A review of retinal blood vessels extraction techniques: challenges, taxonomy, and future trends

Abstract

The visual exploration of retinal blood vessels assists ophthalmologists in the diagnoses of different abnormalities of the eyes such as diabetic retinopathy, glaucoma, cardiovascular ailment, high blood pressure, arteriosclerosis, and age-related macular degeneration. The manual inspection of retinal vasculature is an extremely challenging and tedious task for medical experts due to the complex structure of an eye, tiny blood vessels, and variation in vessels width. Several automatic retinal vessels extraction techniques have been proposed in contemporary literature, which assist ophthalmologists in the timely identification of an eye disorders. However, due to the fast evolution of such techniques, a comprehensive survey is needed. This survey presents a comprehensive review of such techniques, strategies, and algorithms presented to date. The techniques are classified into logical groups based on the underlying methodology employed for retinal vessel extraction. The performance of existing techniques is reported on the publicly accessible datasets in term of various performance measures, providing a valuable comparison among the techniques. Thus, this survey presents a valuable resource for researchers working toward automatic extraction of retinal vessels.

1 Introduction

The human fundus RGB photographs are the projection of the inner surface of an eye. Retinal images are widely used in a noninvasive way by medical experts in the analysis of visual impairments [1]. In the early 1960s, optical photography made an important contribution in the medical prescription [2]. The first photographic retina images indicating blood vessels were discovered by German ophthalmologist, Gerloff in 1891. Gullstrand first developed the fundus camera in 1910 [3]. Herbert and Michael expressed that the advancement of ocular photographs and expanding research in fundus photograph examination may be because of the necessities of clinical procedure to discover better and less expensive methods of diagnosing, classifying, and screening fundus abnormalities [4].

The fundus image of the retina uncovers its structural format such as retinal vessels map, optic disk (OD), fovea, macula, and pathological structures such as microaneurysms (MAs), hemorrhages, exudates, cotton wool spots, if exist [5]. The optic nerve head begins from OD (observed as a bright oval shape) and is the passage point for major vessels to the eye. Macula (exposed as a dull zone without vessels) with the fovea at its middle is responsible for central and high-resolution eyesight. The blood vessel network has a high contrast due to high frequency stretched over the entire fundus image. The retinal veins provide blood supply to the retina and transmit the data signals from the retina to the brain [6].

Retinal vessel extraction and characterization of morphological attributes such as diameter, shape, tortuosity, and bifurcation can be utilized in screening, evaluation, and medication purposes of different eye abnormalities [7]. The retinal vasculature variation is used as an indication for detection of different abnormalities such as DR, hypertensive retinopathy, retinal artery/vein occlusion, and other disorders, for example stroke, hypertension and diabetes, induced modifications in the structure of retinal vessels. Evaluation of retinal blood vessel attributes, for example variation in width, is utilized to analyze hypertension, while bifurcation points and tortuosity can help with the identification of cardiovascular ailments and DR [8]. Most of the retinal diseases that produce alterations in the retinal vessel structure would cause visual impairments. Eyesight can be prevented largely by the detection of such pathologies at the initial stage [9].

Automatic precise identification and diagnoses of eye abnormalities and their timely medication are important in restraining blindness. However, any computerized examination of the retinal vessel structure needs its accurate extraction first. Traditionally, manual segmentation is performed by trained experts that is very hard and time-consuming task [10]. Moreover, the complexities in the imaging procedure, such as the lower contrast difference between vessel and non-vessel pixels, unbalanced background illumination and the vessel width variation, brightness and shape, reduce significantly the coincidence among segmentations performed by various human graders [11]. These facts stimulate the advancement of automatic techniques for retinal vessel extraction without human interference.

Various retinal vessel segmentation techniques in the literature are summarized in different survey and reviews. For example, Mabrouk et al. [12] introduced a concise review of retinal detection and registration techniques, which restrict explanation to the detection of the centers and boundaries of retinal vessels. John et al. [13] and Oliver et al. [14] discussed different frameworks for screening of DR in fundus images. Fraz et al. [6] presented an extensive review of the different methods and algorithms for extraction of retinal vessels. They provide a comprehensive study of various retinal vessel detection algorithms; pros and cons of different techniques; current patterns and future guidelines and outline the open issues. They discussed a limited number of retinal image datasets. The most recent survey discusses the techniques for retinal vessel detection and challenges associated with retina evaluation [7].

In this article, we have discussed the retinal image acquisition and different abnormalities associated with it. Comparison tables of all openly accessible datasets are added. The target of this article is to inspect various retinal vessel detection schemes; to show a comparison of different techniques utilized for retinal vessel extraction; to disclose the pros and cons of various methods; to discuss future guidelines and outline the open issues. We have covered a wide range of published papers from the beginning until now and also divide these articles into different logical categories based on the image processing techniques to make it easier for readers. All the retinal datasets developed until now are also summarized in this paper.

The following are the major contributions:

• Table 1 is added for comparison of all openly accessible retinal image datasets. In the literature, Fraz et al. [6] discussed only eight (8) datasets, while this article has compared thirteen (13) openly accessible datasets.

  Table 1 An overview of freely accessible retinal databases with a semantic segmentation ground truth

• We have classified the retinal vessels segmentation approaches into more logical groups than Fraz et al. [6] for the ease of readers.

• Fraz et al. [6] selected and compared only 69 papers from peer-reviewed publications, while we have compared and evaluated 149 papers from the peer-reviewed journals.

• In the 4th column of Table 3, we have summarized results for medical examination and applications.

• In Sect. 4, we have added Fig. 6 to show frequency of databases, which is frequently used for validation of different retinal vessel extraction algorithms.

This article consists of the following sections: retinal image analysis (Sect. 2), published literature that investigated retinal vessels extraction techniques (Sect. 3), and discussion and the conclusion (Sect. 4).

2 Retinal image analysis

This section discusses the acquisition and abnormalities of retinal images and the datasets used for validation of retinal vessels extraction techniques.

2.1 Acquisition of retinal image

Retinal photographs are captured by a digital fundus camera attached with a low-power microscope. Pupil of the human eye is used as entry/exit points for fundus camera illumination and imaging light beams on the inner surface of the eye called the retina. Visible landmarks of the retina are OD, blood vessels, macula, and fovea. The fundus camera operates in different modes to scrutinize the human eye retina. In color mode, the retina is inspected in full color under the flashing of white light. In red-free mode, a filter is applied to improve the appearance of vessel network to observe vascular disorders within the retina. It is used to examine disorders such as hemorrhages, exudates, nerve fiber layer defects, and epiretinal membranes. In angiography mode, fluorescent dye is vaccinated into the blood stream to capture retina and surrounding tissue blood vessel network for the analysis of various eye diseases. It is helpful for screening abnormal vessels and hyperpermeable vessels leading to ocular tumors and central serous chorioretinopathy, respectively [6]. Figure 1 represents the anatomical structure of the retina, an image taken from the HRF dataset.

Fig. 1

Anatomical structure of the retina

2.2 Retinal image abnormalities

Various crucial diseases that disturb the retinal vascular network appear in the eye, the cerebrum, or the cardiovascular structure. The following is the outline of the most dominating sicknesses, which can be inspected through retina photograph.

2.2.1 Diabetic retinopathy

Diabetes is a severe, fatal disorder that arises either when the pancreas does not generate sufficient insulin (a hormone that controls glucose), or when the body cannot appropriately handle the insulin it generates [15]. Excess of glucose in blood, a common effect of uncontrolled diabetes, may after sometime lead to severe harm to the heart, eyes, kidneys, blood vessels, and nerves and outcomes in a retinal complication of diabetes called DR [16]. Type 1 and type 2 diabetes are two types of diabetes categorized by the lack of insulin and ineffective use of insulin in the body, respectively. Type 1 diabetes is currently inevitable, and its cause is unknown. Indications of type 1 and type 2 diabetes include constant hunger, vision impairments, excessive urination, weight reduction, tiredness, and thirst [17].

DR is a common complication of diabetes and dominant reason for visual impairments in the people of working age of western nations [18]. Analyzing retinal photographs for the existence of DR to a great extent is a pattern recognition task. The common attributes of DR are MAs, exudates and cotton wool spots, etc. The hyperglycemia in the eye harms the retinal vessel walls. Consequently, new tiny vessels are formed which eventually bleed and damage the retina, a condition called proliferative DR. The damage to the retinal vessel barrier causes fluid leakage, diabetic macular edema (DME) and harm to photoreceptors [13].

2.2.2 Glaucoma

Glaucoma arises when the fluid called aqueous humor inside eye does not drain normally. This fluid piles up and increases pressure inside the eye, which harms the eye’s optic nerve that gets worse over time. This can result in visual impairment and blindness [19]. Timely analysis and medication seem to reduce the uncertainty of visual impairment due to glaucoma [2]. Secondary glaucoma is a condition that occurs due to tumor, eye injury, retinal blood vessel blockage, or leakage, swelling of the eye and the use of steroids. The major symptoms of glaucoma are blurred vision, sudden vision loss, and nausea along with extreme eye inflammation. In general, there is no treatment for glaucoma, yet it can be restricted. Eye drops, pills, laser techniques, and surgical operations are utilized to keep or moderate further harm from happening [20].

2.2.3 Cardiovascular disease

Cardiovascular disorder appears in the retina in various ways. Hypertension and atherosclerosis produce variations in the proportion between the width of retinal arteries and veins, famous as the A/V ratio. Narrowing of the arteries and broadening of the veins can decrease A/V ratio and result in an increase in the risk of stroke and myocardial infarction [21]. Moreover, the systemic vascular disorder can lead to arterial/venous occlusions, known as central/branch arterial occlusions (CRAO, BRAO) and central/branch venous occlusions (CRVA, BRVO) [2].

