Attention aware fully convolutional deep learning model for retinal blood vessel segmentation

Abstract

Recent reports indicate a rise in retinal issues, and automatic artery vein categorization offers data that is particularly instructive for the medical evaluation of serious retinal disorders including glaucoma and diabetic retinopathy. This work presents a competent and precise deep-learning model designed for vessel segmentation in retinal fundus imaging. This article aims to segment the retinal images using an attention-based dense fully convolutional neural network (A-DFCNN) after removing uncertainty. The artery extraction layers encompass vessel-specific convolutional blocks to focus the tiny blood vessels and dense layers with skip connections for feature propagation. Segmentation is associated with artery extraction layers via individual loss function. Blood vessel maps produced from individual loss functions are authenticated for performance. The proposed technique attains improved outcomes in terms of Accuracy (0.9834), Sensitivity (0.8553), and Specificity (0.9835) from DRIVE, STARE, and CHASE-DB1 datasets. The result demonstrates that the proposed A-DFCNN is capable of segmenting minute vessel bifurcation breakdowns during the training and testing phases.

Keywords

Deep learning fundus image fully-convolutional neural networks blood vessel segmentation artery vein classification

1 Introduction

Glaucoma and diabetic retinopathy are two eye conditions that seriously affect the optic nerve and eventually result in blindness. Patients with glaucoma experience intra ocular pressure in their eyes, some of whom may feel it and others who may not. Glaucoma’s irreversible harm has a significant impact on the number of blind people worldwide. Through the analysis of fund us images, it is possible to determine how many people are blind worldwide due to these serious eye disorders [1]. Quick recognition and regular treatment will reduce optic diseases and avoids growth in the blindness population. Manual segmentation includes Optic Disc (OD) Segmentation, after separation of the red and blue channel inputs by pre-processing. A dynamic intensity threshold is employed to determine the dissimilar threshold values from areas. Then identification of all the objects is made using eccentricity. Thus, it holds only the value of <0.9 to remove semi-circular reflections at the border of the images.

Blood vessels Removal and In-painting, blood vessels are removed using blue and red channels as input. Equation (1) is used to create a brightness channel to luminance blood vessels.(1). Then the bottom-hat morphological operation is applied with the help of a disk-shaped framing component with 26 as its size. $L = 0.299 * R + 0.587 * G + 0.114 * B$ (1)

The binary images of the segmented optic disc and the blood vessel are multiplied pixel-wise by using Equation (2). Finally, blood vessels are removed. $IR = I_{BV} * I_{OD}$ (2)

Optic Cup Segmentation, the variance maximization method is used to repaint the eliminated areas in the optic disc. In this method region bounding box is used. The pixels in this region are to be centred with a 7×7 window which is equivalent to the blood vessels present in the optic disc. Then rotate the 7×7 window in 12 different directions to predict the disparity of pixels in this domain. The direction with the highest discrepancy is anticipated, and that greater variance value replaces the central pixel. By following this procedure, a green channel that has been in-painted is obtained. The disc cup is then segmented using an intensity-based threshold.Image created with threshold replaces each pixel if intensity x (i, j) is less than A. It is defined as: $I (i, j) = {\begin{matrix} 1 & ifx (i, j) > A \\ 0 & others \end{matrix}$ (3)

An automatic segmentation method is of three different types: i) optic disc segmentation ii) optic cup segmentation and iii) vessel segmentation [2]. An algorithmic way of segmenting vessels in the retina will be an advantage to a great extent by reducing time and minimizing complexity. The evolution of CNN-based vessel classification is done with connectivity patterns in neurons and with multiple convolutional layers [3, 4]. Encoder-improved Unet enhances network performance without computational burden. In the Encoder additional layer is carried out after each convolutional layer except the first stage. A new layer is introduced between feature maps of the first convolution layer output and feature maps of the max pooling output. In between obtained output and the second convolutional layer, the feature maps second addition is included.

