Analysis of Vessel Segmentation Based on Various Enhancement Techniques for Improvement of Vessel Intensity Profile

It is vital to develop an appropriate prediction model and link carefully to measurable events such as clinical parameters and patient outcomes to analyze the severity of the disease. Timely identifying retinal diseases is becoming more vital to prevent blindness among young and adults. Investigation of blood vessels delivers preliminary information on the existence and treatment of glaucoma, retinopathy, and so on. During the analysis of diabetic retinopathy, one of the essential steps is to extract the retinal blood vessel accurately. This study presents an improved Gabor filter through various enhancement approaches. The degraded images with the enhancement of certain features can simplify image interpretation both for a human observer and for machine recognition. Thus, in this work, few enhancement approaches such as Gamma corrected adaptively with distributed weight (GCADW), joint equalization of histogram (JEH), homomorphic filter, unsharp masking filter, adaptive unsharp masking filter, and particle swarm optimization (PSO) based unsharp masking filter are taken into consideration. In this paper, an effort has been made to improve the performance of the Gabor filter by combining it with different enhancement methods and to enhance the detection of blood vessels. The performance of all the suggested approaches is assessed on publicly available databases such as DRIVE and CHASE_DB1. The results of all the integrated enhanced techniques are analyzed, discussed, and compared. The best result is delivered by PSO unsharp masking filter combined with the Gabor filter with an accuracy of 0.9593 for the DRIVE database and 0.9685 for the CHASE_DB1 database. The results illustrate the robustness of the recommended model in automatic blood vessel segmentation that makes it possible to be a clinical support decision tool in diabetic retinopathy diagnosis.


