DAVS-NET: Dense Aggregation Vessel Segmentation Network for retinal vasculature detection in fundus images

Mohsin Raza; Khuram Naveed; Awais Akram; Nema Salem; Amir Afaq; Hussain Ahmad Madni; Mohammad A. U. Khan; Mui-zzud- din

doi:10.1371/journal.pone.0261698

Abstract

In this era, deep learning-based medical image analysis has become a reliable source in assisting medical practitioners for various retinal disease diagnosis like hypertension, diabetic retinopathy (DR), arteriosclerosis glaucoma, and macular edema etc. Among these retinal diseases, DR can lead to vision detachment in diabetic patients which cause swelling of these retinal blood vessels or even can create new vessels. This creation or the new vessels and swelling can be analyzed as biomarker for screening and analysis of DR. Deep learning-based semantic segmentation of these vessels can be an effective tool to detect changes in retinal vasculature for diagnostic purposes. This segmentation task becomes challenging because of the low-quality retinal images with different image acquisition conditions, and intensity variations. Existing retinal blood vessels segmentation methods require a large number of trainable parameters for training of their networks. This paper introduces a novel Dense Aggregation Vessel Segmentation Network (DAVS-Net), which can achieve high segmentation performance with only a few trainable parameters. For faster convergence, this network uses an encoder-decoder framework in which edge information is transferred from the first layers of the encoder to the last layer of the decoder. Performance of the proposed network is evaluated on publicly available retinal blood vessels datasets of DRIVE, CHASE_DB1, and STARE. Proposed method achieved state-of-the-art segmentation accuracy using a few number of trainable parameters.

Citation: Raza M, Naveed K, Akram A, Salem N, Afaq A, Madni HA, et al. (2021) DAVS-NET: Dense Aggregation Vessel Segmentation Network for retinal vasculature detection in fundus images. PLoS ONE 16(12): e0261698. https://doi.org/10.1371/journal.pone.0261698

Editor: Le Hoang Son, Vietnam National University, VIET NAM

Received: June 9, 2021; Accepted: December 7, 2021; Published: December 31, 2021

Copyright: © 2021 Raza et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: Datasets used in this study are publicly available on following links: DRIVE: https://drive.grand-challenge.org; STARE: https://cecas.clemson.edu/~ahoover/stare/probing/index.html; CHASE: https://blogs.kingston.ac.uk/retinal/chasedb1.

Funding: The author(s) received no specific funding for this work.

Competing interests: The authors have declared that no competing interests exist.

1 Introduction

Early detection of potential blindness diseases is vital to treat their progression and avoid vision loss, for instance, Aging based Mocular Degeneration (AMD), Diabetic Retinopathy (DR) and Hypertension Retinopathy (HR) [1]. Similarly, timely detection of Hypoxemia and Glaucoma is useful for availing cost effective remedies. It is widely understood that these diseases impact the structure of retinal blood vessels [2]. Therefore, clinicians diagnose these diseases by observing the visible changes in the structure of blood vessels in retinal images [3, 4]. That is a cumbersome process and hence is not practically viable to perform on a larger scale owing to the limitation of skilled labour and timing consuming nature of the process.

Consequently, Computer-aided diagnostic (CAD) systems have taken a deep root in eye diagnosis owing to their fast processing and ability to scan through large datasets of fundus images [5–7]. These computerized techniques start by employing segmentation strategies to extract patterns of blood vessels [8, 9]. That is followed by the use of automated classifiers to evaluate and analyze the extracted vessels for detection of variations in the characteristics of blood vessels [10]. Thus, leading to automated diagnosis of the eye. In this regard, the role of computerized vessel segmentation strategies is vital because the classifier’s effectiveness in eye disease highly depends on the accuracy of the segmented vessels [11, 12].

Retinal vessel segmentation has attracted significant attention from engineers and scientists, resulting in a wide range of state of the art methods [13–19]. However, effective segmentation of retinal vessels is still an open problem due to various challenges which involve sharp variations in vessel size, shape, and orientation, not to mention the low intensity, branching, and vessel crossovers. Consequently, identification of vessels and differentiating those from irregularities (arising due to a disease or other similar phenomenon) is a difficult task. That is further aggravated by the presence of various types of noise and artifacts due to fundus imaging modalities.

Earlier, classical image segmentation strategies were tailored to detect and segment out vessel patterns. These techniques identify vessels based on width, size, shape and orientation of vessels and hence are referred to as unsupervised methods [14–16, 20–22]. However, these methods can only capture limited types of vessels due to sharp variations in their shapes and sizes. Moreover, these techniques can not fully comprehend and eradicate the problem of low illumination and poor contrast regions in retinal fundus images. Although, contrast enhancement techniques are used as a pre-processing step that partially address the issue but they intensify the noise or artifacts present in the image [11, 23] which led to the use of noise removal as an additional pre-processing step in some recent unsupervised methods [24, 25].

Supervised methods, on the other hand, use trained Support Vector Machines (SVM) [10, 26] and Neural Networks (NNs) [27, 28] to identify vessels based on learned features from fundus images. Compared to SVM, NN can model the interrelationship between features in a much better way that has led to their increased use in this regard.

Deep learning techniques, which employ multi-layered NNs, have particularly yielded much higher rates of accuracy albeit at high computational cost [29–31]. Traditional DNNs do well to learn the inherent structures within the image that allow them to recover the structure of vessels in a much better way when compared with classical techniques. Deep Neural Networks (DNNs) have the ability to learn inherent and deep structures within the retinal images from a large sized fundus image dataset, allowing the detection of fine vessels [32, 33]. For this purpose, deep learning based techniques employ CNNs to extract desirable features which are finally used to identify vessels. Moreover, deep features allow these techniques to move past the problem of noise and artifacts. However, the problem with these methods is their lack of robustness when detecting less significant or minor vessels. This problem is due to the loss of important information due to pooling operations that restrict their efficacy. Consequently, recent vessel segmentation techniques employ semantic segmentation, where each pixel is classified as a vessel or the background. That provides the high precision needed to detect tiny vessels, such as vessels consisting of only few pixels.

This work proposes a novel network architecture, namely Dense Aggregation Vessel Segmentation network (DAVS-Net), for robust semantic segmentation of retinal vessels that is capable of detecting minor vessels owing to its pixel wise segmentation operation. The proposed architecture employs dense concatenation block that permits immediate transfer of spatial information within layers leading to the identification of pixels from the desired class. In addition, we propose an encoder-decoder framework that allows faster convergence by directly transferring the edge information from initial layer of the encoder to the last layer of the decoder. Moreover, the proposed network requires only a few trainable parameters as apposed to a large number of trainable parameters required in existing methods because of low-quality retinal images with different image acquisition conditions and intensity variations. The proposed DAVS-Net achives state of the art performance that is demonstrated publicly available retinal blood vessels datasets of DRIVE, CHASE_DB1, and STARE.