The addition of residual or attention blocks also improves performance [5, 6]. Proposed fully convolutional neural network to segment vascular tree is done on colour fundus images i) Includes attention mechanism into the Unet model architecture ii) Discussed loss function and parameter used during training process iii) Appliance of a trained network [7]. Fully convolutional networks as semantic segmentation with skip connections works in deep shallow architecture. The Unet model [8] was designed with a contraction path and expansion path that will localize the image, applied in major bio-medical segmentation tasks. Proposed dense U-net with the patch learning process, undergoes patch training process of fundus images. In general, all the deep CNN-based structures are moving towards automated applications of medical imaging. Fully automated segmentation of retinal vessels follows the contraction path and expansion path. The model is assisted with skip connections to pass hierarchical representation between different layers [9]. The performance of the article’s suggested SineNet model is assessed using well-known public datasets as DRIVE, STARE, and CHASEDB.The cross-database study is done for vessel segmentation which may be helpful in clinical applications.

The contribution of this article includes, the proposed A-DFCNN model is consequent from existing Unet utilizing subsequent changes: Downsampling and upsampling stages remain with a similar number of kernels when stages get increased. Assign the same weight for down-level and high-level features. ii) Uncertainty quantification research is done for the AV classification. The construction of the paper is as follows: Section 2 expresses the state of art comparison with other deep learning models. The section 3 describes the proposed approach with AV classification and architecture model. The section 4 comprises the output attained, and the performance of the proposed work. Section 5 concludes the overall work done in fundus images.

2 State-of-the-art comparison

Fully convolutional neural networks (FCNN) are used for vessel segmentation with stationary wavelet transform (SWT) exposing the structure of the retinal vascular map. Additional channels are incorporated into the fully convolutional neural network structure. SWT conserves pixels by enriching the input of FCNN. The segmentation of vessels can be carried out either globally or locally. In some circumstances, performance can be enhanced by region-wise semantic segmentation combined with transfer learning. Fundus image segmentation using upsampling and downsampling techniques in FCNN will reveal thin and thick vascular characteristics [10, 11]. In fundus image segmentation usage of moderate vessel features by the decoder can represent tiny objects pixel-wise. There are unfair pixels between vessels and pixels in the surroundings. Accordingly, it is proposed to use a round-wise feature aggregation method with a low stride value that has rich features with tiny blood vessels [12]. ResNet with scale space approximation layers is used for the multi-scale representation of fundus images, with down-sampling and up-sampling processes [13].

New Cross connected CNN model developed for vessel segmentation improves the overall performance of segmentation, as well as a strong segmentation algorithm, was proposed. Performance was assessed in terms of accuracy, prediction speed, and specificity. Feature fusion with skip connection produces great results in DRIVE and STARE datasets [14]. Retinal vessel segmentation in the Dense-Unet model using a random transformation-based training approach was proposed. Random transformation applied for image augmentation and patch-based learning is done to improve the training effectiveness. The dice loss function in fundus image segmentation addresses the data imbalance problem between segmented and ground truth images. An attention mechanism is proposed in future work for the segmentation of tiny vessel features [8, 15]. Skip connection-based residual deep learning networks combine features of various methods. Skip connections extract necessary information from fundus images with a 10-fold validation strategy. SCB (skip connection block) has a fully connected layer followed by ReLU activation. It has advantages in the feed-forward network and weight-updating method of back propagation [16, 17]. Modified encoder-based Unet architecture is proposed with two additional layers at each stage which extracts useful edge information of tiny vessels [6, 18]. The addition of attention blocks and residual blocks also improves performance comparatively while using Unet architecture alone. Attention ResUnet can be seen as an upsampling/ downsampling method that builds by attention blocks. The performance is improved by adding a new type of BSE block (Attention mechanism) with residual structures.Skip connection in residual blocks plays a major role in the modification of the depth of a network without a change in the performance of a network [19, 20]. Table 1 shows the various methods of proposed work.

Table 1
Different method of the retinal blood vessel segmentation

Year /citation Representation Architecture

2018 [10] Fully convolutional neural networks

2018 [11] Fully convolutional deep learning architecture

2019 [12] Round-wise features aggregation

2019 [13] Scale-space approximated CNN

2019 [8] Dense Unet with patch-based learning

2019 [16] Trilogy of skip connection deep networks

2020 [14] Convolutional neural network with cross-connected

2020 [18] Residual Unet for fundus image segmentation

2020 [6] RCED-Net (Residual conn based encoder-decoder network)

2021 [3] CNN with skip connections

2021 [9] Fully convolutional deep learning architecture

2021 [5] Encoder Enhanced Atrous (EEA) Unet

2021 [7] Convolutional kernels and modified Unet model

2021 [19] BSER Unet (Before activation squeeze excitation residual Unet)