Introduction
In the healthcare industry, biomedical images are the primary data source and, simultaneously, utmost hard for analysis. Artificial intelligence (AI) is a technique to automatically analyze and prevent the elevated risks of developing chronic conditions and help patients avoid long-term health problems. As these automated structures become widespread in the healthcare industry, they may bring about progressive changes for radiologists, clinicians, ophthalmologists, and even patients using imaging technology to monitor the treatments. us, it is vital to develop an appropriate prediction model and link carefully for measurable events such as clinical parameters and patient outcomes to analyze the severity of the disease. is permits clinicians to obtain warnings about potential measures before they occur, making more choices about how to progress with a decision to prevent the disease's progression.
e retinal blood vessel is the only part that may catch a straight noninvasively in vivo. Blood vessel plays a significant role in automatic identification as it contains the issue of screening systems. Accurate segmentation and vessel length analysis, orientation, and thickness can make clear the assessment of retinopathy of prematurity, identification of reduction arteriolar, and evaluation of vessel width for the recognition of ailments such as diabetes, arteriosclerosis, hypertension, and so on [1][2][3]. We can read about many ideas to improve the blood vessel segmentation by computing the contrast of retinal blood vessels and background. Computerized evaluation of vasculatures has been extensively recognized as the initial stage in the progress of a computer-aided investigative scheme for ocular ailments. Several suggested rules have been recommended for vessel segmentation [4]. Few commonly suggested algorithms for vessel extraction are discussed. Few authors have proposed a vessel segmentation process using a matched filter. Images at different scales are convolved with the filter, and the highest output is noted at each pixel [5][6][7]. With an assumption of elongated vessels, Staal et al. presented an idea of ridge-based vessel segmentation in which image ridges are transformed to control line elements [8]. It has recommended an adaptive local multi-threshold probing algorithm. For multithreshold probing, they have used different thresholds in series for computation [8]. In the literature, it has introduced automatic vessel tree segmentation by combining shifted filter responses (COSFIRE) [9,10]. e literature suggests using BCOSFIRE and generalized matrix learning vector quantization (GMLVG) to detect the blood vessel [11]. Mapayi et al. have discussed and compared vessel segmentation based on global thresholding [12]. Many filters are introduced by various researchers for retinal blood vessel segmentation, such as median, Gaussian, matched, Gabor, Cake, steerable, Frangi, and many more [5,[13][14][15][16][17][18][19].
Afterwards, many extensions of the existing techniques in several directions are recommended for blood vessel segmentation by various researchers [20][21][22]. Few of the various extended filters are discussed. e original median filter is extended as an improved median filter (IMF), hybrid median filter (HMF), and weighted median filter (WMF) for vessel segmentation [23]. Several expansions of matched filter (MF) are utilized; for example, MF is integrated with pulse coupled neural networks, and the Otsu algorithm is applied for segmentation [24]. It has been suggested in the literature to improve the matched filter through the ant colony algorithm and through Clifford matched filter [25,26]. A zero mean Gaussian matched filter is introduced based on a first-order derivative of the Gaussian filter [27]. An upgraded version of the matched filter is suggested through an optimization technique [28,29]. e matched filter is upgraded through another optimization technique, that is, genetic algorithm [30]. Another recommended way to improve the matched filter is using particle swarm optimization [31,32].
Correspondingly, for the Gabor filter, there are many expansions presented in the literature. A multi-scale, multidirectional Gabor wavelet transform and created feature vector consisting of pixel intensity and maximum response achieved for Gabor filter at various scales are recommended. Afterward, they utilized a classification algorithm known as linear minimum squared error (LMSE) [33]. Two-dimensional Gabor wavelet with a Gaussian mixture model is presented to classify a pixel as a vessel or a nonvessel [34]. Two different approaches are compared for blood vessel extraction. In the first approach, they have employed Gaussian filtering for preprocessing, LoG filtering to enhance the retinal image, and adaptive thresholding for the segmentation task. In the second approach, they have utilized unsharp masking for preprocessing, Gabor wavelet to enhance the retinal image, and global thresholding for the segmentation task [35]. A technique is suggested for the green channel noise reduction of the retina by employing a low pass radius filter and followed by the Gabor filter and a Gaussian fractional derivative for enhancement of blood vessels [36]. Gabor filter is extended by integrating Gabor, Frangi, and Gaussian filters with top-hat transform [37]. A new technique is introduced to design a set of 180 Gabor filters with variable scales and elongated variables by applying an optimization approach known as competitive imperialism algorithm (CIA) for vessel segmentation [38]. A new hybrid scheme is suggested by combining the existing techniques in which multi-scale vessel enhancement (MSVE), morphological operations, bottom-hat transform, and image fusion are combined for blood vessel extraction [36]. Gabor filter and Hessian method are used together for enhancing the features. en K-mean clustering is utilized for vessel extraction [39]. A new improved curvelet transform technique is suggested to detect thick and thin blood vessels for extraction [40]. A hybrid method by combining two different existing techniques such as lateral inhibition and differential evolution is used for vessel segmentation [41]. Existing supervised and unsupervised machine learning techniques are utilized for vessel segmentation by employing image features [42]. To enhance the performance of the original Frangi filter, it is combined with the existing probabilistic patch-based denoiser for vessel segmentation [43].
Newly deep learning that is a supervised approach has been effectively employed for biomedical image processing that includes retinal blood vessel segmentation. Wang et al. have suggested context spatial U-Net for the segmentation of blood vessels [44,45]. Chen et al. have discussed many deep learning approaches for vessel segmentation in their review paper, where better results are achieved [46]. Many machine learning algorithms are available in the literature for various disease detections [47][48][49][50][51][52]. However, deep learning applications depend on an enormously huge database. Moreover, annotated data sets are not readily available compared to other imaging fields. Annotation of medical data is a costly, complicated, and lingering process and thus experts need more time.
Additionally, an annotation may not always be possible for rare health issues. Consequently, the availability of medical data is a significant obstacle for deep learning approaches. Although deep learning methods have achieved substantial achievement, decent theory for deep learning algorithms is still absent. Models of deep learning offer good results, and the researchers are utilizing continuously deprived of having an understandable knowledge of attaining higher results and the work process. Another critical challenge is the legal association of black box utility. It can be a barrier because healthcare experts would not depend on it. If the results achieved are wrong, then who could be accountable. Because of this sensitive issue, hospitals may not be convenient with the black box, that is, how it could draw that particular result from the ophthalmologist. erefore, understanding deep learning techniques and their hidden layers working for a given problem is a great challenge for researchers. Furthermore, in the event of the source of data changes, the problem occurs in network response, which most researchers do not address. at will be the influence of modification in a data acquirement device because this may give on to variations in features of images like colour intensity levels or illumination. us, the absence of generalize ability will harm the performance of deep learning networks. Accordingly, it is concluded that deep learning networks still deliver higher performance results depending on huge image databases. Consequently, it needs large storage and memory with excess training time for the networks.
e insufficient availability of large biomedical imaging data sets is another hurdle in developing a deep learning network [53,54].
Consequently, to enhance the performance of any model whether supervised or unsupervised, the quality of the image has a great impact on the performance model. Few factors in the image like uneven illumination or camera position can affect the image contrast, resulting in inadequate features in the image. us, image enhancement is a very important part of preprocessing and the proper selection of enhancement techniques can improve the effectiveness of the existing models to a great extent. As a consequence, it is essential to research the relationship between image enhancement and the existing models. us, in this work, an unsupervised approach, that is, traditional Gabor filter, is chosen to improve its performance by employing various enhancement techniques.
Six different enhancement algorithms are used in the proposed work. e advantages and disadvantages of a particular enhancement algorithm are difficult to describe because of the reliable and consistent measures for the evaluation of the superiority of the enhanced image. us, based on the experimental results, the best-integrated model is derived.
After an extensive study of the literature, it is noted that many existing techniques are taken into consideration for modifying and improving their performances. erefore, existing methods can still be considered for fundus image segmentation by upgrading them and boosting their computational ability. Gabor filters are found to be effectively suitable in the segmentation of retinal images because of oriented features as the vessels of the retina are linked and piecewise linear [34]. Furthermore, Gabor filters can be adjusted to particular frequencies and thus can be adjusted to enhance the blood vessel. Although from the literature it is observed that there are many techniques available for vessel extraction utilizing various filters and enhancement methods, still, a lot can be done to ameliorate further. Designing a particular enhancement technique is infeasible, as it generates a visual artifact-free output. Selecting a specific enhancement scheme is hard since parameters in assessing output quality are not available. Furthermore, usually enhancement algorithms rely on authentic parameter selection. is prompted a recommendation a robust enhanced Gabor filter by integrating it with various enhanced techniques such as gamma corrected adaptively with distributed weights (GCADW) [55,56], homomorphic filter [57,58], joint equalization of histogram (JEH) [59,60], unsharp masking filter, adaptive unsharp masking filter [61,62], and particle swarm optimization (PSO) based unsharp masking filter [63,64].
Additionally, it is noted in the literature survey that the researchers have combined two to three existing techniques for the improvement of original approaches. In this work, we have presented an idea to improve blood vessel segmentation through illumination-robust Gabor filter by combining it with six enhancement techniques. e main contributions of the suggested approach are covered in few steps as follows: (a) Initially, the existing Gabor filter is used to enhance the fundus image, and hysteresis thresholding is applied for vessel segmentation. (b) In the second step, different enhancement techniques are combined individually with the Gabor filter to make its illumination robust and to improve its performance followed by hysteresis thresholding for vessel segmentation. (c) In the final postprocessing step, a morphological cleaning operation is performed to clean undesired pixels that may lead to more false positives.
e suggested methods are assessed on DRIVE and CHASE_DB1data sets, and based on the results, the bestintegrated model is finalized. Table 1 shows the summary of the advantages and disadvantages of various vessel segmentation methods.

