An Efficient and Light Weight Deep Learning Model for Accurate Retinal Vessels Segmentation

Detecting eye diseases early can make a difference when trying to treat them. Existing diagnostic systems are not only prone to inaccurate judgments, but are also difficult and require a longer time from experts. Artificial intelligence (AI) based on deep learning (DL) has attracted global interest recently because of its effectiveness and accuracy in detecting eye diseases. There are several challenges in diagnosing eye diseases based on retinal fundus imaging. Most of the previous models in the literature targeted these challenges and tried to improve the evaluation metrics, whereas very little attention has been given to reducing the computational complexity of the developed model. The reduced computational complexity is highly desired, as the aim of developing an AI-enabled automated diagnostic system is to use for population-scale screening programs. This observation motivated us, and this work aimed to use a lightweight DL model based on ColonSegNet for retinal vessel segmentation. The performance of the model was assessed using three open-access fundus image datasets: DRIVE, CHASE_DB1, and STARE, and it achieved sensitivity, specificity, accuracy, AUC, and MCC performance of (0.8491, 0.9774, 0.9659, 0.9850, and 0.7960), (0.8607, 0.9806,0.9731, 0.9869, and 0.8014), and (0.8573, 0.9813, 0.9719, 0.9873, and 0.8069) respectively. Furthermore, the proposed method has five M trainable parameters, making it lightweight and capable of deployment on low-end hardware devices. These results outperform several lightweight and computationally heavy methods. The reduced number of parameters, computational complexity, and improved segmentation performance support its use in automated diagnostic systems for retinal vessel segmentation.


I. INTRODUCTION
Due to the rapid progress in computing hardware, especially graphical processing units (GPUs), and their cheap availability in the last two decades, efficient deep learning (DL) frameworks have been explored for solving different problems in various fields of science [1]. Artificial intelligence (AI), based on DL, has attracted global interest in recent years. Medicine and healthcare are not exceptions, where DL has been applied to medical imaging analysis, which has shown robust diagnostic ability in detecting different medical The associate editor coordinating the review of this manuscript and approving it for publication was Alvis Fong . diseases. In ophthalmology, DL achieves robust classification performance in detecting and diagnosing various eye diseases, for instance, diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma [2].
AI began in the 1950s, followed by machine learning (ML) in the 1980s and DL in the 2010s. AI is the ability of a machine to mimic human behavior and make decisions based on its learning. Thus, AI has become the fourth industrial revolution in the history of mankind and is the most popular topic for researchers [2]. ML is a subset of AI that involves learning and making predictions and/or decisions without being explicitly programmed [3]. With the rapid progress in computing hardware, especially GPUs, DL is the most recent VOLUME 11, 2023 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ ML subset that simulates human brain cells called neurons that learn by themselves. Figure 1 shows the relationship between DL, ML, and AI. ML algorithms are categorized as supervised, unsupervised, and reinforcement learning. In supervised learning, the input and output data are labeled whereas in unsupervised learning, the data are not labeled, and it is learned without supervision. Supervised methods are learned based on the features of the input images that are manually marked. In contrast, unsupervised methods discover hidden features and do not require manually segmented images. In reinforcement learning, there is no input data, and it depends on the actions of the algorithm. Reinforcement learning seeks long-term cumulative rewards to achieve an optimal solution. On the other hand, supervised and unsupervised learning typically look for instant rewards [3], [4].
Traditional ML comprises two independent stages: feature extraction and classification. Features such as pixel value, texture, and shape are extracted manually (handcrafted feature). The classification is built independently of feature extraction and has many types of classification algorithms, for instance, the support vector machine (SVM) algorithm.
The key to an effective model construction depends on the identification or extraction of accurate features. In the DL algorithm, feature extraction is automatic. The features learn from training without human intervention. DL has outstanding accuracy, sometimes beyond human perception [3]. The working principles of the DL and ML algorithms are shown in Figure 2.