2.2.4 Age-related macular degeneration

AMD is a syndrome that influences individuals beyond 50 years old and is the major source of inevitable vision loss in elderly individuals on the planet [22]. The systemic risk factors consist of male gender, hypertension, hyperlipidemia, chronic kidney disease (CKD), Hepatitis B surface antigen (HBsAg) positivity, liver cancer, coronary heart disease, and increased serum white blood cell levels, etc. [23]. The two noteworthy modes are dry and wet AMD, of which dry AMD ordinarily prompts slow loss of visual acuity. Wet AMD is the most optically dangerous mode, classified by an increase of a choroidal vascular structure into the macula coupled with expanded vascular permeability. The expansion in vascular permeability prompts to unusual fluid collection inside or beneath the retina that result in visual impairments when it associates the middle of the macula [2].

2.3 Openly accessible retinal image datasets

Image datasets possess significant importance in pattern recognition studies, as the developed techniques need to be validated on sample images. An overview of all the freely accessible retinal image datasets is presented in Table 1. The majority of retinal vessel extraction techniques are assessed on two openly available standard datasets, namely DRIVE and STARE.

3 Categorization of retinal image vessel extraction algorithms

This survey classifies retinal vessel extraction techniques based on the image processing approaches applied and the systems utilized. The retinal vessel extraction methods are classified into five core classes: (I) pattern recognition (PR) approaches, (II) vessel tracking/tracing-based methodologies, (III) model-based techniques, (IV) hardware implementation-based approaches, and (V) hybrid approaches. Some of these classifications are further isolated into subsections. A graphical representation of the classification of retinal blood vessel extraction techniques is shown in Fig. 2.

Fig. 2

Classification of retinal vessel segmentation techniques

This review consists of 149 articles considered from peer-reviewed journals. The articles categorized in each of the aforesaid classifications are outlined in Fig. 2 for reference. The categorization has been performed based on the image processing methodology utilized for retinal vessel extraction. The structure of retinal vessels provides important information for ophthalmologists. Sections 3.1–3.5 provide a detailed explanation of the techniques given in Fig. 2 alongside their analytical discussion.

Figure 3 indicates the number of the papers in each subcategory of retinal vessel extraction techniques. The yearly decomposition of selected papers is shown in Fig. 4.

Fig. 3

Classification of papers based on techniques used for retinal vessel extraction

Fig. 4

The yearly distribution of the selected articles

The performance metrics for retinal vessel detection techniques is summarized in Table 2.

Table 2 Performance metrics for retinal vessels extraction

3.1 Pattern recognition approaches

There are PR techniques to handle the identification of patterns and consistency in data. In the vessel segmentation scenario, PR approaches handle the identification of retinal vasculature, vessel attributes, and other background segments. PR algorithms are distributed into two groups: supervised and unsupervised techniques. Supervised approaches exploit some prior marked statistics to isolate vessel and non-vessel pixels, while unsupervised approaches achieve the vessel extraction without any prior marking information. PR unsupervised approaches for retinal vessel extraction are further classified into four groups: (1) multiscale-based techniques, (2) mathematical morphological methods, (3) matched filtering approaches, and (4) region growing methods. Supervised approaches are further categorized into two classes: (1) neural network-based approaches and (2) ensemble classification.

3.1.1 Unsupervised approaches

Multiscale-based techniques

Multiscale techniques are based on the variation of image scales to perform vessel extraction. The fundamental point of interest of this procedure is to increase the execution speed. In multiscale approaches, a scale space is generated from the input image for extracting different structures. Major vessel structures are segmented from low-contrast photographs, while fine vasculatures are detected from high-contrast ones which increased robustness. Multiscale methods are frequently utilized in vessel extraction in order to address the issue of vessel width variation in fundus images. The articles categorized in the aforesaid classification are outlined in Table 3 for reference.

Table 3 Multiscale-based retinal vessel segmentation approaches summarization

Frangi et al. [24] proposed vessel enhancement filtering process based on geometrical attributes. The filtering process searches for geometrical attributes with elongated structure. Vessels contrast enhancement is achieved by computing all eigenvalues of the Hessian matrix of the image. A multiscale approach is used due to variation in vessel width. The vesselness measure can be utilized as a preliminary step for vessel extraction of this type of photographs, which is further used for contrast enhancement and detection of tiny vessels [25]. This algorithm introduced a new LRV test in place of Hessian and MF, used in [26]. The LRV is constructed from a 6-D measurement array of MFR and edge responses and their associated confidences. A multiscale MFR are combined to segment vessels with varying size. The edge responses are helpful in differentiating between offset boundaries near abnormalities and genuine vessels. The confidence measures give priority to the contour of the intensity surface rather than the magnitude of the responses, improving the capacity of the LRV to distinguish low-resolution vessels. The six-component dimension arrays are recorded to compute the resultant LRV utilizing pdf’s learned from training information. Each component is intended to fix the issue of identifying tiny and low-resolution vessels, while preventing responses to other retinal photograph segments. Finally, the new LRV is utilized by vessel tracing scheme to extract a complete network of vessel centerline.

Martínez-Pérez et al. [27] coupled processes of the scale-space representation, first and second derivatives of the intensity image and region growing technique. The scale-space representation provides detail about diameter, elongated structure, and directions of retinal vessels. Two significant statistical attributes of elongated structures depend on the first and second derivative data which provide weights to pixels with a high probability of relating to vessels. The region growing approach permits vessels to grow where the values of the maximum principal curvature lie within a wide interval, and the growth is restricted into segments with low gradients. This permits fast evolution of regions outside the boundaries. Finally, when the margins between clusters are defined, the technique grows vessel and non-vessel clusters at once without the gradient constraint. This technique is further extended using multiscale feature extraction [28]. They investigate the observations that the intensity of the photograph is relational to the volume of blood in the light track corresponding to that pixel.

The MSLO methodology is presented to segment digital fundus images and suppress noise [29]. This was pursued by thresholding to obtain an initial vessel extraction map that was further refined by utilizing a simple RG method. Results with the application of simple median filter are obtained for comparison with MSLO.

Vlachos et al. [30] presented a novel MSLTA and morphological postprocessing steps for extraction of retinal vascular structure. An optimizing process begins with the approximation of a confidence matrix from a cluster of seeds which is obtained from the image’s histogram and proceeds to a particular state of the cross-sectional profile gets to be invalid. The multiscale photograph sketch is estimated after coupling the individual photograph sketches along scales, holding the confidence for each pixel to relate to a vessel. The combination of multiscale line tracking results produces a multiscale confidence image map. The preliminary vessel tree is extracted after map quantization of the multiscale confidence map. Map quantization and median filtering procedures remove noisy lines and connection breaches growing the accuracy of the confidence array. Finally, morphological postprocessing eliminates the residual artifacts.

Li et al. [31] proposed multiscale vessel detection approach by convolving the responses of MF at three different scales, which further enhances the vessels while suppressing the noise. MF exhibit strong responses at the vessels on multiple scales, while the background noise does not show higher response and can be easily suppressed. Further, good width estimation can be easily achieved by suitable selection of scale parameters. The method is suitable for small and tiny vessels detection, concurrent extraction of vessels of variant widths, low computational complexity and easier to implement.

A multiscale weighted 2-D medialness function-based technique has been introduced for extraction of retinal vessels in [32]. This approach consists of three noteworthy steps: vessel medialness detection, vessel midline recognition, and vessel reconstruction. The outcomes of the medialness function are first convolved by the eigenvalues of the Hessian matrix of the photograph. Next, midlines of vessels are detected utilizing noise elimination and reconnection operations [25]. In conclusion, vessel diameter is computed and retinal vessels are extracted.

Yu et al. [33] introduced a multiscale enhancement and second-order local entropy-based technique for the extraction of retinal vessel tree. After preprocessing, vessel network is produced by calculating the eigenvalues of the second derivatives of Gaussian filtered photograph at different scales [24]. The second-order local entropy thresholding is utilized to extract the vessel tree. Finally, a rule-based decision step is used to compute the geometric shape variance between vessels and background to decrease FPR.

An automatic retinal vessel extraction framework based on multiscale line identification [34] was presented by Nguyen et al. [35]. This approach depends on the variation of the length of a basic line detector in order to achieve multiscale line detectors. Line responses at different scales are linearly coupled to eliminate the drawbacks of each individual line detector and to obtain a final binary image.

A novel multiscale vessel enhancement and segmentation technique using a complex continuous wavelet transform (CCWT) is proposed by Fathi et al. [36]. The CCWT differentiates the elongated elements from the boundary structures utilizing real and imaginary parts of the applied wavelet. The resultant binary image is segmented by integrating the modulus values of the wavelet coefficients in the appropriate scales and utilizing an adaptive thresholding technique along with suitable length filtering.

Azzopardi and Petkov presented a trainable COSFIRE filters for detection of vascular junctions in the DRIVE dataset retinal photographs [37]. The setup of a COSFIRE filter relies upon the spatial procedure of contour measures that exists along concentric circles of given radii around a stated ROI. Gabor filters are used for the identification of elongated structures and boundaries. The weighted geometric mean is utilized to combine responses of filters.

Akram and Shoab introduced a novel multilayered thresholding approach for detection of retinal vascular network [38]. The vessels contrast is boosted by utilizing 2-D Gabor filter. The vessel extraction mask is produced utilizing multilayered thresholding approach which traces vessel boundaries and inspect their connectivity by using various thresholds iteratively. Moreover, it removes all false boundaries and vessel segments prior to vessel tracking.

To enhance the vessels, MFMK algorithm is used [39]. A multiscale hierarchical decomposition is utilized on the normalized enhanced photograph to remove the noise and localize vessels. The final binary vessel tree is extracted by using local adaptive thresholding.

Ganjee et al. [40] proposed an enhanced retinal vessel extraction technique depend on high level features for abnormal photographs. The initial vessel network is detected based on low level features using multiscale MFR [41]. The final binary vessel map is obtained based on high level features.