2021 [20] Attention ResUnet

2021 [15] Dense Unet

Year /citation	Representation Architecture
2018 [10]	Fully convolutional neural networks
2018 [11]	Fully convolutional deep learning architecture
2019 [12]	Round-wise features aggregation
2019 [13]	Scale-space approximated CNN
2019 [8]	Dense Unet with patch-based learning
2019 [16]	Trilogy of skip connection deep networks
2020 [14]	Convolutional neural network with cross-connected
2020 [18]	Residual Unet for fundus image segmentation
2020 [6]	RCED-Net (Residual conn based encoder-decoder network)
2021 [3]	CNN with skip connections
2021 [9]	Fully convolutional deep learning architecture
2021 [5]	Encoder Enhanced Atrous (EEA) Unet
2021 [7]	Convolutional kernels and modified Unet model
2021 [19]	BSER Unet (Before activation squeeze excitation residual Unet)
2021 [20]	Attention ResUnet
2021 [15]	Dense Unet

3 The proposed approach

3.1 Materials

The proposed system is assessed using DRIVE, CHASE_Db1, and STARE fundus image databases. The DRIVE record [21] comprises 40 RGB retinal images with a 584×565 pixel resolution. It uses half of its fundus images for testing and training processes respectively. The STARE database [21] has 20 retinal fundus images with a 605×700 pixel resolution. The CHASE_Db1 record [22] has 28 fundus images taken from an NM-200 fundus camera with an image resolution of 999×960 pixels. Figure 1 shows the proposed work and Fig. 2 are depicted as a DRIVErecord of input fundus image and artery vein reference.

Fig. 1

Proposed workflow diagram.

Fig. 2

DRIVE Record: (a) Input fundus image and (b) artery vein reference model.

3.2 Artery vein classification

Let model with input image x and parameters θ to a function that maps probabilities with spatial extent as the source image: $\overset{\land}{R} = f (i, θ)$ (4)

Probabilities of the mapping function $\overset{\land}{R} = [R_{artery}^{\land} \overset{\land}{R}_vein]$ , labelling objects to vessels are evaluated as follows: $X_{artery} = sigmoid (R_{artery}^{\land})$ (5) $X_{vein} = sigmoid (R_{vein}^{\land})$ (6)

A classification pixel of artery and vein is significant for a multi-label system [23, 24]. The training dataset has an image mask of segmented vessels. $X_{vessel} = X_{artery} + X_{vein} - X_{artery} X_{vein}$ (7)

For classification networks uncertainty of loss, and attenuation occurs in the neural network. Input-reliant uncertainty is common in many classification processes [26]. The defined regular classification technique helps to influence transitional hetero regression. This chosen model works well with neural networks for heterogeneous categorization. Uncertainty is detained by adapting the innovative technique to calculate average and dispersed data of logits: $\overset{\land}{[R}, σ] = f (i, θ)$ (8)

To evaluate dispersed data, logits of the subsequent layer are included in the amount of produced dataset [27 –38]. An activation function is included in the resulting layer to ensure evaluated dispersed data is positive or not. The likelihood of samples is evaluated as follows: $\overset{\land}{X} = sigmoid (\overset{\land}{R} + σ ⊙ \in), \in \sim S (0, Z)$ (9)

Where, ⊙ is set as the product and ∈ are labelled in the conclusion.

The main implication method for AV remains the same with the exemption. The loss is evaluated from models NA labels, minimum loss is nothing but a mean of individual labels. Uncertainty is identified in the view of parameters as random variables $X (R | i, D) = \int X (R | i, θ) X (θ | D) d θ$ (10)

Where D is a pair of input and output, intractable parameter X (θ | D) is used in fully connected deep neural networks.

3.3 Architecture

The structure used is A-DFCNN. A-DFCNN outperforms other fully convolutional network architectures in AV vessel structure prediction. Layer Input is added pixel-wise with an output of the previous layer in dense convolutional blocks shown in Fig. 3. The model has normalization, activation function, convolution operation, and dropout to produce feature maps. It is similar to other downsampling and upsampling structures where downsampling will compress the input and the transition of skip connection is done between the encoder and decoder. The up-sampling process has dense convolutional blocks and transitions.

Fig. 3

Attention-based dense fully-convolutional neural network model.