This paper is organized to provide background of the problem in Section II post the introduction in Section I. The proposed methodology is discussed in Section III followed by the results and discussion in Section IV. Finally, conclusions and scope for future work are discussed in Section V.

2 Background and related work

Semantic segmentation is regarded as a fundamental application in computer vision where pixel-wise classification is performed for all the pixels present in the image. This approach is able differentiate between pixels belonging to objects and those belonging to the background leading to the detection of tiniest objects. Consequently, semantic segmentation is well suited for retinal vessel segmentation since detection of tiniest of vessels is vital for analysis and diagnosis of retinal disease.

The conventional deep learning-based methods [34] effectively learn structures of significant objects but lack robustness to identify the minor ones. Thus, the DNNs used for segmentation are not local enough in their operation and as a consequence, they do not classify each pixel for detection of a vessel leading to loss of minor and tiny vessels. Deep networks for vessel detection use many convolutional and pooling layers which cause vanishing gradient problems. This loss of spatial information degrades the overall performance of pixel-wise classification. To overcome the vanishing gradient problem residual networks (Res-Nets) [35] were introduced that used the residual skip connection to improve the performance and manage the gradient during the training process. However, Res-Nets caused the feature transfer impedance problem that was later covered by Dense-Net [36] through deep feature concatenation.

Another factor affecting the segmentation of tiny vessels is the compromised quality of fundus images typically caused by the limitation of varying acquisition conditions. Hence, robust segmentation of retinal vessels is an open problem with a focus on the detection of minor vessels which provide critical additional information for automated eye diagnosis.

3 Proposed methodology

In this work, we propose DAVS-Net architecture for robust semantic segmentation of retinal vessels from the fundus image by effectively capturing the minor vessels along with the significant ones. Proposed architecture seeks to address limitations of traditional deep learning techniques which employ a number convolutional layers followed by pooling operations that means local information about each and every pixel is not readily available. As a result, these networks work well to detect significant vessels although it means that identification of minor vessels becomes challenging. This issue needs special attention as detection of smaller vessels is critical to accurate eye disease diagnoses.

To address this issue, proposed dense aggregation network, for semantic segmentation of retinal vasculature, feeds on the desirable properties of the DenseNet [36] that is famous for its classification performance. That is because of the use of the dense concatenation which alleviates the feature latency problems and provide higher accuracy compared to ResNet [35], and VGG [37]. Considering the effectiveness of feature concatenation benefits, the connectivity of the DAVS-Net is inspired from Dense-Net. The key differences of proposed DAVS-Net and DenseNet are mentioned in Table 1.

Download:

Table 1. Difference between DenseNet and proposed DAVS-Net.

https://doi.org/10.1371/journal.pone.0261698.t001

3.1 Overview of proposed architecture

The proposed DAVS-Net is designed to take advantage of the deep feature that allows to skip the pre-processing and does not require any enhancement in the quality of the input image. That is because the deep feature allows to allows to import and combine high-frequency information from the corresponding layers thus circumventing the imaging artifacts and bring to light the main features of the image. Owing to that, DAVS-Net is capable of detecting vessel pixels from noisy and low-quality images and non-uniform illumination. The overall principle of the proposed method is summarized in Fig 1. Moreover, the pixel wise segmentation operation and the marking of blood vessels yields the much needed accuracy for vessel detection. The output of the proposed method is a binary image with a representation of ‘1’ for vessel pixels and ‘0’ for the background.

Download:

Fig 1. Flow diagram of the proposed method.

https://doi.org/10.1371/journal.pone.0261698.g001

3.2 Working principle of the DAVS-Net

Proposed DAVS-Net considers dense connections as a means to boost accuracy of the semantic segmentation. To this end, the problems faced by traditional deep learning techniques are addressed using its following key features:

Fewer convolutional layers and pooling layers are used to reduce the spatial information loss.
Dense concatenation of the features is used within the dense block to enable the network for providing immediate spatial information transfer between the layers.
The edge information transfer from the first layers of the encoder to the last layer of the decoder is used for faster convergence of the network.

The connectivity principle of DAVS-Net is demonstrated in Fig 2 that presents the layout of the deep feature concatenation for the candidate encoder-decoder block.

Download:

Fig 2. Connectivity principle of DAVS-Net.

https://doi.org/10.1371/journal.pone.0261698.g002

The encoder consists of three dense blocks with two convolutional layers in each block. Similar structure is used for the decoder as well. We describe both encoder and decoder in detail in Section 3.3 and 3.4. Here, we discuss the connectivity of principle of the proposed DAVS-Net (as given in Fig 2) that leads to formulation of deep feature.

Specifically, the dense block of the encoder, shown on the left side of Fig 2, receives an input feature F_i while the dense block of the decoder, depicted on the right side of Fig 2, receives an input feature U_i. The feature is obtained after two convolutional operations, namely E-Conv-A and E-Conv-B. The spatial loss is recovered by deep feature concatenation of these two convolutional layers. The dense feature is obtained by concatenating the feature of the outputs and of E-Conv-A and E-Con-B, as given below: (1) where ‘*’ denotes the depth-wise concatenation.

We next employ a bottleneck layer, termed Bottle-Neck, to limit number of channels after a Batch Normalization (BN) and a Rectified Linear Unit (ReLU) operations that results in the feature , as follows (2)

Similarly, the decoder applies a convolution on the input U_i through the convolutional layer D-Conv-A resulting in feature . This feature U_i is then fed to the second convolutional layer D-Conv-B resulting in the feature . The spatial loss is recovered by concatenating the deep feature from these two convolution layers and the third feature that comes from the encoder by an external dense path. Thus, the dense feature is an enriched feature by the concatenation of three features of the outputs D-Conv-A, D-Conv-B, and E-Conv-A as given below: (3)

Just like in encoder, the increase in the number of channels for feature may lead to memory consumption that is resolved through the Bottle-Neck layer after BN and ReLU operations yielding to the feature , as follows: (4)

Now, comparing both and , although both are empowered features by dense connectivity but the is the resultant feature of three features concatenation which also includes the important edge information. Owing to that enrichment, the proposed DAVS-Net is able to perform finer segmentation without any prior need for preprocessing. To ensure the segmentation of small objects, feature enhancement is done at dense block level as shown in Fig 3, that presents the complete architecture with the dense feature concatenation. The DAVS-Net is keeping the feature map size before upsampling at 80 × 80 for an input image of 640×640, that is enough to represent the valuable features for vessel segmentation.

Download:

Fig 3. Architecture of DAVS-Net used for vessel segmentation in our work.

https://doi.org/10.1371/journal.pone.0261698.g003

3.3 DAVS-Net encoder

DAVS-Net is a densely connected fully convolutional network that uses a total of 6 dense blocks for both encoder and decoder as shown in Fig 3. The encoder consists of three dense blocks with each block containing two convolutional layers. Each encoder dense block starts with a convolutional layer and ends with a pooling layer that is used to reduce the size of the feature map. As an example, the first encoder dense blocks is with two convolutions of 64 channels, and the output of both convolutions are merged by a depth-wise concatenation layer generating 128 channels.