A new adaptive thresholding method that depends on local homogeneity information for extraction of the retinal vascular network is presented in [42]. Image enhancement and sharpness is performed by using the Unsharp filter, which is pursued by an average filter for smoothening the image. The vessels are localized by utilizing adaptive thresholding depend on local homogeneity information. A combination of morphological postprocessing steps is used on the inverted thresholded photograph to remove the remaining misclassifications.

Ravichandran and Raja [43] proposed a local entropy-based thresholding approach for identification of retinal vessel network. The retinal image is enhanced by applying CLAHE followed by wiener filtering to eliminate background noise. The 2-D Gabor filter is applied for further vessel enhancement. Finally, thresholding is applied to obtain a binary map.

A novel unsupervised retinal vasculature extraction technique that contains inpainting filter, called neighborhood estimator before filling, is exhibited to inpaint exudates in a way that adjacent FPR are considerably decreased during vessel enhancement [44]. The preprocessing steps are adapted from [45]. Multiscale Hessian eigenvalue approach is used for retinal vascular enhancement. Finally, percentile-based thresholding is used to obtain binary image.

Fraz et al. [46] designed a novel QUARTZ approach depending on multiscale line detector for handling and evaluating of retinal photographs. It contains sections for vessel extraction, diameter computation, and angular variation of the individual vessel midline pixel with sub-pixel accuracy, calculating local vessel direction, OD localization, arteriole/venule categorization, tortuosity computation, and exporting the numerical estimations in different output file formats.

A novel approach based on cake filter for extraction of the retinal vascular network is presented in [47]. This technique first joins the real element of orientation score utilizing the cake filter and joins with the adaptive threshold value to get retinal vessel map.

The curvelet transform and kernel-based FCM framework is suggested by Kar et al. [48] for retinal vessel segmentation and optic disk removal. Retinal vasculature enhancement and denoising is performed by applying curvelet transform. The kernel-based FCM approach is used to localize vessel structure. For pathological images, LoG filter along with matched filtering is also applied to differentiate the step and the ramp-like signal from that of vessel structure.

A new framework for multiobjective retinal vessel classification is a combination of FPSA and PFCM [49]. The FPSA is applied to localize the retinal vessel network utilizing PFCM fitness function. In the second level of optimization, the attained cluster centers are much enhanced utilizing PS as local search, but now the goal is to find the tinny or vessels with small widths.

Al Shehhi et al. [50] suggested a graph-based method for extraction of retinal vasculature. Preprocessing steps are used to enhance contrast and to create essential features to enhance vessel structure due to the sensitivity of vessel patterns to multiscale/multiorientation structure. Graph-based segmentation is used to reduce computational processing required for region of interest into most semantic objects.

Khan et al. [51] introduced morphological Hessian-based approach for retinal vasculature extraction utilizing region-based Otsu thresholding. The preliminary process consists of CLAHE and morphological operations for contrast enrichment and low frequency noise elimination, correspondingly [52]. The Hessian matrix and eigenvalues are utilized in a novel way utilizing multi scales to excerpt wide and thin vessel-enhanced photographs, separately. The modified form of Otsu thresholding is employed to segment vessel and background pixels from both enhanced photographs. In conclusion, length filtering is utilized to remove the undesirable region/segment, background pixels, disease aberrations, and noise, to acquire a final segmented photograph.

Table 4 shows the performance metrics stated by different multiscale techniques of retinal vessel extraction, with a high accuracy stated by Moghimirad [32] both on the DRIVE and STARE datasets.

Table 4 Performance analysis of multiscale techniques

Mathematical morphological (MM) approaches

The basic idea of morphological approaches is to probe a photograph with a small segment or template known as a structuring element (SE). The SE is placed at all probable positions in the photograph, and it is contrasted with the respective local pixels. A morphological filter will store shapes like its SE shape while modifying different shapes. A morphological filter will store shapes similar to its SE shape while modifying dissimilar shapes. Most morphological filters are successful at eliminating both linear and nonlinear noise forms. The MM-based techniques are summarized in Table 5.

Table 5 Summarization of Mathematical Morphological (MM) based retinal vessel segmentation methods

Zana and Klein [53] proposed MM- and curvature evaluation-based scheme for extraction of vessel-like structure. Morphological operations are used to prominent vessels with respect to their morphological attributes, and then, cross-curvature evaluation is utilized to distinguish vessel and background pixels. Ayala et al. [54] presented an average of fuzzy sets by utilizing the average distance of Baddeley–Molchanov and the mean of Vorob’ev. The coupling of Zana’s method and Baddeley–Molchanov average produced good results. The vessel extraction operations suggested by Zana and Klien have been reconsidered and revised employing these new averages. All these operations generate grayscale photographs with improved vessels contrast after low level processing. A binary image is obtained by applying a threshold to this enhanced image. The authors preserve the image acquired before utilizing the final threshold which is assumed as a fuzzy set demonstrating the vessels.

Pixel processing-based technique is presented for retinal vessel segmentation in [52]. This approach consists of three stages: preprocessing stage, vessel midline identification, and vessel extraction. In preprocessing stage, background is normalized and thin vessels are enhanced by utilizing a bank of line detection filters at different orientations. In second stage, a bank of DoOG filters is used for extraction of vessel centerline followed by an RG process for reconnection of vessel centerline. In the last stage, a modified top hat transform with different SEs is utilized to enhance vessels of variable widths, followed by morphological reconstruction to extract binary vessel map. A final binary photograph is acquired by an iterative seeded RG operation of the centerline photographs with the arrangement of four binary networks.

Yang et al. [55] presented a hybrid technique that consists of combination of MM and a fuzzy clustering technique pursued by a purification process. Morphological top hat operation is used for vessels enhancement and noise suppression. The binary vessel map is extracted by fuzzy clustering.

Fraz et al. [56] introduced a differential filtering and morphological processing approach for detection of retinal vessel. The FoDoG is used for detection of midlines. A multidirectional morphological top hat operation is performed to acquire the structure and direction map of the vessel along with bit plane slicing of a vessel-enhanced grayscale photograph. The final binary photograph is obtained by combination of midlines with these maps.

The FDCT and multistructure MM is applied for vessel extraction [57]. The vessel contrast is enhanced by using FDCT, and the boundaries of vessels are identified by utilizing a multistructure morphological operation. The wrongly detected boundaries are eliminated by morphological procedure. Connected components analysis (CCA) is coupled with simple thresholding to obtain a final binary image.

Rossant et al. [58] proposed a morphological approach for quantitative analysis and extraction of retinal images. A morphological filter is used for vessel enhancement, and top hat transformation is utilized to extract two attribute images with prominent vessel contrast. Path opening is used for each attribute image so that the main elongated attributes are preserved, whereas the other ones are suppressed. The final segmentation photograph is achieved by thresholding and fusing the two resultant photographs.

The process of vessel midlines identification and morphological bit plane slicing are coupled by Fraz et al. [59] to segment retinal vessels. The FoDoG filter is utilized for extraction of centerlines in four directions, and then, analysis of derivative signs and derivative values are averaged out. The structure and direction map of vessels is acquired by using a multidirectional morphological top hat process pursued by bit plane slicing of the vessel-enhanced grayscale photograph. The midlines are coupled with these maps to attain the final binary image.

The modified form of [59] is presented for detection of retinal vessels in [60]. The DoOG filter is employed for vessel centerlines identification, and top hat transformation is used for vessel enhancement. A retinal vessel binary image is produced by a bit plane slicing. An iterative RG algorithm fuses the main skeleton and the photographs acquired from bit plane slicing of vessel direction-dependent morphological filters.

Xu et al. [61] designed a system for detection of retinal vessels that consists of two stages: shape-based morphology and a simplification of current tree-based connected processes. The first stage is used to extract retinal vessels. The second stage is an extension of the constrained connectivity framework to non-increasing constraints.

Automatic retinal vascular segmentation method based on directional morphological operations and fuzzy categorization is presented by Sigurðsson et al. [62]. The extraction of vessel attributes is achieved by applying path openings filter. The resulting attributes are utilized to perform a data fusion task depending on fuzzy set theory. Therefore, pixel categorization can be simply achieved to make a vessel tree.

Imani et al. [63] introduced a combination of MCA and MWT for extraction of retinal vascular network. This framework supposes that individual signal is a linear mixture of many morphological significant elements. The MCA technique with suitable transforms is used to isolate vessel and non-vessel pixels. The MWT is utilized for retinal vessel enhancement. The final binary image is extracted by adaptive thresholding.

The performance evaluation of the methods based on the MF is shown in Table 6, where the maximum accuracy is shown by Imani et al. [63] for both the DRIVE and STARE databases utilizing MCA technique.

Table 6 Performance metrics for Mathematical Morphology based methodologies

Matched filtering approaches

MF convolves a 2-D template with the retinal photograph for extraction of the vascular network. The template is created to demonstrate an element in the photograph at some anonymous location and direction, and the MFR demonstrates the existence of the element. In order to create the template, the following vessel attributes are considered: vessel elongated structure, bifurcation and branching, crossover points and the vessel width variation. The large size and the utilization of the convolution kernel at various rotations cause a computational complexity. Moreover, the kernel reacts ideally to vessels when the vessel profile matches to the kernel. The retinal background disparity and existence of abnormality in the retinal photograph also upturn the FPR because the abnormality can display the same neighborhood elements as the vessels. A MFR methodology is discovered successful when utilized as a part of any other processing methods. The matched filtering-based algorithms are encapsulated in Table 7.

Table 7 Summarization of vessel tracking methodologies

Chaudhuri et al. [64] extracted retinal vessel tree using a feature identification operator depending on the optical and spatial attributes of vessels. A Gaussian profile is utilized to estimate the shape of the cross section of a vein. The MF is applied to extract piecewise linear regions of retinal vessels. Twelve diverse kernels in all possible orientations are employed to explore the vessel segments. The template is rotated in 15 degree additions to fit into vessels of various directions. The highest MFR is chosen for individual pixel and is thresholded to obtain a resultant vessel map. Finally, postprocessing steps are used to trim and classify the vessel regions.