Gabor Filter.
Gabor filters are influential techniques that have been extensively utilized for multi-scale and multidirectional analysis in image processing. Because of its directional selectiveness ability to detect oriented features, it is extended by proposed fine-tuning. As a result, precise frequencies and scale are shown in filter performance as lowlevel oriented edge discriminators. e features from the Gabor filter can be extracted from the original image as described below [34]: where f(x, y) is the original image and P f (x, y) is the impulse response of the 2-D Gabor filter. e symbol ⊗ represents the convolution sum.

Gamma Corrected Adaptively with Distributed Weights
(GCADW). GCADW utilizes cumulative distribution function (cdf ) and employed normalized gamma function to it. ey have achieved a modified transformation curve where histogram statistics are available. Accordingly, Computational Intelligence and Neuroscience substantial adjustment can be done in the lower gamma parameter.
us, they have formulated adaptive gamma correction (AGC) to process intensity in consecutive increments of the original trend. AGC is defined as follows: where l is the intensity of the input image, l max is the maximum intensity of the input and c is the varying adaptive parameter. e low intensity can be increased substantially without decreasing the high intensity by applying AGC technique. Additionally, weighting distribution (WD) function applied is also employed for the modification of statistical histogram to some extent for the reduction of adverse effect. e WD function is defined as follows: where a is an adjustable parameter, pdf max is the probability density function with a maximum value of the statistical histogram, and pdf min is the probability density function with minimum value. Considering the (3), the revised cdf is approached as follows: where the sum pdf w is computed as follows: In conclusion, the value of parameter gamma derived from cdf equation (3) is altered as follows: 2.3. Homomorphic Filter. Many suggested approaches are available to enhance images utilizing a homomorphic filter [57]. Information missing in dark regions can be identified by equalizing the light variations onto the image. An image can be denoted as a product of two components as seen in the following equation: where L (x, y) is the illumination and R (x, y) is the reflectance components of the original image. e filter function for the homomorphic filter chosen is as follows: where k controls the steepness and is taken as constant, P 0 is the frequency of cut-off value, the measured distance of the origin Fourier transform is represented as P(u, v), and c l , c h are the low-and high-frequency gain, respectively.