The concatenation layer leads to increase in depth of the feature map that requires more memory as well as processing power. The issue is addressed through bottleneck layer that reduces memory consumption by selecting higher minibatch size in each dense block which results in limiting the channels after the concatenation. Moreover, a constant convolution operation is required to segment the image using a convolutional neural network (CNN). Consequently, the DAVS-Net encoder performs the constant convolutional operation on the image and the feature. That travels through the network in a feed-forward fashion until the image is represented by the tiny features.

Another problem with CNN is that max-pooling operation (post convolution) causes spatial information loss. In DAVS-Net, loss of the useful information is covered by the deep feature concatenation. Thus, in the proposed architecture, the encoder is composed of three dense block with 6 convolution layers and three max-Pool layers and the final feature map is 80×80 for a 640×640 input image. The DAVS-Net encoder structure in terms of the dense block is listed in Table 2, which describes the feature empowerment inside each encoder dense block and shows how the bottleneck layer reduces the depth of the feature map. The number of trainable parameters is also shown in the table for the layers in the encoder.

Download:

Table 2. DAVS-Net encoder-decoder I/O feature map sizes.

Where EDB, EDB-C, EDB-Cat, DDB, DDB-C, DDB-Cat represent encoder dense block, encoder dense block convolution, encoder dense block concatenation, decoder dense block, decoder dense block convolution, decoder dense block concatenation, respectively. The layer with shows that layer includes rectified linear unit (ReLU), and batch normalization (BN) after.

https://doi.org/10.1371/journal.pone.0261698.t002

3.4 DAVS-Net decoder

The decoder in DAVS-Net employs the reverse operation to the encoder as shown in Fig 3, whereby each dense block starts with an Max-Unpool layer which is responsible for gradually increasing the size of the feature map. After each unpooling layer, two convolutions follow the same concatenation and bottleneck principle. The depth-wise concatenation layer in each decoder block receives three inputs from first convolution, second convolution and direct information from the outer dense connection of the respective encoder block. The outer dense paths start from the first convolutional layer of the encoder dense block and terminate at the concatenation layer of each decoder dense block. These outer dense paths provide the immediate edge information from encoder to decoder to reduce the latency.

Specifically, the DAVS-Net decoder receives an input of 80×80-pixel from the encoder and provides the final feature map of the size equal to the size of input image. The bottleneck layer in each decoder block is used to reduce the number of channels to avoid memory issues. The last bottleneck layer in the decoder (third decoder dense block) is responsible for reducing the depth of the feature map. That also works as a class mask layer whose number of channels is equal to the number of classes.

This study is based on two classes “Vess” and “BG” representing vessel pixel and backgroud pixels; therefore, the number of channels in the last bottleneck layer is set at 2. The DAVS-Net pixel classification layer in combination with ‘Softmax’ function is responsible to assign a label to each pixel in the image from the available class based on prediction. Table 2 provide the layer layout of the DAVS-Net decoder with respective feature map sizes.

Table 3 presents the architectural differences of the proposed method with similar state-of-the-art networks. That essentially demonstrates that proposed architecture requires less convolution operations reduced channel depth as compared to some of other state of the architectures. Additionally, we also utilize dense connectivity, unpooling and bottleneck layers to further enhance the architecture of the proposed DAVS-Net over the comparative state of the art techniques.

Download:

Table 3. Comparison of architectural differences with similar state-of-the-art networks.

https://doi.org/10.1371/journal.pone.0261698.t003

4 Detection of diabetic and hypertensive retinopathy

It is mentioned in [41] that both diabetic and hypertensive retinopathy cause changes in retinal vessels. The diabetic retinopathy can swell the retinal vessels or even can create new blood vessels (increase in the vessel pixels), where the hypertensive retinopathy causes the shrinkage of retinal blood vessels (decrease in number of vessel pixels). The accurate segmentation of these vessels can provide an opportunity to detect changes in the retinal vessels (increase or decrease in number of vessels). This increase or decrease in number of vessel pixels can be used for diagnostic purposes for analysis of diabetic and hypertensive retinopathy. The disease progression can also be analyzed by comparing the masks of successive visits.

5 Experimental results

The experiments were conducted on a machine with Intel(R) Xeon(R) W-2133 CPU 3.60GHz processor, 96GB RAM, and Nvidia 2080TI GPU. For our implementation, the MATLAB was employed. We employed the ADAM optimizer with an initial learning rate of 1e⁻³, an exponential decay rate of 0.9, and mini-batch size of 10 images. The proposed DAVS-Net is trained from the scratch without weight initialization or migration from other frameworks. A weighted cross-entropy loss is used as an objective function for training in all of our experiments. This decision is based on the fact that the “background” pixels in each retinal image heavily outnumber the “foreground” pixels. We use median frequency balancing to calculate class association weights here [34].

Because the retinal vessel segmentation data sets used here are quite small in size, we used data augmentation to generate enough data for training. We used rotation and contrast enhancement to enhance the data. Each training image is rotated by 1 degree for the rotations. The contrast has been improved by randomly increasing and decreasing the image brightness. This results in 7600 images for the DRIVE and CHASE DB data sets, as well as 7000 images for each of the STARE data’s leave-one-out trails.

5.1 Materials

We have evaluated the performance of our proposed method on the basis of the following three fundus retinal image datasets which are publically available.

STARE: A group of twenty fundus images collected in the USA [41].
DRIVE: A collection of retinal images obtained from aged diabetic patients in Netherland [42].
CHASE_DB1: A collection of retinal fundus images based on fourteen pediatric subjects [43].

Segmentation of blood vessels is performed on retinal images in DRIVE dataset using manual procedure. In comparison of the three datasets, there is a binary mask revealing FOV for DRIVE dataset but it is not available for STARE and CHASE_DB1. For the STARE and CHASE_DB1 datasets, binary masks are manually generated by well-known techniques [44]. DRIVE and CHASE_DB1 have their individual and distinct training and testing datasets. In STARE, two subsets of randomly selected images are taken for training and testing purpose. As given in literature, a “leave-one-out” method is commonly implemented to separate training and testing sets [44]. In this method, a model is trained on ‘n-1’ samples and tested on the remaining sample to avoid overlapping. This process is iterated for ‘n’ times to complete the dataset, “leaving out” each sample at least once for the whole dataset. We have implemented this “leave-one-out” method for STARE dataset to train the model. Details of three selected datasets in our experiments are summarized in the Table 4.