The approach depends on a MF technique combined with a preliminary information regarding retinal vessel attributes to extract the vessel edges, trace the centerline of the vessel, and detect suitable characteristics of medical concern [65]. The computational efficiency in straight vessel clusters is achieved by an adaptive densitometric tracing approach depending on local neighborhood information.

Al-Rawi et al. [66] modified Chaudhuri et al. [64] approach by utilizing a comprehensive search optimization process on the DRIVE dataset to search an optimal parameters for MF size, the standard deviation and threshold limit. This system beats the Chaudhuri et al. framework. Zhang et al. [67] also improved the Chaudhuri et al. [64] MF for retinal vessel segmentation that uses a local vessel cross section evaluation utilizing double-sided thresholding to decrease FPR to nonlinear boundaries.

Hoover et al. [26] designed a technique for the identification of retinal vessels based on local and region-based attributes of vessels utilizing a threshold probing method on an MFR photograph. The approach investigates the MFR photograph and uses thresholding with iterative probing for individual pixel as vessel or background [64]. Pixels that are not categorized as vessels from probes are reprocessed for further probing. The technique is validated utilizing ground truth photographs and evaluated against simple thresholding of the MFR. This approach achieved 15 times decrement of FPR over the simple MFR and up to 75% increment of TPR has been noticed.

The amplitude-modified second-order Gaussian filter is introduced by Gang et al. [68] for retinal vessel extraction. It evidences that the vessel diameter can be approximated in a linear connection with the spreading factor of the matched Gaussian filter when the magnitude coefficient of the Gaussian filter is appropriately allocated. Not only the vessel diameter computation gives the width of retinal vessel, but it is also beneficial for enhancing the MF to increase the TPR. An adaptive local thresholding technique based on a verification-based multi threshold probing approach is designed for vessel extraction [69].

Sukkaew et al. [70] described an automatic process for retinal vessel detection in low-contrast and noisy retinal photographs of premature infants that uses a statistically optimized LoG edge detection filter, Otsu thresholding, medial axis transform skeletonization followed by pruning, and edge thinning for vessel extraction.

Cinsdikici and Aydin [71] proposed a combination of MF and ANT colony technique for retinal vessel extraction. The MF and ANT algorithms are applied in parallel on the preprocessed photograph. The length filtering is used on the combined results to extract the complete vessel map. This approach did not produce fruitful results on the pathological images. The methodology is evaluated on the DRIVE dataset.

The MF reacts not only to vessels but also to background boundaries which will result in an increase of FPR. A novel MF with FDOG is presented by Zhang et al. [41] to address this issue. The combination of the zero-mean Gaussian MF and the FDOG is applied to extract the vessels. The application of different scales make it possible to segment both wide and small vessels, followed by a logical OR operation to fuse the outcomes in an effective manner. The MF-FDOG considerably decreases the FPR created by the MF and identifies many fine vessels overlooked by MF. The MF-FDOG has also the advantage of extraction the vessels in abnormal photographs.

Amin and Yan [72] presented a high-speed retinal vessel extraction technique based on phase congruency which is a soft categorization of vessels and constant to contrast variations. The phase congruency of a photograph is computed by utilizing a set of log-Gabor wavelets, and a final image is acquired by thresholding.

Kaba et al. [73] presented a combination of bias correction, MF and the EM retinal vessel extraction technique. The STARE dataset is utilized for evaluation of normal and abnormal retinal images. A new self-adaptive MF utilizing a nonlinear synergistic arrangement of the vesselness filter and the MF for the extraction of retinal vessel network are proposed in [74].

Zhang et al. [75] described an automatic retinal vessel extraction technique based on new MF using SoGD in so-called orientation scores. By locally matching the multiscale SOGD filters with data in orientation scores, they are able to improve elongated structures positioned in various direction planes accordingly. Both junctions and thin vessels are well-maintained because of the suggested multiscale and multiorientation filtering technique.

A novel MF using Gumbel PDF as its kernel is designed by Singh and Srivastava [76]. The image contrast enchantment is performed by using CLAHE. The MFR image is used as an input to entropy-based optimal thresholding. Length filtering is further used for removing artifacts to generate a final binary image.

Table 8 encapsulates the consequences of different morphological-based retinal vessel extraction methods. A highest accuracy is stated by Al-Rawi et al. [66] and Zhang et al. [75], when tested on the DRIVE and STARE datasets, respectively.

Table 8 Performance metrics evaluation of techniques based on matched filtering

Region growing methods

The basic theory of the conventional growth region is to collect pixels that have similar characteristics together to form a region. Its efficiency is based on the location of seed points and growth conditions. On the other hand, the traditional technique requires to select seed points manually in order to guarantee the stability. Because the blood vessels have a very wide gray-level distribution, making the traditional region growing method and binary extraction technique hard to precisely segment the vascular map, for the region growing conditions also, depends on range of the image gray level. If each seed point uses the same threshold value, the growth is easy to stop when the blood vessels become more slender. These category methods are summarized in Table 9.

Table 9 Region growing-based retinal vessel segmentation techniques summarization

Ahmad Fadzil et al. [77] proposed a region growing approach using FoDoG to extract retinal blood vessel network. The retinal vascular structure is extracted by using a number of processes containing mean filter, CLAHE, bottom-hat morphological operation, contrast stretching. A seed-based region growing (SRG) and gradient-based region growing (GRG) are applied for vessel reconstruction. The DRIVE dataset is used for the evaluation of this methodology.

A parallel application for the vasculature extraction of high-resolution retinal photographs is introduced by Palomera-Pérez et al. [78]. This approach depends on a data splitting, which permitted a faster handling of those photographs. Two approaches for data partitioning are presented and evaluated: a horizontal partitioning for attribute extraction, and a mixed (horizontal and vertical) partitioning for region growing. ITK serial version is used for extraction high-resolution retinal images.

Jiang et al. [79] suggested a region growing vessel extraction approach depends on spectrum information. This scheme used Fourier transform on the ROI consisting pf vessel information to acquire its spectrum knowledge, according to which its initial feature direction will be detected. Then, coupled boundary knowledge with initial feature direction records the vessel network middle points as the seed points of RG extraction. Finally, the advanced RG approach with branch-based growth scheme is utilized to extract the vessels. The framework is evaluated on different medical images along with standard retinal datasets the DRIVE and the STARE.

Zhao et al. [80] introduced retinal vasculature extraction depending on level set and RG process. The CLAHE and 2-D Gabor filters are used to improve the blood vessels contrast, and an anisotropic diffusion filter is utilized to eradicate the uneven illumination in the photograph and maintain vessel edges. At the end, the RG approach and a region-based ACM with a level set operation are utilized to identify retinal vessels, and their outcomes are superimposed to obtain the resultant binary image.

A structure-based level set technique with an automatic seed point nomination for detection of retinal vascular map is presented in [81]. Additionally, this approach presents an improved zero-level contour regularization term which is more suitable than the ones presented by other approaches for retinal vessel extraction. The model is tested on their own dataset [82], as well as openly accessible the DRIVE and the STARE datasets.

An automatic seed identification based on a novel BHS measure and a new EDSRG technique for extraction of the retinal vascular network are proposed by Panda et al. [83]. The BHS approach directly computes a binary symmetry decision at every pixel without the measurement of continuous symmetry map and image thresholding. This framework contains four distinct steps: preprocessing, selection of seed (midline) pixels utilizing the BHS approach, postprocessing to remove false midline pixels, and vessel extraction utilizing EDSRG algorithm. The disadvantage of this approach is dependency on the edge.

Lázár and Hajdu introduced a unique RG technique which utilizes a hysteresis thresholding approach with the response vector similarity of contiguous pixels [84]. The performance comparison of different RG algorithms is summarized in Table 10. A maximum accuracy is reported on the DRIVE dataset by Panda et al. [83] and on the STARE dataset by Zhao et al. [80].

Table 10 Performance metrics for comparison of region growing algorithms

3.1.2 Supervised classification

Supervised classification needs hand-marked best quality level photographs for training, and each pixel is depicted by a feature classifier based on neighborhood or global data of the photograph. The constraint for this approach is that an arrangement of main distinctive attributes must be identified for training and characterization techniques. This retinal vascular map is accurately marked by an eye care experts. However, as mentioned in [26], there is noteworthy dissimilarity in the extraction of vessels even among professional graders. As supervised techniques are modeled depending on prior information, their efficiency is generally higher than that of unsupervised ones and can show superior consequences for normal fundus photographs. The supervised classification methods are listed in Table 11.

Table 11 Encapsulation of supervised retinal vessel segmentation techniques

Niemeijer et al. [10] excerpt a feature array for every pixel that contains of the green channel of the color photograph and the reactions of a Gaussian MF and its first- and second-order derivatives at multiple scales of pixels. Subsequently, the kNN technique is used to compute the probability of the pixel relating to a vessel [85]. The resultant vessel map is acquired by thresholding the probability map.

The 2-D Gabor wavelet technique for extraction of retinal vessel network is suggested by Soares et al. [86]. It utilizes a feature array made of the pixel’s magnitude, and 2-D Gabor wavelet transform responses are captured at various scales. A GMM classifier is utilized to isolate every pixel as either a foreground or background pixel.

Ricci and Perfetti [34] proposed segmentation of retinal vascular map by utilizing line operators and SVM. A line detector is used on the green plane of a color photograph, and the response is thresholded to acquire unsupervised pixel sorting. Additionally, two orthogonal line detectors are also used along with the gray level of the target pixel to extract a feature array for supervised sorting utilizing an SVM. Comparing with other supervised approaches, the methodology (1) needs less features, (2) feature localization is computationally simpler, and (3) less samples are required for training.