Joint Equalization of Histogram.
Joint histogram equalization is an approach where modification of histograms and enhancement of contrast in digital images are implemented [59]. e entire joint histogram equalization process is explained below. By using a neighbouring window of Z 2 , the gray value pixel g(p, q) is calculated and defined below: where the expression h(a, b) represents the existence of the gray level pair numbers f(p, q) and g(p, q) around the correspondent spatial location (p, q) of the images I and I correspondingly. It signifies the count function. Because a and b is considered whatever conceivable numeral value among 0 and C − 1, the number of pixel pair groupings feasible are C × C. us, the joint histogram H will comprise C × C entries. By utilizing the count function, the cumulative distribution function can be achieved as follows: Two-dimensional CDF value is utilized to produce the output pixel intensity enhanced in contrast. e equalized value of the intensity pairs (p and q) in the output image can be achieved through the histogram equalization method as follows:

Unsharp Masking Filter.
Local contrast enhancement can be done using unsharp masking. is technique creates a mask of the original image utilizing a negative image. Afterward, the original positive image is combined with the unsharp mask to produce an image that is less blurry than the original. Usually, a linear or nonlinear filter that magnifies the high-frequency components of a signal is said to be an unsharp masking filter.

Adaptive Unsharp Masking
Filter. Lin et al. have suggested an adaptive unsharp masking filter to enhance the colour images [61]. ey have improved the colour brightness by stretching the colour channel and enhanced the contrast by expanding the edge. Especially, a hyperbolictangent function scale is established that regulates the gain, and sharpness is enhanced. e model relates to the intensity of the original image and recognized edges.
Input pixel magnitude and image coordinate edge are the two components of the hyperbolic scheme. e gain factor considered for the input-based component is defined as follows: where p � 1, . . ., P and q � 1, . . ., Q represent width and height of image, respectively, and λ g pq gain of pixel of I channel, as the RGB image is initially converted to HSI image. While g pq � 1/3(R + G + B) pq relates to the I channel of HSI colour space, where R, G, and B correspond to the red, green, and blue components of the colour image, respectively. us, to gain adjustment on the detected edge, the subsequent scheme is applied.
where λ d pq is gain factor defining strength of reconstructed edge d pq . By multiplying the above two schemes, the complete gain adjustment scheme is obtained and described as follows: Additionally, the measurement of sharpness is evaluated. It is computed from the measurement of the neighbourhood pixel gradient as described below: where Δx pq � g pq − g p+1,q and Δy pq � g pq − g p,q+1 represent the horizontal and vertical gradients across the image, respectively, and N is the total number of pixels. Additionally, an image is evaluated by its colourfulness [65]. Capturing an object under uneven lighting conditions may deteriorate the measurement. e colourfulness is given as below: where e colour is computed with respect to the standard deviations σ rg and σ yb of colour differences Δ rg and Δ yb , respectively, where where μ RG and μ YB signify the means of Δ RG and Δ YB , respectively.

Particle Swarm Optimization (PSO) Based Unsharp
Masking Filter. Particle swarm optimization (PSO) is implemented to design the kernel and gain on an unsharp masking filter [63]. e authors have applied a symmetrical kernel through PSO for extraction of the edge with an optimum scale factor for the enhancement of colour image. PSO is utilized to achieve filter kernel settings and after maximizing the information's content to minimize high ranged pixels.
Computational Intelligence and Neuroscience 5

Proposed Methodology
e main aim of the proposed approach is to enhance retinal Gabor transformed images for blood vessel segmentation.
is work starts with a 2-D Gabor filter for enhancing the retinal images to obtain the Gabor feature images. Subsequently, the Gabor feature images are enhanced by combining various enhancement techniques such as GCADW, homomorphic filter, JEH, unsharp masking filter, adaptive unsharp masking filter, and PSO unsharp masking filter. Segmentation is executed by utilizing hysteresis thresholding. e proposed methodology covers several stages. e proposed approach is summarized in Figure 1 as blocks that elaborate on each step. e entire proposed segmentation process is described in detail as follows.