Download:

Table 4. Summary of datasets used in the experiments.

https://doi.org/10.1371/journal.pone.0261698.t004

5.2 Evaluation criteria

Models for vessel segmentation are actually binary classifiers that necessarily differentiate vessels from the background for the given set of retinal fundus images. Performance of these segmentation classifiers is evaluated with “ground truth” images marked by ophthalmologists. Based on the following four parameters, we utilized the three metrics given in equations 5, 6, and 7 [36], for the performance evaluation of our proposed system.

True Negative (TN): Classifier correctly found as non-vessels,
False Positive (FP): Classifier incorrectly found vessels which are actually non-vessels,
True Positive (TP): Classifier correctly found as vessels,
False Negative (FN): Classifier incorrectly found non-vessels which are actually vessels.

(5) (6) (7) where Sp, Se, and Acc are representing the specificity, sensitivity, and accuracy, respectively. Accuracy is the ratio between correctly detected pixels (vessels and non-vessels) and the total pixels in the mask (FOV only). While specificity and Sensitivity demonstrate that how much accurately a model identifies the non-vessel and vessel pixels respectively. Furthermore, performance of the classifier is also assessed by some other parameters such as area under the Receiver Operating Characteristic (ROC), Area Under the Precision-Recall Curves (AUCPR), and False Positive Rate (FPR). Whenever, we have imbalanced distribution, ROC is a feasible assessing parameter for the classification of objects [45]. The AUC and AUCPR measures are used to analyze the objective efficiency of classification.

5.3 Comparison with state-of-the-art

The visual results of our simulation on the three datasets are shown in Figs 4–6, respectively. In each figure, moving from left to right, the first column shows the original images, the second column shows the ground truth images and the third column shows the segmented images.

Download:

Fig 4. Visual results on the CHASE_DB1 dataset.

From left-to-right: input images, ground truth, result obtained by our proposed method.

https://doi.org/10.1371/journal.pone.0261698.g004

Download:

Fig 5. Visual results on the DRIVE dataset.

From left-to-right: input images, ground truth, result obtained by our proposed method.

https://doi.org/10.1371/journal.pone.0261698.g005

Download:

Fig 6. Visual results on the STARE dataset.

From left-to-right: input images, ground truth, result obtained by our proposed method.

https://doi.org/10.1371/journal.pone.0261698.g006

To evaluate and compare our results with those of state-of-the-art models, we have presented and summarized the results in tabular forms. Results obtained by our simulation on CHASE_DB1 are compared in Table 5. As given in the table, dice and Jaccard Se, Sp and Accuracy of our models are 0.8144, 0.9843 and 0.9726 respectively.

Download:

Table 5. Performance comparison of our proposed model on CHASE_DB1 dataset with other existing models.

https://doi.org/10.1371/journal.pone.0261698.t005

In Table 6, the results of our proposed model, implemented on DRIVE dataset, are compared with those of state-of-the-art. Se, Sp and Accuracy of our model is 0.8286, 0.9824 and 0.9689 respectively.

Download:

Table 6. Performance comparison of our proposed model on DRIVE dataset with other existing models.

https://doi.org/10.1371/journal.pone.0261698.t006

Similarly, results achieved from the implementation of our model on STARE dataset are compared in Table 7. From this experiment, Se, Sp and Accuracy of our model are 0.8238, 0.9866 and 0.9744 respectively.

Download:

Table 7. Performance comparison of our proposed model on STARE database with other existing models.

https://doi.org/10.1371/journal.pone.0261698.t007

From the comparisons with state-of-the-art, it is obvious that our proposed model outperformed other existing models with respect to well-known evaluation metric accuracy on three well-known and publicly available datasets.

6 Conclusion

Diabetic retinopathy is one of the top ophthalmic diseases which lead to blindness in the diabetic patients. Accurate segmentation of retinal blood vessels significantly helps the ophthalmologist for screening and detection of diabetic retinopathy. In proceeding to the diagnosis of this disease, we proposed a segmentation network, DAVS-Net, for the segmentation of retinal blood vessels. Dense concatenation of features in the dense block enabled the network to acquire and transfer spatial information from the image. Fast convergence of the network is achieved through the edge information transfer from encoder layers to decoder layers. There are three main design attributes of DAVS-Net; Firstly, quality of features is improved by feature concatenation, whereas memory requirements are controlled by the bottleneck layers in dense block. Secondly, number of convolution layers is reduced in all six blocks of the network to minimize the spatial information loss. Thirdly, DAVS-Net employs dense paths for feature empowerment which aids in extraction of minor information from the image. We evaluated proposed network on three publicly available datasets and surpassed the existing state-of-the-art methods in terms of accuracy and computational efficiency. This method can be used as a second opinion system to aid medical doctors and ophthalmologists for the diagnosis and analysis of diabetic retinopathy. In the future, we will further increase the accuracy of blood vessels segmentation with consideration of other retinal diseases along.