Anzalone et al. [87] designed a modular system for the identification of retinal vessels. The vessel enhancement scheme is adopted from [28]. This method consists of two core blocks: The first block is responsible for contrast enhancement by applying CLAHE, while the second block performs thresholding and some cleaning steps to extract the binary image.

Osareh and Shadgar [88] utilized Gabor filters at different scales for vessel extraction, and the attributes are detected utilizing PCA. The elements for Gabor filters are excellently adjusted by conducting tests. The image components are categorized as vessels and background utilizing the corresponding feature arrays by the GMM and SVM classifiers.

For vessels detection, an arrangement of different image processing methods with SVM classifier is presented by Xu and Luo [89]. After preprocessing, the input photograph is further handled by wavelets at different scales for feature abstraction. The line detectors are utilized to detect tiny vessels. A 12-D feature array for individual residual pixel in the binary retinal photograph excluding large vessels is generated. SVM is utilized to differentiate thin vessel regions from the background. A tracing algorithm depends on the coupling of vessel orientation, and the eigenvector of the Hessian matrix is utilized for thin vessel growth to acquire a vessel map.

FABC-based approach is proposed for automatic detection of retinal vascular network [90]. They used a 41-D feature array utilizing statistics of the local intensity arrangement, spatial attributes, and geometry at different measures. Categorization depends on the FABC trained on ground truth patterns of vessel and background pixels. This methodology is validated on the DRIVE dataset.

You et al. [91] introduced an arrangement of the radial projection and the semi-supervised self-training technique utilizing SVM for vessel extraction. The radial projections are used for the localization of vessel midlines, the tiny and low-contrast vessels. The vessel contrast is enhanced by an improved steerable complex wavelet. The SVM classifier is utilized to detect the main attribute of vessels. The extracted vascular map is acquired by the combination of the two. The methodology is very efficient in extracting tinny and poor resolution vessels but prone to inaccuracies in case of abnormalities.

A new supervised technique to investigate the performance of Leung–Malik filters in classifying retinal vessels is presented by Varnousfaderani et al. [92]. It consists of two-level hierarchical learning framework to extract vessels in abnormal fundus photographs. The retinal disorders, vessels, and non-vessels are demonstrated by expert classifiers in the first level. The results of the expert classifiers are combined to identify vessels in the second level. Roychowdhury et al. [93] suggested a novel three-step retinal vessel extraction algorithm. In the first step, the green channel of a retinal photograph is preprocessed to segment a binary photograph after high-pass filtering, and other binary photograph from the morphologically reconstructed enhanced image for the vessel regions. Subsequently, the mutual segments of both binary photographs are considered as the major vessels. In the second step, all residual pixels in the two binary photographs are categorized by utilizing a GMM classifier. In the third postprocessing step, the major segments of the retinal vessels are coupled with the categorized vessel pixels.

Orlando et al. [94] introduced a discriminatively trained FC-CRF model for extraction of retinal vessels. Attributes of this approach are trained without any human intervention by utilizing a structured output SVM. This methodology is capable of reconstructing the retinal vessel network very accurately than other conventional CRF by using features detected from the images and fully connected pairwise potentials.

The performance comparison of supervised techniques of retinal vessels is demonstrated in Table 12, where an assessment according to accuracy, the Lupaşcu et al. [90] is exposed to beat the other compared techniques on the DRIVE dataset, while Ricci and Perfetti [34] achieved maximum accuracy on the STARE dataset.

Table 12 Performance comparison for supervised approaches

Neural network approaches

The benefit of utilizing CNN-based approaches is that it reduces the computational cost due to massively parallel processing. It has also appropriate for hardware application on a chip-set architecture based on the CNN Universal Machine (CNNUM) paradigm. Table 13 shows summary of the neural network-based methods.

Table 13 Summarization of neural network-based retinal vessel segmentation methods

ANNs have been widely inspected for extracting fundus images attributes such as the vessel structure and perform categorizations which depend on geometric probabilities rather than objective reasoning [95]. Sinthanayothin et al. [96] proposed MLP-NN for extraction of retinal vessels depends on the inputs from a PCA of the photograph and boundary recognition of the intensity. The results of this approach are compared with the manually labeled retinal components by an expert ophthalmologist.

A back propagation technique for the extraction of vessels in angiography is presented by Nekovei and Ying [97]. This approach uses NN directly to pixels without prior attributes extraction. The pixels of the small sub-window, which slides across the angiogram photograph, are directly fed as input to the network. The gold standard photographs of hand-marked angiograms are utilized as the training set to set the network’s weights. An improved form of the common delta-rule is utilized to acquire these weights.

Yao and Chen [98] introduced a PCNN and fast 2-D Otsu approach for detection of retinal vascular network. The binary vessel map is acquired via evaluating region-based attributes. The STARE dataset is used for performance evaluation of this method.

Lupaşcu and Tegolo [99] proposed an unsupervised neural network self-organizing map (SOM) and K-means clustering approach for retinal vasculature extraction. AdaBoost classifier is used to divide map units into two groups [90]. SOM and K-means clustering produces good results on a small number of classes and is also very fast. This framework is tested on the DRIVE dataset.

A new supervised scheme based on neural network (NN) is proposed by Marin et al. [45] for extraction of retinal vasculature. This technique utilizes a 7-D feature array that depends on gray-level and moment invariant-based attributes. A multilayer feed forward NN is used for training and categorization. The NN is trained on the DRIVE dataset only, but it shows robustness with various image settings and on different image datasets.

A supervised framework based on different settings of NN known as LNNDP is suggested by Vega et al. [100]. Preprocessing, feature extraction, classification, postprocessing are the major steps of this approach. The LNNDP approach is different from the conventional NN approaches in the computation executed by the single neuron. The STARE data are used for validation of this methodology. Vega el al. [101] proposed another adapted version of the hyperboxes partition technique [102] for detection of retinal vasculature. The LNNDP an advance version of [100] is used for extraction of retinal vessels.

Andersson et al. [103] introduced modified gradient descent (MGD) search for level set-based detection of retinal blood vessels. This approach used two MGD approaches: One utilizes a momentum term, and others depends on resilient propagation (Rprop). The MGD approaches are utilized to train the learning systems based on ANNs. The framework is validated on the real and synthetic data. The DRIVE dataset is used for retinal vessel extraction.

Anitha and Hemanth [104] suggested an EKFN-based automatic technique retinal detection of retinal vascular map. This automatic approach also contains hybrid feature detection methods which considerably enhance the competence of the proposed NN. This system is tested on their own collected datasets.

The ANN-based approach is proposed by Franklin and Rajan [105] to detect retinal vasculature. The MLP-NN is used for identification of retinal blood vessels, for which the inputs are obtained from Gabor and moment invariants-based features. The disadvantage of this approach is missing of tiny blood vessels in some cases. The images from the DRIVE dataset are used for confirmation and evaluation of this methodology.

A novel supervised technique for retinal vasculature extraction is redesigned as a cross-modality data transformation problem [106]. A wide and deep NN with robust training aptitude is presented to model the transformation, and an efficient training tactic is used. The segmentation process does not require artificially designed features, reducing the influence of particular features. This approach is also robust to noisy and pathological images.

Sironi et al. [107] introduced a learning-based filtering system used for detection of curvilinear structures along with two learning-based approaches for acquiring separable filter banks. The first one directly learns separable filters by varying the regular objective function. The second one acquires a basis of separable filters to estimate an existing filter bank, and not only acquires the equivalent performance of the original, but also significantly decreases the quantity of filters, and thus convolutions, needed. The performance is evaluated on the DRIVE dataset. The comparison of different techniques based on neural network is shown in Table 14.

Table 14 Performance metrics for neural network-based methodologies

Ceylan and Yaşar [108] proposed a complex ripplet-I transform and complex valued ANN for detection of retinal vascular map. The ripplet-I transform estimates the features matrix from the retinal images of standard datasets, which is further used as an input to the complex valued ANN. The inverse transformation is applied to resize the output coefficients of ANN.

A novel supervised DNN-based approach is presented for extraction of retinal vascular map [109]. The DNN is trained on a large (up to 400,000) sample of examples preprocessed with global contrast normalization, zero-phase whitening, and augmented using geometric transformations and gamma corrections.

Table 14 demonstrates the performance evaluation for the methodologies based on neural network. The Ceylan and Yaşar [108] reported the highest accuracy on the both standard datasets: DRIVE and STARE.

Ensemble classification approaches

Ensemble classification is the procedure by which various classifiers are strategically created and pooled to address a specific machine learning task. The ensemble classification approaches are listed in Table 15.

Table 15 Summarization of ensemble classification-based retinal vessel segmentation methods

Fraz et al. [110] suggested a supervised classification approach utilizing an ensemble classifier of bagged decision trees for extraction of retinal vessels. They utilized decision trees as the categorization model and the outputs of these frail classifiers are pooled by means of bootstrap aggregation also known as Bagging. They utilized 8-D feature array containing the outputs from a filter set comprising the filter templates of dual-Gaussian and Gabor functions and the line strength computation, which deals the central vessel reflex successfully. This approach is further modified utilizing 9-D feature vector for detection of retinal vascular map [111]. The feature vector contains the vessel tree acquired from the orientation investigation of the gradient vector field (GVF), the morphological procedure, line strength computation and the Gabor filter response which translates data to deal with the both healthy and unhealthy fundus photographs successfully. This concept is further utilized by applying 13-D feature vector containing of the responses from a filters set comprising the filter templates of dual-Gaussian, second-order derivative of Gaussian and Gabor functions, along with the line strength measures and morphological procedure, which deals with the central vessel reflex successfully [112]. The CHASE_DB1 database is utilized for the performance validation of the proposed framework.

A feature and ensemble learning-based technique utilizing a combination of two classifier CNN and RF is proposed [113]. CNN works as a trainable hierarchical feature extractor, and ensemble RFs work as a trainable classifier.