DL learns the representations of data by using multiple hidden processing layers. DL can discover complex structures in large datasets and change its internal parameters in each layer by using the backpropagation technique [3]. Recently, state-of-the-art results produced for image recognition and analysis using DL. Identifying the pixels of organs and distinguishing them from different diseases is one of the most challenging tasks in analyzing medical images. Image segmentation is the first step toward developing effective strategies for image processing-based disease diagnoses [1]. For this purpose, researchers have designed segmentation algorithms to improve the accuracy of segmentation for different organs, including the brain, eyes, chest, and others.

A. AUTOMATED TOOLS FOR EYE DISEASE CLASSIFICATION
Accurate diagnosis of the structure of the retinal vascular tree and the optic cup/ disc is essential for identifying  various eye diseases, including DR, glaucoma, and AMD [6]. DR is defined as a defect in the retinal vasculature caused by long-term diabetes. DR leads to blindness before the age of 50 years, where there are almost no signs at the initial stages of the disease [7]. Glaucoma, the second major cause of blindness, is a chronic disease that affects the optic nerves, if it is not diagnosed and treated in a timely manner, it may destroy blood vessels and optic nerves and cause blindness [8]. AMD is one of the most common irreversible causes of vision loss. The pathogenesis of AMD remains unclear, but it affects people aged 65 years and older [8]. The differences between the normal and abnormal retinal fundus images are displayed in Figure 3.
Relying on expert ophthalmologists and optometrists to manually segment and analyze images of the retina is a traditional method, which is not only prone to inaccurate judgments due to fatigue but also difficult and requires a longer duration from experts [6]. The tremendous developments in digital imaging and computer vision have led to the possibility of using image processing to diagnose eye diseases. Therefore, AI techniques based on automated or semi-automated computer-aided diagnosis (CAD) systems play a crucial role in health monitoring, medical informatics, and medical imaging. CAD systems can significantly reduce the burden on public healthcare resources. Therefore, CAD systems based on DL have become a research trend that surpasses traditional segmentation methods in terms of accuracy [3]. Figure 4 presents the working framework of the different health-monitoring models.

B. KEY RETIINAL IMAGE FEATURES AND PATHOLOGIES
This section defines the primary retinal features and pathologies and discusses their significance in automated retinal image analysis. The human retina consists of different retinal features, such as blood vessels, fovea, macula, optic disc (OD), optic cup (OC), exudates (EXs) (hard and soft), hemorrhages (HMs), and microaneurysms (MAs). A retinal fundus image captures a photo of the retina by using a fundus camera. These images help to diagnose eye diseases that occur in the retina. The macula is located at the center of the retina. The macula receives images and light signals. In the central part of the macula, there is a dark brown or red fovea that does not contain any blood vessels [10]. AMD begins with characteristic yellow deposits (drusen) in the macula [8].
OD is the bright yellow part of the retina. The OD shapes vary from one person to another, but they appear almost circular. In OD, there are no photoreceptors, namely cones and rods, and the OD is known as the blind spot. The OC is smaller than the OD and is the bright central part of the OD. Glaucoma causes changes in the shape of the OD, which can be identified by measuring the retinal part of the OD and OC. In glaucoma, the size of the OC increases more than the normal size [8], [10].
The blood vessels (arteries and veins) that look like trees with roots and branches appear in images of the retina [5]. Many chronic eye diseases can be detected by observing structural changes in the retinal blood vessel diameters. DR occurs when fluid leaks into the retina because blood vessels are damaged inside the retina. Different signs of DR, such as MAs, HMs, and EXs, can be detected. MAs occur  because of leakage from tiny blood vessels in the retina. MAs are circular in shape, red in color, and small in size. MAs are considered the first sign of DR. HMs occur because of rupture of the MAs walls and look like red dots. EXs occur due to leakage of the arteries and veins, which contain lipids and proteins and create yellow spots on the retina. EXs lead to complete blindness if lipid accumulation is near or on the macula [7]. A healthy retina, along with its main components, and examples of MAs, HMs, and EXs appeared in Figure 5.