References

1. Mohamed Q, Gillies MC, Wong TY. Management of diabetic retinopathy: a systematic review. Jama. 2007;298(8):902–916. pmid:17712074
- View Article
- PubMed/NCBI
- Google Scholar
2. Srinidhi CL, Aparna P, Rajan J. Recent advancements in retinal vessel segmentation. Journal of medical systems. 2017;41(4):70.
- View Article
- Google Scholar
3. Cheung CYl, Zheng Y, Hsu W, Lee ML, Lau QP, Mitchell P, et al. Retinal vascular tortuosity, blood pressure, and cardiovascular risk factors. Ophthalmology. 2011;118(5):812–818. pmid:21146228
- View Article
- PubMed/NCBI
- Google Scholar
4. Niemeijer M, Staal J, van Ginneken B, Loog M, Abramoff MD. Comparative study of retinal vessel segmentation methods on a new publicly available database. In: Medical imaging 2004: image processing. vol. 5370. International Society for Optics and Photonics; 2004. p. 648–656.
5. Soomro TA, Gao J, Khan T, Hani AFM, Khan MA, Paul M. Computerised approaches for the detection of diabetic retinopathy using retinal fundus images: a survey. Pattern Analysis and Applications. 2017;20(4):927–961.
- View Article
- Google Scholar
6. Ravudu M, Jain V, Kunda MMR. Review of image processing techniques for automatic detection of eye diseases. In: 2012 Sixth International Conference on Sensing Technology (ICST). IEEE; 2012. p. 320–325.
7. Irshad S, Akram MU. Classification of retinal vessels into arteries and veins for detection of hypertensive retinopathy. In: 2014 Cairo International Biomedical Engineering Conference (CIBEC). IEEE; 2014. p. 133–136.
8. Khawaja A, Khan TM, Khan MAU, Nawaz J. A Multi-Scale Directional Line Detector for Retinal Vessel Segmentation. Sensors. 2019;19(22). pmid:31766276
- View Article
- PubMed/NCBI
- Google Scholar
9. Khan MA, Khan TM, Naqvi SS, Khan MA. Ggm classifier with multi-scale line detectors for retinal vessel segmentation. Signal, Image and Video Processing. 2019;13(8):1667–1675.
- View Article
- Google Scholar
10. Wisaeng K, Hiransakolwong N, Pothiruk E. Automatic detection of retinal exudates using a support vector machine. Applied Medical Informatics. 2013;32(1):33–42.
- View Article
- Google Scholar
11. Soomro TA, Khan MA, Gao J, Khan TM, Paul M. Contrast normalization steps for increased sensitivity of a retinal image segmentation method. Signal, Image and Video Processing. 2017; 11(8):1509–1517.
- View Article
- Google Scholar
12. Khan TM, Bailey DG, Khan MA, Kong Y. Efficient hardware implementation for fingerprint image enhancement using anisotropic Gaussian filter. IEEE Transactions on Image processing. 2017;26(5):2116–2126. pmid:28237927
- View Article
- PubMed/NCBI
- Google Scholar
13. Khan MA, Soomro TA, Khan TM, Bailey DG, Gao J, Mir N. Automatic retinal vessel extraction algorithm based on contrast-sensitive schemes. In: 2016 International conference on image and vision computing New Zealand (IVCNZ). IEEE; 2016. p. 1–5.
14. Franklin SW, Rajan SE. Computerized screening of diabetic retinopathy employing blood vessel segmentation in retinal images. biocybernetics and biomedical engineering. 2014;34(2):117–124.
- View Article
- Google Scholar
15. Khan MA, Khan TM, Aziz KI, Ahmad SS, Mir N, Elbakush E. The use of fourier phase symmetry for thin vessel detection in retinal fundus images. In: 2019 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT). IEEE; 2019. p. 1–6.
16. Mehmood M, Khan TM, Khan MA, Naqvi SS, Alhalabi W. Vessel intensity profile uniformity improvement for retinal vessel segmentation. Procedia Computer Science. 2019;163:370–380.
- View Article
- Google Scholar
17. Khan MA, Khan TM, Bailey DG, Soomro TA. A generalized multi-scale line-detection method to boost retinal vessel segmentation sensitivity. Pattern Analysis and Applications. 2019;22(3):1177–1196.
- View Article
- Google Scholar
18. Soomro TA, Khan TM, Khan MA, Gao J, Paul M, Zheng L. Impact of ICA-based image enhancement technique on retinal blood vessels segmentation. IEEE Access. 2018;6:3524–3538.
- View Article
- Google Scholar
19. Khan MA, Khan TM, Soomro TA, Mir N, Gao J. Boosting sensitivity of a retinal vessel segmentation algorithm. Pattern Analysis and Applications. 2019;22(2):583–599.
- View Article
- Google Scholar
20. Fan Z, Lu J, Wei C, Huang H, Cai X, Chen X. A hierarchical image matting model for blood vessel segmentation in fundus images. IEEE Transactions on Image Processing. 2018;28(5):2367–2377. pmid:30571623
- View Article
- PubMed/NCBI
- Google Scholar
21. Soomro TA, Khan MA, Gao J, Khan TM, Paul M, Mir N. Automatic retinal vessel extraction algorithm. In: 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA). IEEE; 2016. p. 1–8.
22. Khan TM, Khan MA, Rehman NU, Naveed K, Afridi IU, Naqvi SS, et al. Width-wise vessel bifurcation for improved retinal vessel segmentation. Biomedical Signal Processing and Control. 2022;71:103169.
- View Article
- Google Scholar
23. Toufique Soomro MAUKTMKMPNM Junbin Gao. Role of Image Contrast Enhancement Technique for Ophthalmologist as a Diagnostic Tool for the Diabetic Retinopathy. In: IEEE International Conference on Digital Image Computing: Techniques and Applications (DICTA), At Gold Coast, Autralia; 2016. p. 1–8.
24. Khawaja A, Khan TM, Naveed K, Naqvi SS, Rehman NU, Nawaz SJ. An improved retinal vessel segmentation framework using frangi filter coupled with the probabilistic patch based denoiser. IEEE Access. 2019;7:164344–164361.
- View Article
- Google Scholar
25. Naveed K, Abdullah F, Madni HA, Khan MA, Khan TM, Naqvi SS. Towards Automated Eye Diagnosis: An Improved Retinal Vessel Segmentation Framework Using Ensemble Block Matching 3D Filter. Diagnostics. 2021;11(1):114. pmid:33445723
- View Article
- PubMed/NCBI
- Google Scholar
26. Tuba E, Mrkela L, Tuba M. Retinal blood vessel segmentation by support vector machine classification. In: 2017 27th International Conference Radioelektronika (RADIOELEKTRONIKA). IEEE; 2017. p. 1–6.
27. Khan TM, Abdullah F, Naqvi SS, Arsalan M, Khan MA. Shallow Vessel Segmentation Network for Automatic Retinal Vessel Segmentation. In: 2020 International Joint Conference on Neural Networks (IJCNN). IEEE; 2020. p. 1–7.
28. Khan TM, Robles-Kelly A, Naqvi SS. A Semantically Flexible Feature Fusion Network for Retinal Vessel Segmentation. In: International Conference on Neural Information Processing. Springer, Cham; 2020. p. 159–167.
29. Khan TM, Robles-Kelly A. Machine Learning: Quantum vs Classical. IEEE Access. 2020;8:219275–219294.
- View Article
- Google Scholar
30. Cherukuri V, Bg VK, Bala R, Monga V. Deep retinal image segmentation with regularization under geometric priors. IEEE Transactions on Image Processing. 2019;29:2552–2567. pmid:31613766
- View Article
- PubMed/NCBI
- Google Scholar
31. Liskowski P, Krawiec K. Segmenting retinal blood vessels with deep neural networks. IEEE transactions on medical imaging. 2016;35(11):2369–2380. pmid:27046869
- View Article
- PubMed/NCBI
- Google Scholar
32. Imtiaz R, Khan TM, Naqvi SS, Arsalan M, Nawaz SJ. Screening of Glaucoma disease from retinal vessel images using semantic segmentation. Computers & Electrical Engineering. 2021;91:107036.
- View Article
- Google Scholar
33. Khan TM, Robles-Kelly A, Naqvi SS, Muhammad A. Residual Multiscale Full Convolutional Network (RM-FCN) for High Resolution Semantic Segmentation of Retinal Vasculature. In: Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR International Workshops, S+ SSPR 2020, Padua, Italy, January 21–22, 2021, Proceedings. Springer Nature; 2021. p. 324.
34. Badrinarayanan V, Kendall A, Cipolla R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2017;39(12):2481–2495. pmid:28060704
- View Article
- PubMed/NCBI
- Google Scholar
35. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2016. p. 770–778.
36. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017. p. 4700–4708.
37. Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition; 2015.
38. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. vol. 9351; 2015. p. 234–241.
39. Arsalan M, Oqais M, tahir Mahmood, Cho SW, Park KR. Aiding the Diagnosis of Diabetic and Hypertensive Retinopathy Using Artificial Intelligence-Based Semantic Segmentation. Journal of Clinical Medicine. 2019;8(9).
- View Article
- Google Scholar
40. Guan S, Khan AA, Sikdar S, Chitnis PV. Fully Dense UNet for 2-D Sparse Photoacoustic Tomography Artifact Removal. IEEE Journal of Biomedical and Health Informatics. 2020;24(2):568–576. pmid:31021809
- View Article
- PubMed/NCBI
- Google Scholar
41. Hoover A, Kouznetsova V, Goldbaum M. Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging. 2000;19(3):203–210. pmid:10875704
- View Article
- PubMed/NCBI
- Google Scholar
42. Staal J, Abràmoff MD, Niemeijer M, Viergever MA, Van Ginneken B. Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging. 2004;23(4):501–509. pmid:15084075
- View Article
- PubMed/NCBI
- Google Scholar
43. Fraz MM, Barman SA, Remagnino P, Hoppe A, Basit A, Uyyanonvara B, et al. An approach to localize the retinal blood vessels using bit planes and centerline detection. Computer methods and programs in biomedicine. 2012;108(2):600–616. pmid:21963241
- View Article
- PubMed/NCBI
- Google Scholar
44. Soares JV, Leandro JJ, Cesar RM, Jelinek HF, Cree MJ. Retinal vessel segmentation using the 2-D Gabor wavelet and supervised classification. IEEE Transactions on medical Imaging. 2006;25(9):1214–1222. pmid:16967806
- View Article
- PubMed/NCBI
- Google Scholar
45. Li Q, Feng B, Xie L, Liang P, Zhang H, Wang T. A cross-modality learning approach for vessel segmentation in retinal images. IEEE transactions on medical imaging. 2015;35(1):109–118. pmid:26208306
- View Article
- PubMed/NCBI
- Google Scholar
46. Zhang J, Dashtbozorg B, Bekkers E, Pluim JPW, Duits R, Romeny BM. Robust Retinal Vessel Segmentation via Locally Adaptive Derivative Frames in Orientation Scores. IEEE Transactions on Medical Imaging. 2016;35(12):2631–2644. pmid:27514039
- View Article
- PubMed/NCBI
- Google Scholar
47. Jin Q, Meng Z, Pham TD, Chen Q, Wei L, Su R. DUNet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems. 2019;178:149–162.
- View Article
- Google Scholar
48. Yin P, Yuan R, Cheng Y, Wu Q. Deep Guidance Network for Biomedical Image Segmentation. IEEE Access. 2020;8:116106–116116.
- View Article
- Google Scholar
49. Wang D, Haytham A, Pottenburgh J, Saeedi O, Tao Y. Hard Attention Net for Automatic Retinal VesselSegmentation. IEEE Journal ofBiomedical and Health Informatics. 2020.
- View Article
- Google Scholar
50. Ma W, Yu S, Ma K, Wang J, Ding X, Zheng Y. Multi-Task Neural Networks with Spatial Activation for Retinal Vessel Segmentation and Artery/Vein Classification. In: Medical Image Computing and Computer Assisted Intervention; 2019.
51. Guo S, Wang K, Kang H, Zhang Y, Gao Y, Li T. BTS-DSN: Deeply supervised neural network with short connections for retinal vessel segmentation. International Journal of Medical Informatics. 2019;126:105–113. pmid:31029251
- View Article
- PubMed/NCBI
- Google Scholar
52. Wu Y, Xia Y, Song Y, Zhang D, Liu D, Zhang C, et al. Medical Image Computing and Computer Assisted Intervention. In: Vessel-Net: Retinal Vessel Segmentation Under Multi-path Supervision; 2019.
53. Wang B, Qiu S, He H. Dual Encoding U-Net for Retinal Vessel Segmentation. In: Medical Image Computing and Computer Assisted Intervention; 2019.
54. Gu Z, Cheng J, Fu H, Zhou K, Hao H, Zhao Y, et al. CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE Transactions on Medical Imaging. 2019;38(10):2281–2292. pmid:30843824
- View Article
- PubMed/NCBI
- Google Scholar
55. Chen LC, Zhu Y, Papandreou G, Schroff F, Adam H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In: European Conference on Computer Vision; 2018. p. 833–851.