Gabor Filter.
e approach activates by employing 2-D Gabor filters on the retinal images to achieve the Gabor transformed images. e Gabor transformed images are having object regions with enhanced boundaries. e Gabor kernel used in this work for producing P f (x, y) is defined as follows: where the scale and orientation are determined by u and v, respectively. z � (x, y) .represents the norm operator. e wave vector k u,v � k v e jφu , where k v � k max /λ v and φ u � πu/8. e parameter λ denotes the space among the filters in the frequency domain. e oscillatory part of the kernel is defined inside the first square bracket term, while the second term compensates the dc value of the kernel. e values of the Gabor kernel are selected as follows. λ � √2 for better intensification near the transition regions. e value of k max � π/2, width � 60, height � 60, and σ � π/3. With parameters k max and λ, Gabor filters of eight orientations are generated. Retinal Gabor transformed images are generated by convolving the retinal images with Gabor filters. Only the magnitude part is taken into consideration because the phase part is time varying in nature.  (i) GCADW enhancement is computed with three vital steps. e detailed mathematical computations of GCADW are described in (2)-(4) and (6). e steps of GCADW are summarized as follows:   (10), in which 256 × 256 is the order of the matrix. Utilizing (12), the equalized joint histogram is computed, and improved enhanced images are produced. e description of the JEH enhanced technique is as follows: Figure 5(a) represents the intensities of a grayscale subimage k of 8 bit and size 6 × 6. By utilizing (11), the average subimage M is accomplished, and Figure 5(b) represents this. e size of the window is considered as three because the higher size window may blur the image, and according to the location, the pixel pairs are generated. For instance, both input and average images are represented in a pixel pair (1, 1) with values (111, 76). Among the pixel pairs, the minimum and maximum value specified by CDF is (109, 81) and (167, 152), respectively. e joint equalized histogram value is achieved by (12). For example, the CDF of (140, 139) pixel pair is 11. e histogram equalized is calculated as follows: In the original subimage, the intensity value is substituted as 140, that is, M at every occurrence of the pixel pair (140, 139) where f smooth (m, n) is a smooth version of f(m, n). e final sharpening image obtained through unsharp masking is given as follows: where k is a scaling constant. Values of k vary between 0.2 and 0.7, with the higher values providing growing amounts of sharpening. (a) RGB input image (b) RGB colour space is converted to HSV colour space (c) PSO iteration count is set as zero (d) Consider kernel element and gain as a particle and initialize the particle (e) Repeat (f ) kernel is generated from each particle (g) Unsharp masking filter operation is carried (h) Compute entropy penalized by over-range ration (i) Update the global best solution and particle motion (j) Particle position is updated until maximum iteration is touched (k) Return optimum solution, that is, global best solution (l) Finally, the edge extraction kernel and the augmentation gain factor are tuned using the PSO optimizer to yield contrast-enhanced images with minimum over-range artifacts.