C. RETINAL IMAGE PROCESSING
Retinal vessel segmentation face many challenges, as shown in Figure 6, where some retinal vessels have similar intensities to that of the background color, which makes the identification process more complicated. Furthermore, accurate vessel segmentation becomes difficult when the contrast of fine retinal vessels in subnormal regions of the retina is low [11]. Another critical issue is the pixels near the center of the vessels are misclassified as background pixels, which occurs because the edges of vessels have more intensity compared to their mid-lines, called the central light reflex [12]. There are different challenges like uneven background illuminations, merging of closely parallel vessels, bifurcations, and crossover regions [5]. However, camera calibration must be considered at the time of capture, and drift in image brightness [6].
Retinal fundus images processing is the primary task in retinal vascular segmentation, involving several steps such as capturing a photograph of the retina, image preprocessing, and postprocessing. A retinal fundus image captures the photo by using a specialized fundus camera with a flash attached to a microscope. The camera is used to capture the back eye structure, such as the macula, OD, and the central and peripheral retina. Fundus photography can be decomposed into different modes, such as using a green filter to remove red light from fundus photographs. The retina can also be examined using full color in color photography [13].
The preprocessing step aims at better segmentation of the retinal vessels through vessel enhancement and noise reduction. The authors in [13] briefly mentioned most of the commonly used preprocessing steps such as image transformation, image filtering, image segmentation, green channel extraction, feature extraction and selection, knowledge-based image enhancement, and image restoration. Post-processing is the last step for better segmentation accuracy, such as morphological opening and closing, along with dilation and erosion [13].
The focus of this work is to develop a lightweight model for accurate retinal vessel segmentation based on ColonSegNet [14]. We selected ColonSegNet because of its many advantages, including low computational complexity, and a low number of trainable parameters, which can be deployed on low-end devices. Hence, it can be used for the early diagnosis of eye diseases at the point of care in hospitals. The following is the structure of this paper. The second section reviews the literature. This section is followed by section three, which describes the proposed method. Section four deliberates the experimental results and provides a comprehensive assessment of the representative models from the literature. Finally, conclusions are presented in Section five.

II. LITERATURE REVIEW
Different retinal vessel segmentation techniques can be categorized into supervised and unsupervised methods. Supervised methods are learned based on the features of the input images and manually marked images. In contrast, unsupervised methods discover the hidden features of blood vessels and do not require manually segmented images [15]. In [13], the authors presented different vessel segmentation techniques for the supervised and unsupervised methods. The authors divided supervised methods into SVM, artificial neural networks (ANNs), and miscellaneous methods. Unsupervised methods are subdivided into the matched filter (MF), mathematical morphology, model-based, and vessel tracking methods.
The SVM [13] is a well-known ML model that is typically used for supervised classification. In [16], the authors presented a new automatic blood vessel detection algorithm for retinal images. The proposed method uses Gabor filters to compute the feature vectors for each pixel. Then, the Gaussian mixture model (GMM) and SVM were used to classify the extracted features. This method was performed using the DRIVE database and achieved a sensitivity, specificity, and AUC of 0.9650, 0.9710, and 0.974, respectively.
ANNs contain a large number of neurons that attempt to simulate the human brain. These neurons are connected to enable the system to predict the target output. The novel method proposed in [17] attempts to address the imbalance segment between thick and thin vessels. It proposes a threestage deep learning model, a thick segmenter, a thin segmenter, and a fusion segmenter. First, the thick segmenter segments the thick vessels, and then they refine by the fusion segmenter, which tackles the central vessel reflex. Subsequently, the effects of lesions are removed from the thick vessel segmentation local features by the thin segmenter and the fusion segmenter. This method was performed on the DRIVE, STARE, and CHASE-DB1 datasets and it achieved accuracies of 0.9538, 0.9638, and 0.9607, and sensitivities of 0.7631, 0.7735, and 0.7641, respectively.
Miscellaneous methods include various supervised techniques for the automatic detection of retinal fundus images, such as 2D Gabor wavelet, ridge-based, and feature based [13]. The authors in [18] presented a method for automated segmentation by classifying each pixel as vessel or non-vessel. They applied the GMM classifier to speed up classification and a 2D Gabor wavelet transform for vessel enhancement and noise removal. This method was performed on both the STARE and DRIVE datasets and achieved accuracies of 0.947 and 0.948 and AUC of 0.961 and 0.967, respectively.