Skip connection: Skip connections will combine high-level and low-level DFCNN characteristics that perform well in terms of sensitivity and accuracy. By modulating the skip concatenate across the contracting path and the expansion process, the DFCNN architecture encourages information synthesis among local and global features. It aids in emphasising the vessel’s characteristics. By using the skip-connections shown in Fig. 3, the dense block Db 1 is joined to the dense block Db 5 for the sample.

Attention mechanism is the process of including parameters in every channel of a convolutional block, so that the method can adaptively fine-tune the weighting of each feature map. Figure 4 consists of five basic units of operations: i) Input added to the convolutional block and the number of channels used for future intervention. ii) Squeeze channels into a single value by average pooling. iii) ReLU function next to a fully connected layer will avoid nonlinearity. iv) Smoothing operation is given by a second fully connected layer. v) Based on the result produced in the network feature maps are weighted.

Fig. 4

Attention mechanism.

4 Experiment and results

The training of the proposed A–DFCNN model is done on NVIDIA Tesla k40c GPU, 32GB RAM, and Intel E5-2630 processor. Four training groups are assigned with 30 epochs, with a learning rate of 0.0001. For pixel-wise vessel segmentation, it produces four measures: TP, TN, FP, and FN viewpoints for the number of true positive and negative samples and the number of false positive and negative samples, which is depicted as Equations 11 to 16. The following metrics are used for assessment: Accuracy, sensitivity, specificity, and area under the curve.The A-DFCNN model is compared on the following datasets CHASE_DB1, STARE, and DRIVE. Every dataset is divided into the training phase, validation, and testing phase. The addition of the Adam optimizer and the binary cross entropy loss function. The stochastic descent method’s Adam optimizer is an enhancement that works well for deep learning applications in image processing and computer vision.

Proposed DFCNN algorithm test image and input model of corresponding segment image pseudo code described. Let H, W is image width and height, P is patch size, O_p is output patch size, O_s is output stride size, and S is sampling step size. The usage of attention mechanism segments tiny vessel features precisely concerning the ground reality images revealed in Fig. 5. Output images are compared with ground reality images in Fig. 6. Attained results from section (d) show that the Attention-based DFCNN network performs well as compared to DFCNN from section (c). The proposed technique’s effectiveness depends on attention maps produced from fundus images. The self-attention mechanism includes different comparison functions such as multiplicative attention and addictive attention.

Fig. 5

Rows 1, and 2 of DRIVE and STARE datasets show a partial view of attention results (a) Input fundus image (b) partial views (c) ground truth (d) A-DFCNN.

Fig. 6

Rows 1, 2 and 3 are DRIVE, STARE and CHASE_DB1 datasets respectively. (a) Input fundus image (b) Ground truth (c) DFCNN (d) A-DFCNN.

$Accuracy = \frac{TP + TN}{TP + TN + FP + {FN}^{'}}$ (11) $Sensitivity = \frac{TP}{TP + {FN}^{'}}$ (12) $Specificity = \frac{TN}{TN + {FP}^{'}}$ (13) $Precision = \frac{TP}{TP + {FP}^{'}}$ (14) $Recall = \frac{TP}{TP + {FN}^{'}}$ (15) $F 1 = 2 \times \frac{Precision×Recall}{{Precision×Recall}^{'}}$ (16)

Numerous works has been done on the segmentation of retinal vessels based on STARE, DRIVE, and CHASE_DB1 datasets depicted in Table 2. Since the proposed technique depends on a self-attention mechanism which provides better results in all the datasets. Individual database results: The DRIVE dataset hathe most challenging dataset which provides less accuracy and sensitivity. STARE dataset has low-contrast images which affect vessel structure prediction. It has high accuracy and specificity but less sensitivity. It is found that the CHASE_DB1 dataset produces better results concerning all aspects. Combined dataset results: In the overview, the combined images from DRIVE, STARE and CHASE_DB1 used for training andthe STARE dataset’s challenging images which affect the segmentation result. Figure 7(a&b) depicted as individual/combined dataset training of proposed model and result image presented.

