Clinical report guided retinal microaneurysm detection with multi-sieving deep learning, IEEE Trans. Med. Imaging PP (99) (2018). P. Chudzik, S. Majumdar, F. Caliv, B. Al-Diri, A. Hunter, Microaneurysm detection using fully convolutional neural networks, Comput. Methods Progr. Biomed. 158 (2018) 185–192. C.I. Sánchez, R. Hornero, A. Mayo, M. García, Mixture model-based clustering and logistic regression for automatic detection of microaneurysms in retinal images, Proc. SPIE 7260 (2009) 7260-7260-8. C. Agurto, V. Murray, E. Barriga, S. Murillo, M. Pattichis, H. Davis, S. Russell, M. Abramoff, P. Soliz, Multiscale AM-FM methods for diabetic retinopathy lesion detection, IEEE Trans. Med. Imaging 29 (2) (2010) 502–512. K. Ram, G.D. Joshi, J. Sivaswamy, A successive clutter-rejection-based approach for early detection of diabetic retinopathy, IEEE Trans. Biomed. Eng. 58 (3) (2011) 664–673. G. Quellec, S.R. Russell, M.D. Abramoff, Optimal filter framework for automated, instantaneous detection of lesions in retinal images, IEEE Trans. Med. Imaging 30 (2) (2011) 523–533. D. Saleh, C. Eswaran, An automated decision-support system for non-proliferative diabetic retinopathy disease based on ma's and ha's detection, Comput. Methods Progr. Biomed. 108 (2012) 186–196. C. Kose, U. Sevik, C. Ikibas, H. Erdol, Simple methods for segmentation and measurement of diabetic retinopathy lesions in retinal fundus images, Comput. Methods Progr. Biomed. 107 (2) (2012) 274–293. M. Tavakoli, R.P. Shahri, H. Pourreza, A. Mehdizadeh, T. Banaee, M.H.B. Toosi, A complementary method for automated detection of microaneurysms in fluorescein angiography fundus images to assess diabetic retinopathy, Pattern Recognit. 46 (10) (2013) 2740–2753. I. Figueiredo, S. Kumar, C. Oliveira, J. Ramos, B. Engquist, Automated lesion detectors in retinal fundus images, Comput. Biol. Med. 66 (2015) 47–65. W. Zhou, C. Wu, D. Chen, Y. Yi, W. Du, Automatic microaneurysm detection using the sparse principal component analysis-based unsupervised classification method, IEEE Access 5 (2017) 2563–2572. S.S. Kar, S.P. Maity, Automatic detection of retinal lesions for screening of diabetic retinopathy, IEEE Trans. Biomed. Eng. 65 (3) (2018) 608–618. S. Wang, H.L. Tang, L.I.A. Turk, Y. Hu, S. Sanei, G.M. Saleh, T. Peto, Localizing microaneurysms in fundus images through singular spectrum analysis, IEEE Trans. Biomed. Eng. 64 (5) (2017) 990–1002. International Diabetes Federation, 2017, IDF Diabetes Atlas, 8th ed., (2017) ISBN: 978-2-930229-87-4, http://www.diabetesatlas.org/. International Diabetes Federation and The Fred Hollows Foundation, Diabetes Eye Health: A Guide for Health Care Professionals, International Diabetes Federation, Brussels, Belgium, 2015 ISBN: 978-2-930229-82-9, http://www.idf.org/eyecare. C. Sinthanayothin, V. Kongbunkiat, S. Phoojaruenchanachai, A. Singalavanija, Automated screening system for diabetic retinopathy, Proceedings of the 3rd International Symposium on Image and Signal Processing and Analysis, 2003. ISPA 2003, vol. 2 (2003) 915–920. M.D. Abramoff, M.K. Garvin, M. Sonka, Retinal imaging and image analysis, IEEE Rev. Biomed. Eng. 3 (2010) 169–208.