MF methods were used to clarify vessel outlines. Retinal vessel segmentation based on a Gaussian matched filter (GMF) and U-net was proposed by the authors in [19]. They applied GMF to strengthen thin vessels. This method was performed on the DRIVE dataset and achieved an accuracy of 0.9636.
Mathematical morphology methods are used to extract image components and recognize specific shapes, such as boundaries and convexity. The authors in [20] identify three main processing phases: preprocessing, vessel center-line detection, and vessel segmentation. They normalized the background and enhancement of the thin vessel using the arithmetic mean kernel in the preprocessing phase. Gaussian filters were used to define a set of connected segments in the central part of vessels. In the last phase, top-hat transforms were used for multiscale morphological vessel enhancement. This method was performed on both the STARE and DRIVE datasets, and achieved accuracies of 0.9579 and 0.9633, respectively.
Model-based approaches aim to extract retinal blood vessels by using explicit vessel models. This method includes two models: deformable and vessel profile models. In [21], the authors used a morphology feature driven deformable model for the extraction of vessels. This model was developed to address complex vessel shapes with low contrast. Instead of gray intensity or its gradient, the morphology measure of the regions works as a stop criterion of the curve to avoid the inhomogeneity caused by gray intensity. This method was applied to both real and synthetic images.
Vessel tracking methods simply track the center line of a vessel. The authors in [22] presented a blood vessel segmentation method based on the Gabor wavelet and line detector. This method consists of image preprocessing, candidate blood vessel extraction, and post-processing. In image preprocessing, Gabor wavelet was applied at a single scale to enhance the tiny vessels and edges of the wide vessels. They then added the complemented green channel to obtain a better contrast. Candidate blood vessel extraction is based on a Multiscale Line detector to overcome the central reflection problem. Finally, the median filter was used for post-processing. This method was performed on the DRIVE, STARE, and HRF datasets and it achieved accuracies of 0.9470, 0.9472, and 0.9559, and sensitivities of 0.7421, 0.8004, and 0.7207, respectively.
In ML, neural networks are used to simulate the human brain. A large number of neurons are connected to each other, and each neuron is considered a small information processing unit. The authors in [15] developed an unsupervised ML method for retinal vessel segmentation. In pre-preprocessing, they applied the particle swarm optimization (PSO) algorithm as an optimizer of contrast limited adaptive histogram equalization (CLAHE) to enhance the contrast of the retinal fundus images. Top-hat transformation on inverted CLAHE images and Wiener filters were used to remove the image noise. They applied Otsu global thresholding to the transformed image and thick vessels to obtain binary images. ISODATA local thresholding was applied to a single enhanced image to obtain a thin vessel binary image. The final stage removes the non-vessel components from the input binary images. This method was performed on the CHASE-DB1 and DRIVE datasets, achieving sensitivities of 0.7776 and 0.7851 and accuracies of 0.9505 and 0.9559, respectively.
With flourishing artificial intelligence, deep neural networks (DNN) have become a research trend that surpasses traditional segmentation methods in terms of accuracy. Convolutional neural network (CNN) is a model generated by the integration of image processing and DL, which has achieved great success in processing 2D or 3D images of diseased organs [23]. The CNN comprises several layers, and each layer has its own importance and functionality. The authors in [24] proposed a method based on a CNN that aims to detect tiny vessels from low contrast images. Pre-processing was applied to overcome uneven illumination and noise, and postprocessing was applied to remove the background noise. Preprocessing and postprocessing were applied to the network to remove extra noise to improve the sensitivity. This method was performed on the DRIVE and STARE datasets, achieving sensitivities of 0.746 and 0.748, and accuracies of 0.946 and 0.948, respectively. DL based medical image segmentation has been performed using numerous architectures, including fully convolutional networks (FCN), U-Net, generative adversarial networks (GNAs), convolutional residual networks (CRNs), and recurrent neural networks (RNNs). In FCN architecture, all layers are fully convolutional, which allows the model to perform a pixel-wise prediction for the entire image in just one forward pass [25]. In this study [26], the authors proposed a multiscale, multi-path, and multi-output fusion FCN (M3FCN). They applied four methods of data preprocessing: grayscale, CLAHE, normalization, and gamma correction. Moreover, they used an overlapping patch reconstruction algorithm to reconstruct the patch segmentation results into a final segmentation. This method achieved F1 scores of 0.8321, 0.8531, and 0.8243, with average accuracies of 0.9706, 0.9777, and 0.9773 on the DRIVE, STARE, and CHASE-DB1 datasets respectively.