Two different state-of-the-art classifiers SVM and ensemble are explored by Welikala et al. [114]. Both delivered remarkable consequences; however, the SVM classifier is chosen as it marginally left behind the ensemble classifier of bagged decision trees.

Zhu et al. [115] proposed an ensemble approach utilizing 36-D feature vector, consisting of local attributes, morphological procedure with multiscale and multiorientation and DVFs. Then, weak classifiers are trained by the classification and regression tree (CART) utilizing information of a feature array. Lastly, an FABC is created by iteratively training for the retinal vessel extraction.

The performance metrics used for analysis of the efficiency of ensemble classification-based approaches of retinal vessels segmentation are shown in Table 16, where Wang et al. [113] achieved a highest accuracy on the DRIVE and STARE datasets.

Table 16 Performance comparison of the ensemble classification methods

3.2 Vessel tracking techniques

The tracking methodologies track the vessels based on manual/automatic selection of some seed positions and subsequently the vessel midline detection steered by local data. The benefit of tracing-based techniques is its competence since only pixels near to the preliminary points are inspected and analyzed. Additionally, significant statistics (i.e., vessel width and bifurcations) are usually detected along with the vessel tree. However, a disadvantage of such techniques is that refined approaches have been introduced due to the complex intensity profile at branching or crossover positions. Since vessel bifurcation or boundary locations are not well modeled, this process often tends to end at these positions and this leads to incompletion in the extraction result. Table 17 shows the summary of all vessel tracking based frameworks.

Table 17 Summarization of matched filtering-based retinal vessel segmentation methods

Recursive tracking-based vasculature extraction in retinal angiograms is presented in [116]. Initially, choosing the starting point and direction, a section within vessel map is identified. After identification of segment, it is discarded in the angiogram photograph. The recognition-elimination plan is utilized to overcome the issue of tracking-path re-entry in those zones where vessels crossover happens. This operation is executed repeatedly to excerpt the retinal vessel network. Human intervention is involved in this approach to identify vessel seed points. An automatic vessel tracking technique is proposed by Liang et al. [65] to detect vessel segment and tortuosity. The method depends on a MF technique along with a preliminary information about retinal vessel attributes to automatically extract the vessel edges, centerline, and other helpful information of clinical interest. However, this approach requires human interference for selection of direction, start and end points.

Chutatape et al. [117] utilizes a combination of Gaussian and Kalman filters for retinal vessel extraction. The second-order Gaussian MF is used to detect the vessel midline center point, and then, the tracing operation is performed, initializing from the boundary of the OD. The Kalman filter is utilized to detect the subsequent vessel segment position utilizing all the former and the current segment attributes.

Tolias and Panas [118] proposed a retinal vessel tracking framework based on the FCM clustering approach. The remarkable characteristics of this approach are that it does not use any boundary data to trace the precise position of the vessels and in result suppressed the noise. Also, the method utilizes only fuzzy image intensity data. Additionally, there is no setting of parameters and initialization required. The technique tracked the major vessels in the image very well and overlooked only tiny and low-contrast vessels.

Ali et al. [119] presented the recursive tracking depend on recursive tracking of the vessels using directional templates. This approach improves the previous work in different directions: (1) No human intervention is required, only few parametric tuning; (2) it is robust to image contrast variations, major and tiny vessels and artifact elimination, and (3) it does not need the vasculature to be linked. The technique is tested on normal and abnormal dilated retinal images.

Non-recursive paired tracking method is presented to trace the vessels in low-contrast retinal images utilizing their boundary network as calculated by the canny edge detector [120]. Tracing proceeds by subsequent edge line while observing the connectivity of its twin edge. Seed generation permits the approach to handle junctions (a main problem in vessel tracking) and jump over damaged edges. Five typical retinal images are used for the algorithm evaluation.

Quek and Kirbas [121] introduced retinal vascular map detection approach based on a wave propagation and trace-back scheme. Every vessel pixel is marked on the photograph with the likelihood by applying a dual-sigmoidal filter. The refractive index photograph is created by the complement of this image, and then, a digital wave is circulated through the photograph from the base of the vascular map. The vascular map is acquired by tracking the wave along the neighborhood perpendicular to the waveform. This methodology allows the location of the particular vessels, as well as the vascular connecting morphology also.

A model-based tracking method for vessel extraction and width computation is proposed in [122]. The approach uses a parametric model of a vessel consisted of a ‘stripe’ which utilizes statistical attributes for parameter descriptions. A measure of match (MoM) computes the resemblance between the model and the input photograph. The starting of seed pixels for vessel tracing is performed utilizing a multiscale filter and distributing the binary output in non-overlapping square chunks and selecting a random nonzero pixel as a seed [24]. This approach does not involve any human interference. The vessel width is additionally recuperated with the characterized model using the strip width parameter in this way assuming direct reliance between vessel width and model width parameter. The DRIVE dataset as a training set is used for tuning of parameters.

Xu et al. [123] designed graph-based technique for tracking of vessel edges and widths in the retinal photographs. This framework is tested on the REVIEW dataset. This method depends on the initial vessel extraction of the image. FPR or disconnection in the vessels may result in an increase in falsification of edge detection and vessel width computation. This model does not deal with the crossing and branching points separately.

Huang et al. [124] suggested an automatic framework for tracking of retinal vascular network. The approach includes OD position, midline detection of the retinal vascular network, individual network marking, and branch order evaluation. The retinal vessel tree is detected by probing scheme without segmentation. Based on the marked terminal points and OD position, the preliminary position and termination positions for individual retinal vascular network are recognized. In light of the results, morphological properties such as statistical and topological attributes are quantitatively computed, particularly depending on branch order evaluation. The improvement is the identification of the neighborhood statistics of a retinal vessel map. The DRIVE dataset is used for validation of this approach.

A probabilistic tracking framework [125] used Bayesian approach with the MAP criterion for extraction of vessel boundary. Bayesian approach with the MAP criterion is used for extraction of vessel boundary. This approach extracted the vessels efficiently on both synthetic and real retinal photographs and outperforms many conventional vessel extraction techniques. The REVIEW dataset is used for validation of this approach.

The novel technique that depends on particle filtering is presented to track the vessel trajectories in fundus image [126]. The tracking process starts by selecting some preliminary positions within a vessel and remains until the end of the vessel or a branching. In bifurcation, relative to the quantity of branches, the new routes are advised by new preliminary positions. Hence, the methodology contains two stages. In the initial step, the OD position is identified and its nearby impact on biasing gray-level distribution is diminished. At that point, using the distinguished OD, some underlying positions are characterized for the vessel tracing method. In the second step, the tracing is implemented to detect the vessel trajectory utilizing the particle filtering technique.

Fraz et al. [127] designed a model-based approach for computation of vessel caliber from fundus photographs. The initial vascular network is extracted from the vessel probability map photograph. The vessel centerline is detected by using the scale-space skeletonization technique on the binary extracted image. The 2-D model is used for the measurements of vessel widths. The CHASE DB1 dataset is used for validation of this algorithm.

A probabilistic tracking-based technique for automatic detection of retinal blood vessels is presented in [128]. This is a tracking based-method using a Bayesian technique with MAP as criterion to extract vessel edge points. This method is tested on the STARE and DRIVE datasets. The REVIEW database is utilized for diameter computation of retinal vessels.

A minimal path-based technique is proposed for retinal vessel extraction and A/V classification [129]. It contains two major stages: a local classification of vessel segments utilizing the color data and a tracking process depends on the minimal path technique which links vessel segments in succeeding circumferences. Finally, the classification results of all connected vessels are combined with a voting system. Thus, the local clustering scheme permits to eliminate the influence of the irregular brightness in the clustering, whereas the tracking process gives a way to confirm the growing of a vessel taking into account the color information along the whole vessel. This approach is validated on the REVIEW dataset.

Bekkers et al. [130] introduced a novel framework for extraction of vessel edges through the orientation score (ETOS) of an image. The ETOS approach can generally be utilized with both invertible and noninvertible orientation scores (NIOS), which are constructed with cake wavelets and Gabor wavelets, respectively. The IOS produced good results. Another technique that depends on vessel centerline tracking through a multiscale set of noninvertible orientation scores (CTOS) is also presented. The CTOS method is very fast, and the multiscale technique makes the method less stable at critical vessel points compared to ETOS. The HRF and REVIEW datasets are used for the validation of COTS and ETOS approach, respectively.

De et al. [131] presented probabilistic graphical models for automatic tracking of retinal vasculature. This approach consists of two-phase processes: In first phase, vessel skeleton is extracted through segmentation followed by second stage to track the vessel trajectories and direction through graphical models. This model is tested on the DRIVE dataset. The graph-based method to trace retinal vascular network is extended in [132].

Poletti and Ruggeri [133] suggested a graph search-based retinal vessel extraction. Two different aspects are analyzed in this approach: In the first one, accuracy, sensitivity, and specificity of the retinal vessel extraction are computed on three different datasets: the DRIVE, STARE, and YARD. In the second analysis, diameter is computed using the REVIEW dataset.

Bayesian theory and multiscale line extraction-based approach is presented to track retinal vascular map [134]. There is no human interference in the selection of initial tracking points and direction. The REVIEW dataset is used for validation of this approach. Matrix-forest theorem of directed graphs is presented for detection of retinal vasculature [135].

Chen and Cohen [136] designed anisotropic fast marching-based geodesic approach for detection of midlines of retinal vessel segments and their edges. This approach is associated with geodesic or minimal path algorithm which is specifically capable of detecting an elongated profile, such as a blood vessel. The DRIVE dataset is used for confirmation of this method. The performance results of the vessel tracking methods are shown in Table 18.