[ref1] 1. Mohamed Q, Gillies MC, Wong TY. Management of diabetic retinopathy: a systematic review. Jama. 2007;298(8):902–916. pmid:17712074
View Article
PubMed/NCBI
Google Scholar

[2] View Article

[3] PubMed/NCBI

[4] Google Scholar

[ref2] 2. Srinidhi CL, Aparna P, Rajan J. Recent advancements in retinal vessel segmentation. Journal of medical systems. 2017;41(4):70.
View Article
Google Scholar

[6] View Article

[7] Google Scholar

[ref3] 3. Cheung CYl, Zheng Y, Hsu W, Lee ML, Lau QP, Mitchell P, et al. Retinal vascular tortuosity, blood pressure, and cardiovascular risk factors. Ophthalmology. 2011;118(5):812–818. pmid:21146228
View Article
PubMed/NCBI
Google Scholar

[9] View Article

[10] PubMed/NCBI

[11] Google Scholar

[ref4] 4. Niemeijer M, Staal J, van Ginneken B, Loog M, Abramoff MD. Comparative study of retinal vessel segmentation methods on a new publicly available database. In: Medical imaging 2004: image processing. vol. 5370. International Society for Optics and Photonics; 2004. p. 648–656.

[ref5] 5. Soomro TA, Gao J, Khan T, Hani AFM, Khan MA, Paul M. Computerised approaches for the detection of diabetic retinopathy using retinal fundus images: a survey. Pattern Analysis and Applications. 2017;20(4):927–961.
View Article
Google Scholar

[14] View Article

[15] Google Scholar

[ref6] 6. Ravudu M, Jain V, Kunda MMR. Review of image processing techniques for automatic detection of eye diseases. In: 2012 Sixth International Conference on Sensing Technology (ICST). IEEE; 2012. p. 320–325.

[ref7] 7. Irshad S, Akram MU. Classification of retinal vessels into arteries and veins for detection of hypertensive retinopathy. In: 2014 Cairo International Biomedical Engineering Conference (CIBEC). IEEE; 2014. p. 133–136.