The authors in [27] proposed U-Net, which was built based on an FCN and has been applied in many studies for medical image segmentation. Shortcut connections between layers are the most important property of U-Net, which provides high resolution features to the de-convolution layers. The U-Net is capable of fast and accurate image segmentation. This paper [28] proposed a ladder-net based on U-Nets. This method has multiple pairs of encoder-decoder branches and uses a shared-weight residual block, while the shared weights greatly reduce the number of parameters. This method achieved a sensitivity, specificity, accuracy, and AUC performance of (0.7856, 0.9810, 0.9561, and 0.9793), respectively on the DRIVE and (0.7978, 0.9818, 0.9656, and 0.9839), respectively on the CHASEDB. The authors in [29] provided a novel method based on dense U-net and a patchbased learning strategy that achieved a few parameters. This method achieved a sensitivity, specificity, accuracy, and AUC performance of (0.7986, 0.9736, 0.9511, and 0.9740), respectively on the DRIVE dataset and (0.7914, 0.9722, 0.9538, and 0.9704), respectively on the STARE dataset.
GANs have recently been applied by the authors in [23] which were important image segmentation models. Simply, it is through training to make two networks compete by generating fake data, and then using a discriminator to determine authenticity. In [30], the authors proposed a multi-label DCNN model (GL-Net). It combines GNAs and enables automatic segmentation of OD and OC. This model consists of two network structures: generator and discriminator. The authors added the L1 distance function and cross-entropy function to increase the accuracy of the segmentation result.
CRNs [31] solve the degradation of accuracy in the depth of the network by allowing the network to redirect derivatives through the network by skipping some layers. In [6], the authors proposed a supervised DL model called RCED-Net for retinal vessel segmentation. RCED-Net used the residual connection between the encoder and decoder based deep convolutional architecture known as SegNet [32]. They applied a special skip connection such that vital edge information would not be lost in the final prediction. This model reduced the training time and the number of tunable hyperparameters. In addition, it was designed to save only max pooling, which enhanced memory efficiency and reduced the testing time. This method achieved a sensitivity, specificity, accuracy, and ROC performance of (0.8252, 0.8440, and 0.8397), RNNs are simply the output going back to the input through a feedback connection, it can memorize the patterns VOLUME 11, 2023 from the last inputs and the inputs should be vectorized, long short-term memory (LSTM) [33] is a well-known type of RNN. This idea is considered a disadvantage for medical image segmentation because of the loss of spatial information. Hence, convolutional LSTM (CLSTM) [34] is a good application for solving this limitation by applying the convolutional operation instead of vector multiplication.
All previous studies achieved astounding results in retinal vessel extract segmentation, which signifies their deployment in diagnostic systems for eye diseases. Regardless of the improved accuracy of DL based supervised approaches, it is considered to have the highest computational complexity owing to the long training process. Numerous critical aspects still exist that require significant attention from researchers, including computational complexity, hyperparameters, and improved segmentation performance. In the current study, we propose a DL model for retinal vessel segmentation attributes with few trainable parameters and computational complexity with improved segmentation performance, making it attractive for deployment on low-end hardware devices.

III. MATERIALS AND METHOD A. MATERIALS
The authors in [13] briefly describe retinal vessel datasets that are available to researchers for analysis and comparison with previous best works from the literature such as structured analysis of the retina (STARE), digital retinal images for vessel extraction (DRIVE), Child Health and Heart Studies in England (CHASE), high-resolution fundus images (HRF), automatic retinal image analysis (ARIA), and ImageRet. STARE dataset consists of 20 vessel images, including normal and different types of retinal diseases. DRIVE dataset consists of 40 retinal images, including normal and DR images. CHASE dataset contained 28 pictures of 14 children from the left and right eyes. HRF database consists of 45 images, including normal, glaucomatous, and DR images. ImageRet for DR contains 219 retinal images, which are divided into two sets DIARETDB0 and DIARETDB1. ARIA has three types of disease images, which are explained in Table 1.