Results and Discussion
e proposed idea is analyzed and examined on DRIVE (Digital Retinal Images for Vessel Extraction) and CHASE_DB1 (Child Heart and Health Study in England) databases.
e DRIVE data set contains 20 coloured fundus images in each training and testing set, an equivalent set of masks, and two manually segmented sets. Computational Intelligence and Neuroscience e first manual segmented image that first ophthalmologist provides is preserved as the ground truth image. In supervised models for training the network, the training data set is generally utilized. e test data set is utilized for the computation in this work as the proposed method is an unsupervised method. e ground truth images of the test data set are utilized for analysis purposes. e CHASE_DB1 data set consists of ground truth images of left and right eyes taken from 28 children.
Comparisons between segmented and ground truth image are verified using metrics, that is, sensitivity (Sen), accuracy (Acc), and specificity (Sp).    TN + FP + FN) , where TP: true positive states as the correct identification of a vessel, TN: true negative states as the correct identification of a background, FP: false positive states as incorrect identification of a vessel, and FN: false negative states as incorrect identification of a background. e measure of ability for verifying the correct vessel pixel is known as sensitivity. In contrast, the measure of the ability to verify accurate nonvessel pixels is known as specificity, and the accuracy displays the conventionality of the segmentation result. Tables 2 and 3 present the performance of blood vessel segmentation using the original Gabor filter in terms of Sen, Acc, and Sp for DRIVE and CHASE_DB1. Traditional Gabor filter delivers Sen, Acc, and Sp as 0.6434, 0.9215, and 0.9470,       In view of the DRIVE database, the suggested approaches report a substantial improvement in Sp and Acc scores. However, the suggested approaches report moderate Sen scores. Considering the CHASE_DB1 database, the suggested approaches report a substantial improvement in Sen, Acc, and Sp scores. It is also noteworthy here that in terms of accuracy, the suggested approaches register a substantial improvement for both the databases. e entire segmentation approach can be categorized into under-segmentation, over-segmentation, or accurate segmentation.
(i) If specificity shows a higher side and sensitivity shows a lower side, then the vessel is undersegmented that may result in the vessel being inaccurately recognized (ii) If specificity shows the lower side and sensitivity shows the higher side, then the vessel is oversegmented and may lead to nonvessel being recognized as a vessel    Overall, it can be deduced from the above discussion that all the enhancement variations that are adopted for Gabor     Figure 9: Flowchart of the final algorithm.
Computational Intelligence and Neuroscience filter yield better performance measures either with regard to Acc or Sen or Sp. Figure 9 illustrates the flowchart of the final algorithm of the suggested approach. e segmented images of the retina 2 and retina 4 of the DRIVE database and retina 5 of the CHASE_DB1database achieved from different suggested approaches are represented in Figures 10-12  unsharp masking filter. Utilizing the global best solution, the unsharp mask image is generated. In the next step, the Gabor filter is applied, and a maximum Gabor enhanced image is generated. In the last step, hysteresis thresholding is applied with morphological cleaning for the vessel extraction.

Conclusions
In this work, six enhancement techniques are individually combined with the Gabor filter to improve the performance of the standard Gabor filter. e proposed techniques are assessed using DRIVE and CHASE_DB1 data sets. All together six algorithms are recommended for the improvement of the traditional Gabor filter. e parameters Sen, Acc, and Sp are taken into account in order to determine the best algorithm. Experimental results are compared with state-of-the-art models. It is observed that the homomorphic filter and unsharp masking filter integrated with the Gabor filter underperforms compared to the standard Gabor filter in terms of sensitivity on the DRIVE database. Similarly, homomorphic and unsharp masking filters combined with the Gabor filter underperform the standard Gabor filter in all performance measures on the CHASE_DB1 database. e best results are attained with a PSO unsharp masking filter with the Gabor filter by delivering an average value of Sen, Acc, and Sp of 0.748, 0.959, and 0.9801 on the DRIVE data set, respectively, and 0.759, 0.961, and 0.984 on the CHA-SE_DB1 data set, respectively. erefore, it is inferred that adding different enhancement techniques before Gabor filter boosts the performance of the traditional Gabor filter and also improves the accuracy, specifically with respect to the tiny vessels.
Consequently, it is observed that though deep learning, a supervised approach is actively implemented for blood vessel extraction in recent research and achieving better results; still, the unsupervised traditional methods can be enhanced to achieve precise vessel segmentation. Also, the results of the suggested approach that is an unsupervised approach outperform many state-of-the-art methods that are coming under the group of unsupervised approaches.
Moreover, it will enable new practical applications, where analysis of low-contrast images in real time is required, for example, robotic microsurgery of the eye.
A drawback of the suggested model is that even though six enhancement techniques are combined with the Gabor filter, but only one integrated model is able to perform better as compared to the other integrated models. is is because enhancement of certain features might be accompanied by undesirable effects that might be led to the loss of valuable image information.
For future studies, we suggest considering different illumination normalization techniques such as small-scale retinex (SSR), multi-scale retinex (MSR), isotropic illumination, wavelet normalization, and so on combined with deep learning approaches for vessel segmentation.

Data Availability
Publicly available data are used in this study.

Conflicts of Interest
e authors declare that there are no conflicts of interest.