[ref8] 8. Khawaja A, Khan TM, Khan MAU, Nawaz J. A Multi-Scale Directional Line Detector for Retinal Vessel Segmentation. Sensors. 2019;19(22). pmid:31766276
View Article
PubMed/NCBI
Google Scholar

[19] View Article

[20] PubMed/NCBI

[21] Google Scholar

[ref9] 9. Khan MA, Khan TM, Naqvi SS, Khan MA. Ggm classifier with multi-scale line detectors for retinal vessel segmentation. Signal, Image and Video Processing. 2019;13(8):1667–1675.
View Article
Google Scholar

[23] View Article

[24] Google Scholar

[ref10] 10. Wisaeng K, Hiransakolwong N, Pothiruk E. Automatic detection of retinal exudates using a support vector machine. Applied Medical Informatics. 2013;32(1):33–42.
View Article
Google Scholar

[26] View Article

[27] Google Scholar

[ref11] 11. Soomro TA, Khan MA, Gao J, Khan TM, Paul M. Contrast normalization steps for increased sensitivity of a retinal image segmentation method. Signal, Image and Video Processing. 2017; 11(8):1509–1517.
View Article
Google Scholar

[29] View Article

[30] Google Scholar

[ref12] 12. Khan TM, Bailey DG, Khan MA, Kong Y. Efficient hardware implementation for fingerprint image enhancement using anisotropic Gaussian filter. IEEE Transactions on Image processing. 2017;26(5):2116–2126. pmid:28237927
View Article
PubMed/NCBI
Google Scholar

[32] View Article

[33] PubMed/NCBI

[34] Google Scholar

[ref13] 13. Khan MA, Soomro TA, Khan TM, Bailey DG, Gao J, Mir N. Automatic retinal vessel extraction algorithm based on contrast-sensitive schemes. In: 2016 International conference on image and vision computing New Zealand (IVCNZ). IEEE; 2016. p. 1–5.

[ref14] 14. Franklin SW, Rajan SE. Computerized screening of diabetic retinopathy employing blood vessel segmentation in retinal images. biocybernetics and biomedical engineering. 2014;34(2):117–124.
View Article
Google Scholar

[37] View Article

[38] Google Scholar

[ref15] 15. Khan MA, Khan TM, Aziz KI, Ahmad SS, Mir N, Elbakush E. The use of fourier phase symmetry for thin vessel detection in retinal fundus images. In: 2019 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT). IEEE; 2019. p. 1–6.

[ref16] 16. Mehmood M, Khan TM, Khan MA, Naqvi SS, Alhalabi W. Vessel intensity profile uniformity improvement for retinal vessel segmentation. Procedia Computer Science. 2019;163:370–380.
View Article
Google Scholar

[41] View Article

[42] Google Scholar

[ref17] 17. Khan MA, Khan TM, Bailey DG, Soomro TA. A generalized multi-scale line-detection method to boost retinal vessel segmentation sensitivity. Pattern Analysis and Applications. 2019;22(3):1177–1196.
View Article
Google Scholar

[44] View Article

[45] Google Scholar

[ref18] 18. Soomro TA, Khan TM, Khan MA, Gao J, Paul M, Zheng L. Impact of ICA-based image enhancement technique on retinal blood vessels segmentation. IEEE Access. 2018;6:3524–3538.
View Article
Google Scholar

[47] View Article

[48] Google Scholar

[ref19] 19. Khan MA, Khan TM, Soomro TA, Mir N, Gao J. Boosting sensitivity of a retinal vessel segmentation algorithm. Pattern Analysis and Applications. 2019;22(2):583–599.
View Article
Google Scholar

[50] View Article

[51] Google Scholar

[ref20] 20. Fan Z, Lu J, Wei C, Huang H, Cai X, Chen X. A hierarchical image matting model for blood vessel segmentation in fundus images. IEEE Transactions on Image Processing. 2018;28(5):2367–2377. pmid:30571623
View Article
PubMed/NCBI
Google Scholar

[53] View Article

[54] PubMed/NCBI

[55] Google Scholar

[ref21] 21. Soomro TA, Khan MA, Gao J, Khan TM, Paul M, Mir N. Automatic retinal vessel extraction algorithm. In: 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA). IEEE; 2016. p. 1–8.

[ref22] 22. Khan TM, Khan MA, Rehman NU, Naveed K, Afridi IU, Naqvi SS, et al. Width-wise vessel bifurcation for improved retinal vessel segmentation. Biomedical Signal Processing and Control. 2022;71:103169.
View Article
Google Scholar

[58] View Article

[59] Google Scholar

[ref23] 23. Toufique Soomro MAUKTMKMPNM Junbin Gao. Role of Image Contrast Enhancement Technique for Ophthalmologist as a Diagnostic Tool for the Diabetic Retinopathy. In: IEEE International Conference on Digital Image Computing: Techniques and Applications (DICTA), At Gold Coast, Autralia; 2016. p. 1–8.

[ref24] 24. Khawaja A, Khan TM, Naveed K, Naqvi SS, Rehman NU, Nawaz SJ. An improved retinal vessel segmentation framework using frangi filter coupled with the probabilistic patch based denoiser. IEEE Access. 2019;7:164344–164361.
View Article
Google Scholar

[62] View Article

[63] Google Scholar

[ref25] 25. Naveed K, Abdullah F, Madni HA, Khan MA, Khan TM, Naqvi SS. Towards Automated Eye Diagnosis: An Improved Retinal Vessel Segmentation Framework Using Ensemble Block Matching 3D Filter. Diagnostics. 2021;11(1):114. pmid:33445723
View Article
PubMed/NCBI
Google Scholar

[65] View Article

[66] PubMed/NCBI

[67] Google Scholar

[ref26] 26. Tuba E, Mrkela L, Tuba M. Retinal blood vessel segmentation by support vector machine classification. In: 2017 27th International Conference Radioelektronika (RADIOELEKTRONIKA). IEEE; 2017. p. 1–6.

[ref27] 27. Khan TM, Abdullah F, Naqvi SS, Arsalan M, Khan MA. Shallow Vessel Segmentation Network for Automatic Retinal Vessel Segmentation. In: 2020 International Joint Conference on Neural Networks (IJCNN). IEEE; 2020. p. 1–7.

[ref28] 28. Khan TM, Robles-Kelly A, Naqvi SS. A Semantically Flexible Feature Fusion Network for Retinal Vessel Segmentation. In: International Conference on Neural Information Processing. Springer, Cham; 2020. p. 159–167.

[ref29] 29. Khan TM, Robles-Kelly A. Machine Learning: Quantum vs Classical. IEEE Access. 2020;8:219275–219294.
View Article
Google Scholar

[72] View Article

[73] Google Scholar

[ref30] 30. Cherukuri V, Bg VK, Bala R, Monga V. Deep retinal image segmentation with regularization under geometric priors. IEEE Transactions on Image Processing. 2019;29:2552–2567. pmid:31613766
View Article
PubMed/NCBI
Google Scholar

[75] View Article

[76] PubMed/NCBI

[77] Google Scholar

[ref31] 31. Liskowski P, Krawiec K. Segmenting retinal blood vessels with deep neural networks. IEEE transactions on medical imaging. 2016;35(11):2369–2380. pmid:27046869
View Article
PubMed/NCBI
Google Scholar

[79] View Article

[80] PubMed/NCBI

[81] Google Scholar

[ref32] 32. Imtiaz R, Khan TM, Naqvi SS, Arsalan M, Nawaz SJ. Screening of Glaucoma disease from retinal vessel images using semantic segmentation. Computers & Electrical Engineering. 2021;91:107036.
View Article
Google Scholar

[83] View Article

[84] Google Scholar

[ref33] 33. Khan TM, Robles-Kelly A, Naqvi SS, Muhammad A. Residual Multiscale Full Convolutional Network (RM-FCN) for High Resolution Semantic Segmentation of Retinal Vasculature. In: Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR International Workshops, S+ SSPR 2020, Padua, Italy, January 21–22, 2021, Proceedings. Springer Nature; 2021. p. 324.