The performance of the model was assessed using three open-access retinal image databases: DRIVE, CHASE_DB1, and STARE. Both DRIVE and CHASE_DB have separate training and testing. The STARE dataset contains only one dataset used for training and testing purposes. For the STARE dataset, we applied ''leave-one-out'' [18] where the model is trained on 'n-1' samples, then what has been trained is tested using one remaining sample.

B. METHOD: ColonSegNet
The authors in [14] developed the ColonSegNet model and applied it to cancer diagnosis. ColonSegNet can detect and localize polyps at 180 fps. This model attributes a few trainable parameters, which make it preferable for embedded AI applications. We aimed to develop a lightweight model for accurate retinal vessel segmentation, which can be used for the early diagnosis of eye diseases at the point of care in hospitals. To the best of our knowledge, this is the first study to use ColonSegNet for retinal vessel segmentation and has achieved significantly improved evaluation metrics in addition to maintaining the reduced computational complexity of the model. We retrained the ColonSegNet model on retinal vessel datasets DRIVE, CHASE_DB1, and STARE.
ColonSegNet architecture has a lightweight network which leads to real-time performance because it is computationally efficient. It is an encoder-decoder architecture that produces segmentation of colonoscopic images in real-time. Figure 7 shows a block diagram of ColonSegNet, which uses a residual block with a squeeze and excitation network. The encoder parts learn how to extract information from the input which then passes to the decoder parts which use this information to generate the segmentation mask. ColonSegNet consists of two encoder blocks and two decoder blocks. Both the first and second encoders consisted of two residual blocks and a 3×3 strided convolution between them. In addition, the first encoder was followed by 2 × 2 max-pooling. A transpose convolution is used at the start of each decoder, which increases the feature map spatial dimensions. Furthermore, each decoder block contains two skip connections from the encoder block. The first skip connection directly joined to the concatenation, and the second skip connection incorporated multi-scale features in the decoder through a transpose convolution. These multi-scale features generate more semantic information in the form of a segmentation mask.

C. IMPLEMENTATION DETAILS
DL techniques have improved segmentation accuracy, and typically require many annotated samples to obtain reliable and well-performing models. Data collection and annotation are manual processes that are difficult and require a longer duration from experts. The most common technique used to increase the number of training samples in a dataset is data augmentation, which involves a set of affine transformations [1]. For our DL models, we performed data augmentation to enhance robust learning, such as horizontal and vertical flips on each training sample. Additionally, we applied ZCA whitening to training images, rotation from 0 to 180-degree, rescaling factor, and shear intensity.  For normalization, we divided the image by the standard deviation.
Retinal images are acquired using a high-resolution fundus camera, and this results in poor contrast between the background and retinal vascular structures. Thus, before accurate segmentation of the retinal vasculature, appropriate pre-processing techniques must be performed. We have applied CLAHE to improve contrast based on the local context of the image. The fundus image is an RGB color image consisting of three channels (red, green, and blue). Extracting the blood vessels using a color fundus image can be accomplished by separating the retinal image into three channels and using only one of them. In this work, we used the blue channel image because it provides contrast between the background and blood vessels to identify the central reflex in the vessels. In this study, we trained the model from scratch for 200 epochs, and a batch size of 4. We used the stochastic gradient descent (SGD) method to optimize our model and set the learning rate to 0.005. For the training stage, Table 2 shows the hyper-parameter settings for the proposed model with other best models.

A. EVALUATION METRICS
Vessel segmentation models are used to distinguish between the vessels and background in retinal fundus images.
The performance of the proposed method is evaluated by manually marked ''ground truth'' images. The four parameters listed below were used to compare the performance of the proposed method. 1) False negative (FN): The vessels that are detected as non-vessels by the classifier. 2) True positive (TP): The vessels that are detected (correctly) as vessels by the classifier. 3) False Positive (FP): Non-vessels that are detected incorrectly as vessels by the classifier. 4) True negative (TN): The non-vessels that are detected correctly as non-vessels by the classifier. The parameters provided above were used to assess the four main evaluation metrics of the developed approach.