Table 2

Performance comparison of proposed model to show how it is varied from existing models in DRIVE, STARE and CHASE_DB1 database

Database	Existing Methods	Year	Acc	Sen	Spc	AUC
DRIVE	Jiang et al. [11]	2018	0.9624	0.7540	0.9825	0.9810
	Feng et al. [14]	2018	0.9528	0.7625	0.9809	0.9678
	Hu et al. [25]	2018	0.9533	0.7772	0.9793	0.9759
	Oliveira et al. [10]	2018	0.9639	0.8405	0.9814	–
	Wang, Chang, et al. [8]	2019	0.9511	0.7986	0.9736	0.9740
	Lv, Yan, et al. [21]	2020	0.9558	0.7941	0.9798	0.9847
	Tang, Yu, et al. [18]	2020	0.9554	0.8160	0.9756	0.9799
	Atli et al. [9]	2020	0.9689	0.7987	0.9854	0.9851
	Samuel et al. [3]	2021	0.9627	0.7827	0.9821	0.9789
	Sathananthavathi et al. [5]	2021	0.9577	0.7918	0.9708	–
	Manuel et al. [7]	2021	0.9547	–	–	0.9837
	Proposed Method	2021	0.9524	0.7529	0.9731	0.9708
STARE	Jiang et al. [11]	2018	0.9734	0.8352	0.9846	0.9900
	Feng et al. [14]	2018	0.9633	0.7709	0.9848	0.9700
	Hu et al. [25]	2018	0.9632	0.7543	0.9814	0.9751
	Oliveira et al. [10]	2018	0.9365	0.6329	0.9924	–
	Wang, Chang, et al. [8]	2019	0.9538	0.7914	0.9722	0.9704
	Lv, Yan, et al. [21]	2020	0.9640	0.7598	0.9878	0.9640
	Tang, Yu, et al. [18]	2020	0.9723	0.7551	0.9903	0.9863
	Atli et al. [9]	2020	0.9682	0.6574	0.9933	0.9748
	Samuel et al. [3]	2021	0.9737	0.8738	0.9812	0.9905
	Sathananthavathi et al. [5]	2021	0.9445	0.8021	0.9561	–
	Manuel et al. [7]	2021	0.9754	–	–	0.9926
	Samuel et al. [3]	2021	0.9737	0.8738	0.9812	0.9905
	Proposed Method	2021	0.9834	0.8553	0.9835	0.9890
CHASE_DB1	Jiang et al. [11]	2018	0.9668	0.8640	0.9745	0.9810
	Oliveira et al. [10]	2018	0.9600	0.7731	0.9813	–
	Lv, Yan, et al. [21]	2020	0.9608	0.8176	0.9704	0.9608
	Atli et al. [9]	2020	0.9678	0.8011	0.9815	0.9833
	Sathananthavathi et al. [5]	2021	0.9340	0.6457	0.9653	–
	Manuel et al. [7]	2021	0.9663	–	–	0.9880
	Proposed Method	2021	0.9674	0.8276	0.9718	0.9611

Fig. 7(a)

Individual and combined dataset training of A-DFCNN model. A row belongs to DRIVE, STARE and CHASE_DB1 datasets, respectively. Columns shows individual and combined dataset training results.

Fig. 7(b)

Result image.

Table 3 shows the Sensitivity, Specificity and Accuracy obtained on each fold and its comparisons. Each is found to be closer in the STARE dataset compared to the DRIVE and CHASE_DB1 datasets. In k-fold cross-validation, there is not much variation in the performance metrics of all three datasets.

Table 3

Comparison of Spc, Sen and Acc on each fold

Testing fold	Acc.	Sen.	Spc.
1st fold	0.96	0.79	0.97
2nd fold	0.97	0.87	0.99
3rd fold	0.94	0.80	0.96
4th fold	0.97	0.81	0.99
5th fold	0.97	0.87	0.99
Average	0.96	0.82	0.98

The ROC values obtained after training aregiven in Figs. 8–10. The common issue in this method is to identify the vessel patterns for segmentation. Figure 11 depicted as visualize F1 score and all three datasets comparison of image segmentation for each image in test case dataset.

Fig. 8

Receiver operating characteristic curve (ROC) showing the maximum result of a proposed model at all segmentation thresholds in STARE dataset.

Fig. 9

ROC in the DRIVE dataset.

Fig. 10

ROC in CHASE_DB1 dataset.

Fig. 11

Comparison of image segmentation results for each image in the dataset a) DRIVE, b) STARE, and c) CHASE_DB1.

5 Conclusion

In this study, the proposed attention mechanism with a dense fully convolutional neural network shows better results in segmentation of optical vessel. The segmentation of vessels can be done with uncertainty evaluation and it is also compared with state of art methods for vessel segmentation. The public datasets CHASE DB1, STARE, and DRIVE were all evaluated to the suggested methodology. In this situation, the A-DFCNN outperforms other ground-breaking models across several datasets. By using skip connections and the attention method, it is proved that A-DFCNN method is achieved better performance in segmenting the low and high-resolution fundus images. This paves the way for novel area of research in the domain knowledge, in addition to the FCNN model.