Table 18 Performance metrics analysis for the vessel tracking based approaches

3.3 Model-based techniques

Vermeer et al. [137] designed a model-based approach for identification of retinal vessels. The performance of this framework is improved by using Laplace and thresholding segmentation step, pursued by a categorization step. The selection of different datasets affects the accuracy of the technique. The approach is evaluated on GDx generated photographs.

Mahadevan et al. [138] suggested a combination of three systems: Huber’s censored likelihood ratio test, the rank-based algorithms [139] and the robust model selection [140], for vessel extraction in noisy photographs [141]. The approach is flexible to combine a range of vessel profile models containing Gaussian, derivatives of Gaussian and dual-Gaussian and different noise types like Gaussian and Poisson noise. The model is validated on simulated and real data. The outcomes are contrasted with the MF and the direct exploratory vessel tracking approach [132]. An improved version of this vessel extraction model is also presented by the authors with the addition of a generalized dual-Gaussian cross-sectional profile for better-quality extraction of vessels comprising a central vessel reflex [142]. The summarization of model based techniques is shown in Table 19.

Table 19 Summarization of model-based retinal vessel segmentation methods

The CNN approaches are a fast mechanism for extraction of retinal vasculature due to its massively parallel processing [143]. The major steps consist of CNN-based histogram equalization and variation, simple adaptive thresholding, and morphological procedure. The MATCNN environment is used for simulation of this methodology. The drawback of this approach is the dependency on different design parameter for opening operator and the varying CNN kernels for the local approximation of the variance. These drawbacks are addressed by a novel CNN-based framework [144] utilizing geometrical information instead of applying generic algorithms [143].This scheme is validated on the DRIVE dataset.

Extraction of retinal vessels by using multiresolution Hermite model (MHM) is presented by Li et al., which utilizes a 2-D Hermite function intensity model in a quad-tree structure over a range of spatial resolutions [145]. The MHM depends on a Hermite polynomial rather than a combination of Gaussian to integrate the central light reflex [146]. A MHM is coupled with an expectation-maximization (EM) optimization approach and statistical associating technique for the modeling and evaluation of retinal vessel network.

Lam and Yan [147] described a new vessel extraction technique for abnormal fundus photographs that depends on the DVFs. In this approach, the midlines are extracted utilizing the normalized GVF, and then, the elongated vessel-like entities are extracted utilizing the Laplacian operator of a pixel. The false distinguished vessel-like elements are trimmed by separation from extracted midlines.

Alonso-Montes et al. [148] suggested a pixel-parallel methodology for detection of vascular network which is adapted from [149] in terms of neighborhood element convolutions and morphological processes coupled with arithmetic and logical processes to be executed and verified in a fine-grain single instruction multiple data (SIMD) parallel processor array [150]. The outward of vessels is sought by parallel dynamic form, the PLS [151].

Lam et al. [152] presented a regularization-based multiconcavity modeling to deal simultaneously both with the healthy and abnormal retinas with bright and dark abrasions. The differentiable concavity measure is designed to deal with bright lesions. The line-shape concavity measure is presented to take out dim injuries which have a power structure disparate from the lengthened vessels in a retina. The locally normalized concavity measure is suggested to handle the randomly scattered noise in a fundus photograph. These concavity measures are coupled together as indicated by their geometrical properties to extract vessels in fundus photographs.

The physical and functional attributes of double wavelength retinal photographs captured at 570 and 600 nm are used for vessel extraction and classification as arteries and veins [146]. In this method, the dual-Gaussian model originally utilized in [153] is expanded to approximate the intensity profile of retinal vessels.

A general illustration for ascending and descending physical vessel profiles with variation in the acuity of edges is suggested by Zhu [154]. An input photograph is converted by 24 log-Gabor filters covering six orientations and four scales in symmetric and asymmetric couples. The equality and unevenness of nearby Fourier parts are figured utilizing the scale-invariant trademark and the Kovesi phase congruency model [155]. The balanced and unbalanced photographs are binaries to get a preliminary segment and a group of edge positions isolating the vessel from non-vessel, respectively, which are further used in an RG procedure for vessel extraction. Kovács and Hajdu [156] proposed a template matching and contour reconstruction for extraction of retinal vessels. It used a generalized Gabor function-based kernels to detect the midlines of vessels. The intensity features of vessel shapes computed in training datasets are remodeled.

Snakes are known as active contour models [157]. A snake is an energy-minimizing spline guided by outward constraint forces and affected by photograph. They lock onto nearby boundaries, localizing them precisely. Several benefits of snakes over traditional attribute extraction approaches are that they are independent and automatic in their finding for a minimal energy state. They can also be simply deployed utilizing outward photograph forces. Snakes are able to trace vibrant elements in temporal as well as the spatial measurements. The key disadvantage of the snakes is that they generally just include boundary data, neglecting other image attributes. Due to this, they often neglect small structures in the procedure of reducing the energy along with the complete route of their contours. Their precision depends on the merging conditions utilized in the energy minimization method; greater precisions need tighter convergence conditions and hence, the greater time complexity. Object tracing, shape identification, segmentation, boundary extraction, and stereo-matching are the major applications of snakes model. The different researchers have inspected the application of active contour models in retinal vessel extraction.

McInerney and Terzopoulos [158] proposed an affine cell image decomposition (ACID) model. The ‘active contour model in ACID’ approach specifically expands traditional snakes, empowering topological flexibility among other attributes. The subsequent topology adaptive snakes are able to fragment some of the most complex-shaped biological features from clinical photographs in an effective and extremely robotic mode. The author has extended the same concept by using ‘deformable organisms’ for automatic vessel extraction, classification and quantitative investigations of physical structures in retina and MR angiograms [159]. Nain et al. [160] presented an implicit deformable model based on image attributes and shape data to develop a region-based active contour that extracts elongated features and penalizes outflows. The approach performed well than the flow based on image statistics only.

Espona et al. [161] utilized the conventional snake along with vessel topological attributes to detect the retinal vascular map. The model is driven by a vessel wrinkle which is really the estimate of vessel midlines distinguished utilizing multilevel set extrinsic curvature based on the structure tensor. The snake is instated and twisted in view of the outside energy characterized by the vessel wrinkle.

The Ribbon of Twin ACM is proposed for detection of retinal vascular map and vessel width computation in [162]. It utilizes two couples of contours to detect every vessel boundary, while preserving diameter stability. The framework precisely traces the vessel boundaries under challenging situations, light reflex phenomena, consisting of noisy fuzzy boundaries, closely parallel vessels and very fine vessels.

Sum and Cheung suggested an improvement in the Chan and Vese [163] approach by including the local photograph dissimilarity into a level set-based ACM to deal with uneven brightness [164]. The framework is confirmed with trials including both synthetic photographs and medical angiograms.

The nonlinear orthogonal projection-based model is presented to detect the properties of vessel map and develops a new local adaptive thresholding technique for vessel extraction [165]. This framework diverges from existing techniques in that it utilizes MF, vessel tracking, or supervised approaches.

Faraz et al. [166] designed a model for quantification and computation of openness of the major temporal arcade (MTA), including Gabor filters to extract retinal vessels and the Hough transform to identify and parameterize parabolic forms. Observing computations of the openness of the MTA and how they vary over time could assist better analysis and improved medication of retinal abnormalities. The model is tested on the DRIVE dataset.

A novel framework for extraction of vessel map that depends on geodesic voting is proposed by Rouchdy and Cohen [167]. The initial start point is selected manually to extract vessel map with no of any additional information is required. This framework uses fourth dimension or with region-based level sets utilizing shape priors to detect both the midlines and edges of the network. This geodesic voting scheme is tested on different biomedical images, including the DRIVE dataset retinal images to extract vessel map.

Guo et al. [168] designed a retinal vessel detection approach utilizing the Mumford level set that depends on Mumford–Shah model. The matched filter function is used for the blood vessel enhancement followed by the binary level set technique and the Mumford–Shah model to perform segmentation.

The parallel active contour model along with alterations in the presegmentation step is presented for segmentation of retinal vessels in normal and abnormal optics images [169]. However, the presegmentation sometimes becomes inefficient to precisely mark the location of curves near the artery walls.

A novel IPACM with hybrid region forms for the retinal vessel extraction is suggested by Zhao et al. [170]. This framework takes the benefit of utilizing various kinds of region statistics, e.g., the intensity knowledge and local phase (LP)-based enhancement tree, which is utilized for its dominance in maintaining vessel boundaries while the particular photograph intensity knowledge will assure a precise attribute extraction. This is a first approach which is validated on fluorescein angiography photographs of the VAMPIRE dataset along with other standard DRIVE and STARE datasets.

Wang et al. [171] suggested region-based ACM, which considers picture intensities and ‘vesselness measures’ as two autonomous random variables with various means and variances, and afterward utilizes the two variables to build a multifeature Gaussian distribution fitting energy to enhance the extraction performance of the designed system.

The MRBIC model for vascular extraction of retinal vessel is described by Rad et al. [172]. To avoid any human intervention and drawback of manual setting, statistical and morphological data of the image are used to select an appropriate IC for level set approaches. This model is evaluated on synthetic and real photographs.

In Oliveira et al. [173] approach, the matched filter, and Frangi’s and Gabor Wavelet filter are combined through weighted mean and median ranking and used to enhance the photographs. The deformable models and the FCM are used for extraction of the retinal vascular network.

The performance evaluation of model-based algorithms for extraction of retinal vessel is shown in Table 20, where the maximum accuracy is stated by Zhang et al. [165] and Kovács and Hajdu [156] on the DRIVE and STARE datasets, respectively.

Table 20 Performance analysis for the model-based techniques

3.4 Hardware implementation-based approaches

The hardware implementation-based approaches are summarized in Table 21.

Table 21 Summarization of hardware implementation-based retinal vessel segmentation methods

Nieto et al. [174] proposed a SIMD design implementation for fast detection of retinal vasculature. It is plotted onto a Spartan 3, amounting to 90 processing elements. The on-chip memory employed was 1.4 MB and saves 8 gray photographs of 144 × 160 pixels. The operational rate is 53 MHz, sanctioning for a 3 × 3 convolution in less than 110 µs. The total time required for segmentation of retinal image having resolution 768 × 584 is 1.4 s.