Equations (1), (2), (3), and (4) denote accuracy, sensitivity, specificity, and MCC, respectively. Acc represents the accuracy, which is the ratio of correctly identified pixels (non-vessels or vessels) to the total number of pixels. Se and Sp represent the sensitivity and specificity, respectively, which measure the detection ratio of the vessel versus nonvessel. Matthew's correlation coefficient (MCC) is a balanced accuracy metric. It is well suited for class imbalance problems, such as retinal vessel segmentation. It produces a high score only if the prediction obtains better segmentation for all four confusion matrix categories (TP, FN, TN, and FP). Additionally, we adopted the area under each curve (AUC) as the fifth evaluation metric for image segmentation, ranging from 0 to 1. The AUC represents the area under (the receiver operating characteristic (ROC)) curve.

B. THE QUANTITATIVE PERFORMANCE COMPARISON OF OUR PROPOSED MODEL WITH STATE-OF
To assess the efficiency of our method, we performed simulations on three well-known open-access databases: STARE, DRIVE, and CHASE_DB1, and compared the obtained results with those of existing state-of-the-art models, which mentioned in Tables 3, 4, and 5 were taken from their respective papers. The three best values in each column of the table are highlighted in green, blue, and red, which denote the best, second best, and third best results, respectively.
As shown in Table 3, after training and testing the proposed method on the DRIVE database, we achieved 0.8491, 0.9774, 0.9659, 0.9850, and 0.7960 for sensitivity, specificity, accuracy, AUC, and MCC, respectively. The sensitivity of 0.9382 [35] is the best, but it is important to mention that the VOLUME 11, 2023   specificity and accuracy of this method are much lower than those of our method. The specificity, accuracy, and AUC of [36] are the best. The sensitivity, accuracy, and AUC of our method are the second best followed by RCED-Net comes in third place in terms of sensitivity and accuracy while the AUC of Ladder-Net [28] lies in the third place. Our obtained specificity is lower than the specificity of [37] and [38] which are the second best and third best among supervised methods, respectively.
For the CHASE_DB1 database, the results presented in Table 4 apear that the proposed approach achieved 0.8607, 0.9806, 0.9731, 0.9869, and 0.8014 for Se, Sp, Acc, AUC, and MCC, respectively. It can be observed that the accuracy and AUC of the proposed method are the best among the supervised approaches. The sensitivity of the proposed method lies in the second place closely behind [35], which achieved the best sensitivity. Our obtained specificity comes in the fifth place after the specificity of [36], and [39], ladder-Net [28], and RCED-Net [6], which are the best, second, third, and fourth best, respectively. Table 5 shows that our proposed method on STARE dataset achieved 0.8573, 0.9813, 0.9719, 0.9873, and 0.8069 for Se, Sp, Acc, AUC, and MCC respectively. The sensitivity and accuracy of our proposed method are the best among all the existing state-of-the-art methods. The AUC is the third best close behind [27] and [43] which are the best and the second best. Our obtained specificity is a little bit lower than the specificity of [27], [38], and [43] U-Net which are the best, the second best, and the third best. The results of MCC from the three tables show that our method has the highest MCC on DRIVE, STARE, and CHASE_DB1 datasets when compared with other supervised strategies. Further, Figure 8, gives the ROC curves for the DRIVE, STARE, and CHASE_DB1 datasets.
Additionally, Table 6 considers the number of parameters, ladder-Net [28] and dense U-Net [29] have the lowest number of parameters, which are the best and second best, respectively. Our proposed method is computationally efficient and becomes the third best with deployment possible on low-end hardware devices. Furthermore, RCED-Net is the fourth best with 9.37M parameters. The performance of the proposed model has been thoroughly assessed on both normal and pathological images, and highly convincing results based on specificity, sensitivity, accuracy, and AUC have been achieved compared to ladder-Net [28] and dense U-Net [29]. It is reasonable to say that the proposed method is much  better than ladder-Net [28] and dense U-Net [29] because of its superiority in evaluation metrics and competitive computational complexity.