Conflict of interest

No, “The authors declare no conflict of interest.”

References

Thanh

T.H.P.

et al., A Real-Time Classification of Glaucoma from Retinal Fundus Images Using AI Technology,}, 2020 International Conference on Advanced Computing and Applications (ACOMP) IEEE 2020.

Raja Siva

P.M.

, Sumithra

R.P.

, Thanusha

Automatic Glaucoma Diagnosis Based on Photo Segmentation with Fundus Images, 2021 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE) IEEE, 2021.

Samuel

P.M.

, Veeramalai

VSSC Net: Vessel specific skip chain convolutional network for blood vessel segmentation, Computer Methods and Programs in Biomedicine 198 (2021), 105769.

Malathi

, Gopinath

M.P.

Classification of pest detection in paddy crop based on transfer learning approach,, Acta Agriculturae Scandinavica, Section B—Soil & Plant Science 71(7) (2021), 552–559.

Sathananthavathi

, Indumathi

Encoder enhanced atrous (EEA) unet architecture for retinal blood vessel segmentation, Cognitive Systems Research 67 (2021), 84–95.

Khan

T.M.

et al., Residual connection-based encoder decoder network (RCED-Net) for retinal vessel segmentation, IEEE Access 8 (2020), 131257–131272.

Gegundez-Arias

M.E.

et al., A new deep learning method for blood vessel segmentation in retinal images based on convolutional kernels and modified U-Net model, Computer Methods and Programs in Biomedicine 205 (2021), 106081.

Wang

et al., Dense U-net based on patch-based learning for retinal vessel segmentation, Entropy 21(2) (2019), 168.

Atli

, Gedik

O.S.

Sine-Net: A fully convolutional deep learning architecture for retinal blood vessel segmentation, Engineering Science and Technology, an International Journal 24(2) (2021), 271–283.

10.

Oliveira

, Pereira

, Silva

C.A.

Retinal vessel segmentation based on fully convolutional neural networks, Expert Systems with Applications 112 (2018), 229–242.

11.

Jiang

et al., Retinal blood vessel segmentation using fully convolutional network with transfer learning, Computerized Medical Imaging and Graphics 68 (2018), 1–15.

12.

Hua

C.-H.

, Huynh-The

, Lee

Retinal vessel segmentation using round-wise features aggregation on bracket-shaped convolutional neural networks, 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) IEEE, 2019.

13.

Noh

K.J.

, Park

S.J.

, Lee

Scale-space approximated convolutional neural networks for retinal vessel segmentation, Computer Methods and Programs in Biomedicine 178 (2019), 237–246.

14.

Feng

et al., CcNet: A cross-connected convolutional network for segmenting retinal vessels using multi-scale features, Neurocomputing 392 (2020), 268–276.

15.

Zhao

et al., A nested U-shape network with multi-scale upsample attention for robust retinal vascular segmentation, Pattern Recognition 120 (2021), 107998.

16.

Hua

C.-H.

et al., Bimodal learning via trilogy of skip-connection deep networks for diabetic retinopathy risk progression identification, International Journal of Medical Informatics 132 (2019), 103926.

17.

Garifullin

, Lensu

, Uusitalo

On the uncertainty of retinal artery-vein classification with dense fully-convolutional neural networks, International Conference on Advanced Concepts for Intelligent Vision Systems, Springer,Cham,2020.

18.

Tang

et al., ResWnet for retinal small vessel segmentation, IEEE Access 8 (2020), 198265–198274.

19.

, Rahardja

BSEResU-Net: An attention-based before-activation residual U-Net for retinal vessel segmentation, Computer Methods and Programs in Biomedicine 205 (2021), 106070.

20.

Rahman

A.A.

et al., Robust segmentation of vascular network using deeply cascaded AReN-UNet, Biomedical Signal Processing and Control 69 (2021), 102953.

21.

et al., Attention guided U-Net with atrous convolution for accurate retinal vessels segmentation, IEEE Access 8 (2020), 32826–32839.

22.