[ref34] 34. Badrinarayanan V, Kendall A, Cipolla R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2017;39(12):2481–2495. pmid:28060704
View Article
PubMed/NCBI
Google Scholar

[87] View Article

[88] PubMed/NCBI

[89] Google Scholar

[ref35] 35. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2016. p. 770–778.

[ref36] 36. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017. p. 4700–4708.

[ref37] 37. Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition; 2015.

[ref38] 38. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. vol. 9351; 2015. p. 234–241.

[ref39] 39. Arsalan M, Oqais M, tahir Mahmood, Cho SW, Park KR. Aiding the Diagnosis of Diabetic and Hypertensive Retinopathy Using Artificial Intelligence-Based Semantic Segmentation. Journal of Clinical Medicine. 2019;8(9).
View Article
Google Scholar

[95] View Article

[96] Google Scholar

[ref40] 40. Guan S, Khan AA, Sikdar S, Chitnis PV. Fully Dense UNet for 2-D Sparse Photoacoustic Tomography Artifact Removal. IEEE Journal of Biomedical and Health Informatics. 2020;24(2):568–576. pmid:31021809
View Article
PubMed/NCBI
Google Scholar

[98] View Article

[99] PubMed/NCBI

[100] Google Scholar

[ref41] 41. Hoover A, Kouznetsova V, Goldbaum M. Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Transactions on Medical imaging. 2000;19(3):203–210. pmid:10875704
View Article
PubMed/NCBI
Google Scholar

[102] View Article

[103] PubMed/NCBI

[104] Google Scholar

[ref42] 42. Staal J, Abràmoff MD, Niemeijer M, Viergever MA, Van Ginneken B. Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging. 2004;23(4):501–509. pmid:15084075
View Article
PubMed/NCBI
Google Scholar

[106] View Article

[107] PubMed/NCBI

[108] Google Scholar

[ref43] 43. Fraz MM, Barman SA, Remagnino P, Hoppe A, Basit A, Uyyanonvara B, et al. An approach to localize the retinal blood vessels using bit planes and centerline detection. Computer methods and programs in biomedicine. 2012;108(2):600–616. pmid:21963241
View Article
PubMed/NCBI
Google Scholar

[110] View Article

[111] PubMed/NCBI

[112] Google Scholar

[ref44] 44. Soares JV, Leandro JJ, Cesar RM, Jelinek HF, Cree MJ. Retinal vessel segmentation using the 2-D Gabor wavelet and supervised classification. IEEE Transactions on medical Imaging. 2006;25(9):1214–1222. pmid:16967806
View Article
PubMed/NCBI
Google Scholar

[114] View Article

[115] PubMed/NCBI

[116] Google Scholar

[ref45] 45. Li Q, Feng B, Xie L, Liang P, Zhang H, Wang T. A cross-modality learning approach for vessel segmentation in retinal images. IEEE transactions on medical imaging. 2015;35(1):109–118. pmid:26208306
View Article
PubMed/NCBI
Google Scholar

[118] View Article

[119] PubMed/NCBI

[120] Google Scholar

[ref46] 46. Zhang J, Dashtbozorg B, Bekkers E, Pluim JPW, Duits R, Romeny BM. Robust Retinal Vessel Segmentation via Locally Adaptive Derivative Frames in Orientation Scores. IEEE Transactions on Medical Imaging. 2016;35(12):2631–2644. pmid:27514039
View Article
PubMed/NCBI
Google Scholar

[122] View Article

[123] PubMed/NCBI

[124] Google Scholar

[ref47] 47. Jin Q, Meng Z, Pham TD, Chen Q, Wei L, Su R. DUNet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems. 2019;178:149–162.
View Article
Google Scholar

[126] View Article

[127] Google Scholar

[ref48] 48. Yin P, Yuan R, Cheng Y, Wu Q. Deep Guidance Network for Biomedical Image Segmentation. IEEE Access. 2020;8:116106–116116.
View Article
Google Scholar

[129] View Article

[130] Google Scholar

[ref49] 49. Wang D, Haytham A, Pottenburgh J, Saeedi O, Tao Y. Hard Attention Net for Automatic Retinal VesselSegmentation. IEEE Journal ofBiomedical and Health Informatics. 2020.
View Article
Google Scholar

[132] View Article

[133] Google Scholar

[ref50] 50. Ma W, Yu S, Ma K, Wang J, Ding X, Zheng Y. Multi-Task Neural Networks with Spatial Activation for Retinal Vessel Segmentation and Artery/Vein Classification. In: Medical Image Computing and Computer Assisted Intervention; 2019.

[ref51] 51. Guo S, Wang K, Kang H, Zhang Y, Gao Y, Li T. BTS-DSN: Deeply supervised neural network with short connections for retinal vessel segmentation. International Journal of Medical Informatics. 2019;126:105–113. pmid:31029251
View Article
PubMed/NCBI
Google Scholar

[136] View Article

[137] PubMed/NCBI

[138] Google Scholar

[ref52] 52. Wu Y, Xia Y, Song Y, Zhang D, Liu D, Zhang C, et al. Medical Image Computing and Computer Assisted Intervention. In: Vessel-Net: Retinal Vessel Segmentation Under Multi-path Supervision; 2019.

[ref53] 53. Wang B, Qiu S, He H. Dual Encoding U-Net for Retinal Vessel Segmentation. In: Medical Image Computing and Computer Assisted Intervention; 2019.

[ref54] 54. Gu Z, Cheng J, Fu H, Zhou K, Hao H, Zhao Y, et al. CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE Transactions on Medical Imaging. 2019;38(10):2281–2292. pmid:30843824
View Article
PubMed/NCBI
Google Scholar

[142] View Article

[143] PubMed/NCBI

[144] Google Scholar

[ref55] 55. Chen LC, Zhu Y, Papandreou G, Schroff F, Adam H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In: European Conference on Computer Vision; 2018. p. 833–851.

Figures

Abstract

1 Introduction

2 Background and related work

3 Proposed methodology

3.1 Overview of proposed architecture

3.2 Working principle of the DAVS-Net

3.3 DAVS-Net encoder

3.4 DAVS-Net decoder

4 Detection of diabetic and hypertensive retinopathy

5 Experimental results

5.1 Materials

5.2 Evaluation criteria

5.3 Comparison with state-of-the-art

6 Conclusion

References