A fully automatic framework that depends on the local radon transform and utilizes only techniques that are well parallelizable on GPUs by means of CUDA is presented for extraction of retinal vessels [175]. They allow the evaluation of huge datasets in a short time span. Preprocessing steps consist of a black top hat pursued by a Gaussian convolution. It is possible to process photographs of resolution 4288 × 2848 pixels in 1.2 s on an NVIDIA Geforce GTX680. The DRIVE dataset is used for performance evaluation of this approach. Its disadvantage is the high power consumption (the GPU platform requires approximately 500 W) to execute the segmentation operation, which is not acceptable in embedded applications.

Koukounis et al. [176] suggested a VLSI implementation for extraction of retinal vascular tree while analyzing different attributes that affect the power utilization, precision and efficiency of the framework. This is an unsupervised approach which uses MF with signed integers to increase the variance between the vessel and background pixels. The system speeds up the operation of attaining a binary vessels map by utilizing parallel processing and efficient resource sharing, achieving real-time performance. The system is tested on a commercial FPGA design and shows distinct performance enhancements when contested with current similar techniques. Additionally, the low power requirement of this design permits the suggested design to be utilized in portable frameworks, as it attains an efficient stability between execution, power utilization, and precision.

Another approach is modified from [176] and completely redefined to be implemented in CUDA-based GPUs [177]. This approach is assembled in a PC platform containing a four-core Intel Core i7 860 CPU with 8 GB of RAM, utilizing a 64-bit Linux GNU compiler and CUDA 5.0 with the -O3 and -sm_20 flags. This method is based on filtering and contour tracking algorithms. This algorithm contains five major steps: a quick approximation of the retinal vessel map and four steps related to contour tracing based on ACM and a final postprocessing step.

Figure 5 summarizes the performance measures for hardware-based implementation techniques of retinal vessel extraction, where the maximum accuracy is obtained by Krause et al. [175] and Argüello et al. [177] on the DRIVE and STARE datasets, correspondingly.

Fig. 5

Performance metrics for hardware implementation-based approaches

3.5 Hybrid approaches

Table 22 summarizes the hybrid methods, which use multiple techniques or classifiers for extraction of retinal vessels.

Table 22 Summarization of hybrid approaches for retinal vessel segmentation

Villalobos-Castaldi et al. [178] introduced a technique that depends on the second local entropy and on the GLCM for extraction of retinal vasculature. The most important factor is the thresholding value. This approach provides a powerful tool to acquire an automatic threshold limit to extract the vessel and relies upon only on the knowledge enclosed in the evaluated photograph. The DRIVE dataset is used for the evaluation of this approach.

A model of hysteresis-classifier scheme for detection of retinal vascular map is used in [179]. It consists of two classifiers: The first one, called the pessimist, works with a practically zero FPR, which with overlapping classes indicates a high false-negative rate (FNR); the second one, called the optimist, works with a practically zero FNR and a high FPR. Then, utilizing the preceding information about the connectivity characteristic of vessels, the pessimist sorting can be utilized to choose true vessels from among the optimist sorting.

Mudassar and Saira [180] suggested a combination of four different techniques; ILCS, EEED, MMF, and CA for extraction of retinal vessels. ILCS, EEED, and CA are new additions, whereas MMF is an enhanced and improved version of the current MF approach. CA is a favorable method.

Saleh Shahbeig [181] presented a fast and automatic morphological-based retinal vessel segmentation system utilizing CT and PCA. CT is utilized to improve the contrast of retinal photographs by making prominent the boundary of photographs in different measures and orientations. An enhanced morphology procedure with multidirectional structure components is utilized to detect the retinal vessel network. CCA and an adaptive filter are utilized to refine appeared frills with the size of smaller than arterioles in photographs.

A combination of HCF with a discriminative learning system is proposed in [1]. The SWT and WLD both extract the curvilinear structures. The Gabor responses and vesselness filter are used for vessel detection. The new vessel attributes as well as conventional filter-based attributes segmented from orientation invariant local context are used an input into an RF approach for pixel level vessel segmentation.

Salazar-Gonzalez et al. [181] introduced a graph cut method for detection of retinal vascular tree. Two different approaches, MRF image reconstruction technique and compensation factor scheme are evaluated in this article for detection of retinal vessels and OD.

A combination of IUWT, texture entropy, and FCM classifier is proposed in [182]. IUWT is used on the preprocessed photograph to denoise the retinal photograph in its frequency field. FCM on texture attributes based on local gray value entropy is used to extract retinal vessel network.

Dai et al. [183] proposed a gray-voting approach for segmentation of retinal vasculature. The gray-voting technique is utilized to improve the contrast of tiny vessels, while a 2-D Gabor wavelet is applied to segment the major vessels. The gray-voting outcomes are fused with the 2-D Gabor filter outcomes as preprocessing result. A GMM classifier is then applied to detect vessel segments from the preprocessing result, while tiny vessel segments are acquired utilizing another gray-voting procedure, which counterparts the vessel segment detection previously used.

The performance metrics for hybrid approaches of retinal vessel extraction is shown in Table 23, where the highest accuracy is attained by Villalobos-Castaldi et al. [178] and Cheng et al. [1] on the DRIVE and STARE datasets, respectively.

Table 23 Performance metrics for hybrid methods

4 Discussion and conclusion

It is observed that the DRIVE and STARE databases are mostly utilized for validation of different retinal vessel extraction algorithms. Figure 6 displays the frequency of the distribution of datasets used by different retinal segmentation techniques.

Fig. 6

The frequency of the distribution of datasets

The supervised approaches show better performance than unsupervised approaches. However, the performance of supervised techniques reduces on the photographs with uneven brightness because of the increase in FPR in some photographs on the boundary of the OD, hemorrhages, and other kinds of disorders that exist in prominent contrast. The automatic vessel extraction of fundus images has been performed by utilizing MF. Various enhancements and adaptations techniques are suggested based on Gaussian MF by Chaudhuri et al. [64]. The MF alone cannot deal with the extraction of vessel in abnormal fundus photographs; therefore it is usually coupled with other techniques [71, 75]. The issue of the central vessel reflex is addressed by coupling the Gaussian model [142, 146] with ACM [162]. The Sofka and Stewart handle the issue of overlying of the background structures like the retinal edge and the OD in the vascular map detection [25]. The multi-concavity modeling [152] and the DVFs [147] are very effective in handling various retinal abnormalities. The union of radial projections with steerable wavelets and semi-supervised arrangement [184] produced efficient results in the extraction of tiny and poor resolution vessels, thus showing maximum sensitivity. The Gabor filters are very helpful in fundus photograph evaluation. Moreover, along with vessel extraction [86, 88] the Gabor wavelet transform has also been used for the robust fractal exploration of the retinal vascular map.

It is noticed that few articles express the performance in terms of accuracy and AUC, whereas the other papers select sensitivity and specificity for evaluating the performance. The methodology having the highest accuracy of each category on the DRIVE and STARE databases is depicted in Figs. 7 and 8, correspondingly. The highest AUC of each category of the both databases is represented in Figs. 9 and 10, correspondingly.

Fig. 7

Highest accuracy of the each category on the DRIVE dataset

Fig. 8

Highest accuracy of the each category on the STARE dataset

Fig. 9

Highest AUC of the each category on the DRIVE dataset

Fig. 10

Highest AUC of the each category on the STARE dataset

Accurate detection of retinal veins is a fundamental step in computer-assisted diagnostic (CAD) analysis and medical screening of retinopathy. Although high quality of fundus images is available, isolation of the veins and retinal background is still a challenging task. Moreover, pathological disorders of the retinal vessel structure can be seen in various ailments, such as diabetes, hypertension, stroke, and glaucoma. Despite the development of different promising supervised and unsupervised techniques, there is still room for improvement. Table 2, summarizes algorithms developed for extraction of retinal vasculature till now. Some of the reviewed techniques handle both healthy and unhealthy retinal images. Some of those are appropriate for investigation of noisy and central reflex images. Most of the published articles are validated on a limited range of databases which contain a small number of photographs. The performance metrics showed in a large portion of the articles are computed on a limited number of photographs of specific morphological attributes. The limited number of photographs in the DRIVE and STARE datasets do not accommodate for the image-associated attributes, for example inter-image and intra-image variation in luminance, uneven contrast and non-uniform background gray-level qualities. The expansion of approaches, valid for photographs obtained from various imaging system, under diverse environmental situations is also an open range for exploration of vessel extraction methodologies.

Vessel extraction is the fundamental step for abnormality detection. There are large volume of photographs obtained from different kinds of fundus camera. Furthermore, the vast majority of the published techniques are validated on the limited number databases, i.e., DRIVE and STARE. The quality of photographs from these databases is constrained to 0.4 and 0.3 megapixels, correspondingly. This poor contrast is satisfactory for specific considerations like fractal dimension or tortuosity, and computing the vessel width usually needs greater resolution photographs to attain superior accuracy. An advancement in image acquisition brings it easy to collect a large number of patient photographs for analysis. Analysis of these photographs needs fast detection techniques strong enough to execute the photographs obtained from different image capture equipments and imaging conditions. The hardware-based implementations approaches suggest a solution providing high processing speed that is essential in real-time applications.

For accurate and efficient CAD system, precision and strength of the extraction method are important. This article offers an extensive report of existing retinal vessel extraction techniques. We have enclosed both initial and latest articles concentrating on extraction of retinal vasculature methodologies and approaches. Our objective is to present the existing extraction methods, provide the reader an outline for the current research, and to recommend the array of retinal vessel extraction techniques found in the literature. The present drifts, the future guidelines, and the open hitches in the automated retinal vessel extraction are also debated.