Owen

C.G.

et al., Measuring retinal vessel tortuosity in 10-year-old children: Validation of the computer-assisted image analysis of the retina (CAIAR) program, Investigative Ophthalmology & Visual Science 50(5) (2009), 2004–2010.

23.

Kendall

, Gal

What uncertainties do we need in bayesian deep learning for computer vision, Advances in Neural Information Processing Systems 30 (2017).

24.

Gal

, Ghahramani

Dropout as a bayesian approximation: Representing model uncertainty in deep learning, International Conference on Machine Learning PMLR 2016.

25.

et al., Retinal vessel segmentation of color fundus images using multiscale convolutional neural network with an improved cross-entropy loss function, Neurocomputing 309 (2018), 179–191.

26.

et al., Automatic artery/vein classification using a vessel-constraint network for multicenter fundus images, Frontiers in Cell and Developmental Biology 9 (2021).

27.

Albert

J.R.

, Stonier

A.A.

Design and development of symmetrical super-lift DC–AC converter using firefly algorithm for solar-photovoltaic applications, IET Circuits Devices Syst 14(3) (2020),261–269 .https://doi.org/10.1049/iet-cds.2018.5292. .

28.

Shunmugham Vanaja

Albert

J.R.

, Stonier

A.A.

An Experimental Investigation on solar PV fed modular STATCOM in WECS using intelligent controller,Int Trans Electr Energ Syst 31(5) (2021), e12845.https://doi.org/10.1002/2050-7038.12845..

29.

Murugesan

, Kaliannan

, Balraj

, Singaram

, Kaliannan

, Albert

J.R.

A hybrid deep learning model for effective segmentation and classification of lung nodules from CT images, Journal of Intelligent and Fuzzy System 42(3) (2021), 2667–26791 10.3233/JIFS-212189.

30.

Albert

J.R.

et al., Investigation on load harmonic reduction through solar-power utilization in intermittent SSFI using particle swarm, genetic, and modified firefly optimization algorithms, Journal of Intelligent and Fuzzy System 42(4) (2022), 4117–4133. DOI: 10.3233/JIFS-212559.

31.

Vanchinathan

, Valluvan

K.R.

, Gnanavel

, Gokul

, Albert

J.R.

An improved incipient whale optimization algorithm based robust fault detection and diagnosis for sensorless brushless DC motor drive under external disturbances,, Int Trans Electr Energ Syst 31(12) (2021), e13251 DOI: 10.1002/2050-7038.13251.

32.

Ramaraju

S.K.

et al., Design and experimental investigation on VL-MLI intended for half height (H-H) method to improve power quality using modified particle swarm optimization (MPSO) algorithm, 42(6) (2022), 5939–5956DOI: 10.3233/JIFS-212583.

33.

Thangamuthu

, Albert

J.R.

, Chinnanan

, Gnanavel

Design and development of extract maximum power from single-double diode PV model for different environmental condition using BAT optimization algorithm, J Intell Fuzzy Syst 43(1) (2022), 1091–1102 https://doi.org/10.3233/JIFS-213241.

34.

Palanisamy

, Govindaraj

, Siddhan

, Albert

J.R.

Experimental investigation and comparative harmonic optimization of AMLI incorporate modified genetic algorithm using for power quality improvement, Journal of Intelligent and Fuzzy System 43(1) (2022), 1163–1176DOI: 10.3233/JIFS-212668.

35.

Albert

J.R.

Design and investigation of solar PV fed single-source voltage-lift multilevel inverter using intelligent controllers, Journal of Control, Automation and Electrical Systems. DOI: 10.1007/s40313-021-00892-w.

36.

Gnanavel

, Muruganatham

, Vanchinathan

, Albert

J.R.

ExperimentalValidation and Integration of Solar PV Fed Modular Multilevel Inverter (MMI) and Flywheel Storage System, 2021 IEEE Mysore Sub Section International Conference, 2021, pp. 147–153. Doi: 10.1109/MysuruCon52639.2021.9641650.

37.

Albert

J.R.

, Stonier

A.A.

, Vanchinathan

Testing and performance evaluation of water pump irrigation system using voltage-lift multilevel inverter, International Journal of Ambient Energy 43(1) (2022), 8162–8175DOI: 10.1080/01430750.2022.2092773.

38.

Albert

J.R.

et al., An advanced electrical vehicle charging station using adaptive hybrid particle swarm optimization intended for renewable energy system for simultaneous distributions, 43(4) (2022), 4395–4407.