C. THE QUALITATIVE PERFORMANCE COMPARISON OF OUR PROPOSED MODEL WITH SegNet-Basic and RCED-Net
To confirm our hypothesis, we qualitatively evaluated the performance of the proposed model. Figure 9 shows a visual comparison of the proposed model based on the three  segmented images from the DRIVE database with those obtained using SegNet-Basic and RCED-Net. The second and third rows indicate the output of SegNet-Basic and RECD-Net, respectively, whereas the fourth row shows our obtained results. Figure 9 shows that the proposed model can detect small vessels. On the other hand, SegNet-Basic could also discover small vessels, but there were some noises. The obtained results of our model were much better than those of SegNet-Basic and RCED-Net.
Moreover, Figure 10 shows the results of our proposed method on two noisy images, which are images number one and two of the STARE dataset. The second and third columns present the segmentation results from SegNet-Basic and RCED-Net, respectively, while the fourth column shows our obtained results. From the obtained images, we found that our model did not perform well for noisy images because the edge information of the vessels could not be maintained with background noise. Figure 11 shows a comparison of the proposed method on three images of the CHASE_DB1 dataset with UP and RECD-Net. It is to be mentioned that the first and second columns present the original image and its ground truth. The third and fourth columns indicate the outputs of the UP [37] and RECD-Net, respectively. Finally, column five presents the resultant images obtained using the developed model. Based on our observations of these images, it can be said that our developed model is quite precise in segmenting vessels in retinal images, even in pathological images.

V. CONCLUSION
Retinal fundus images have been used for training and testing DNN models to develop automated diagnostic tools for eye disease diagnosis at an early stage. There are several challenges in diagnosing eye diseases based on retinal fundus imaging. Most of the previous state-of-the-art studies tried to develop DNN models to deal with some of the challenges. Most previous studies have focused on enhancing evaluation metrics. Computational complexity is an important factor, especially if a diagnostic tool is designed for large-scale screening programs. In the literature, very little attention has been given to reducing the computational complexity of the developed model. Reduced computational complexity is highly desired, as the aim of developing an AI-enabled automated diagnostic system is to use for population scale (large-scale) screening programs, which will be possible if such tools are used at the front disks in a general hospital.
This inspired us to propose a research work in which any general lightweight DNN model could be used for vessel segmentation in retinal images as a proof of concept. We applied a lightweight DNN model for retinal vessel segmentation, based on the ColonSegNet model. We retrained the ColonSegNet model using retinal vessel datasets. The performance of the proposed model achieved a highly convincing result based on sensitivity, accuracy, and AUC compared with other state-of-the-art lightweight and heavy methods, such as dense U-Net, Ladder-Net, RCED-Net, and Basic SegNet. The proposed model is an appropriate choice for deployment in the computationally constrained computing facility at the point of care due to the advantages of being highly robust, reliable, and efficient in terms of segmentation accuracy in addition to being lightweight.

VI. ACKNOWLEDGMENT
RASHA SARHAN ALHARTHI received the B.S. degree in computer engineering from Taif University (TU), Saudi Arabia, in 2017. Currently, she is pursuing the master's degree in computer engineering with King Saud University (KSU), Saudi Arabia. Her research interests include embedded systems, machine/deep learning, healthcare, and image processing. MUSAED ALHUSSEIN received the B.S. degree in computer engineering from King Saud University, Riyadh, in 1988, and the M.S. and Ph.D. degrees in computer science and engineering from the University of South Florida, Tampa, FL, USA, in1992 and 1997, respectively. He is a Professor with the Department of Computer Engineering, College of Computer and Information Sciences, King Saud University (KSU), Riyadh, Saudi Arabia. Since 1997, he has been the Faculty Member with the Computer Engineering Department, College of Computer and Information Science, King Saud University. Recently, he has been successful in winning a research project in the area of AI for healthcare, which is funded by Ministry of Education at Saudi Arabia. He is the Founder and the Director of Embedded Computing and Signal Processing Research (ECASP) Laboratory. His research interests include computer architecture and signal processing with an emphasis on big data, machine/deep learning, VLSI testing and verification, embedded and pervasive computing, cyberphysical systems, mobile cloud computing, big data, healthcare, and body area networks.