Glaucoma disease detection using stacked attention U-Net and deep convolutional neural network

Abstract

Glaucoma is a condition of the eye that is caused by an increase in the eye’s intraocular pressure that, when it reaches its advanced stage, causes the patient to lose all of their vision. Thus, glaucoma screening-based treatment administered in a timely manner has the potential to prevent the patient from losing all of their vision. However, because glaucoma screening is a complicated process and there is a shortage of human resources, we frequently experience delays, which can lead to an increase in the proportion of people who have lost their eyesight worldwide. In order to overcome the limitations of current manual approaches, there is a critical need to create a reliable automated framework for early detection of Optic Disc (OD) and Optic Cup (OC) lesions. In addition, the classification process is made more difficult by the high degree of overlap between the lesion and eye colour. In this paper, we proposed an automatic detection of Glaucoma disease. In this proposed model is consisting of two major stages. First approach is segmentation and other method is classification. The initial phase uses a Stacked Attention based U-Net architecture to identify the optic disc in a retinal fundus image and then extract it. MobileNet-V2 is used for classification of and glaucoma and non-glaucoma images. Experiment results show that the proposed method outperforms other methods with an accuracy, sensitivity and specificity of 98.9%, 95.2% and 97.5% respectively.

Keywords

Medical image segmentation classification convolutional neural network U-Net MobileNet-V2

1 Introduction

The loss of retinal ganglion cells characterises glaucoma, an optic neuropathy that can have devastating consequences if left untreated. Optic nerve head (ONH) anatomical changes, specifically lamina cribrosa sheet thinning and posterior bowing, are clinically observable manifestations of this condition. A clinical decision can only be made after carefully observing and evaluating several clinical features relevant to glaucomatous optic neuropathy. Current methods for diagnosing and monitoring glaucoma involve extensive testing and a large amount of data that can be difficult to interpret, such as a comprehensive eye exam. In addition, early glaucoma patients and healthy subjects share many ocular characteristics. As a result, efforts are being made to create supplementary methods, such as Artificial intelligence systems, that can help tell the difference between true pathology and normal variability, as well as true progression and inter-test variability.

There is currently no cure for glaucoma, but early diagnosis and treatment can help a great deal. It is crucial in the creation of automated methods for early detection of glaucoma [1]. The vitreous, macula, retina, and blood vessels of the optic nerve can all be evaluated in a retinal fundus image. Retinal images were captured with a fundus camera by ophthalmologists. In the case of eye diseases like glaucoma, the retinal image was used for diagnosis. Glaucoma is a leading cause of blindness. The optic nerve head is responsible for sending signals from the retina to the brain [2]. Images of the retina and fundus are shown in Fig. 1.

Fig. 1

Retinal fundus image.

One of the most prominent structural image cues for glaucoma detection is the Cup-to-Disc Ratio (CDR) [3]. Most symptoms occur near the OD. Automated glaucoma detection methods that use segmented discs are sensitive to the accuracy of the segmentation, and even a slight error in outlining OD can have an impact on the diagnosis [4]. On the other hand, localization provides information regarding the precise location of OD within the overall image as well as the context of its surroundings. Automatic methods for the detection of glaucoma that are based on this method of region of interest extraction are more resistant to errors in localization.

The disease pattern in retinal fundus images is subtle and complicated, making automated classification difficult. Because of its readily apparent colour, shape, texture, etc. in natural scene images, segmentation is a relatively straightforward task. In contrast, the most telling signs of illness in medical images are often obscured and only visible to specialists with years of experience in the field. On the other hand, it has been demonstrated that Deep Learning is capable of learning a discriminative representation of data that is able to identify subtle differences [5]. This allows the representation to be both compact and useful. Computer-assisted diagnosis (CAD) tools like these can greatly improve the quality and efficiency of mass screening programmes. Possible benefits of these automated systems include reduced human error, timely service delivery in underserved areas, and the absence of bias and chronic stress in medical professionals.

A CNN automatically figures out how to classify things based on their many features, without having to separate the fuzzy optical cups. Most of these methods, however, don’t have enough training data, which leads to the overfitting problem. Transfer learning is used when there isn’t enough data. This work moves the task of classifying images of nature to finding glaucoma on images of the fundus. Compared to images of nature, the fundus images have a lot of areas that are the same and don’t give any useful information for detecting glaucoma. The redundancy regions could trick CNN into focusing on information that isn’t important. To overcome this problem, we proposed an automatic detection of Glaucoma disease. In this proposed objective is as follows:

The first stage is based on Optic Disc and Optic Cup segmentation which is responsible for segmenting and extracting optic disc from a retinal fundus image using Sacked Attention based U-Net architecture.

While the second stage uses MobileNetV2 to classify the extracted disc into healthy or glaucomatous.

The proposed method is evaluated experimentally and compared to other existing methods.

The remainder of the paper is organized in the following manner: Section II discusses works that are related to this study. Section III goes into wider context about the proposed method. Section IV discusses the experimental design, evaluation measures, and results obtained from the experiment. Finally, a conclusion is provided.

2 Related study

Recently, image processing and machine learning-based computer-aided systems for intelligent glaucoma detection have gained popularity [6, 7]. For a given dataset of retinal images, supervised machine learning algorithms are used for classification of the normal image and glaucomatous image, while unsupervised machine learning algorithms are primarily used for segmenting the disc and cup in the improved retinal image. Researchers have proposed new retinal image glaucoma detection methods in recent years using multi-layer perceptron, Random Forest, and Radial Basis Function classifiers [8, 9]. Heuristics and deep learning are the two primary categories of diagnostic methods that can be used to detect glaucoma. The extraction of features is done through many of the heuristic glaucoma detection methods by using various image processing techniques. Screening techniques for glaucoma have been developed that involve measuring the thickness of the retinal nerve fibre layer (RNFL) [10]. In [11], an attempt was made to facilitate the diagnosis of glaucoma by extracting features of the texture as well as higher order spectral features. In [12], an attempt was made to diagnose glaucoma using energy features that were derived from wavelets. [11, 12] utilised both a support vector machine (SVM) and a naive Bayesian classifier in order to classify the human-created features. However, the aforementioned heuristic approaches only take into account a small subset of the features that can be found on fundus images, which results in a classification accuracy that is not particularly high.

Deep learning-based methods are another class of glaucoma detection techniques. In particular, work on glaucoma detection using deep learning based on automatic segmentation of the optic cup and disc has been reported in [13, 14]. However, without comprehensive education, their work focuses solely on the optic disc and how they may or may not be connected to glaucoma. In addition, [15] suggested a multi-stream CNN that merged the final segmentation result with the full optical images. When it comes to detecting glaucoma, a deep learning approach based on RNFL defect was proposed by [33]. However, [16] was the first to propose a comprehensive CNN approach to glaucoma detection. Chen’s work was built upon in [17], where an improved CNN structure was proposed for glaucoma classification using a combination of global and local features. Both [16, 17] performed a pre-processing step on the primary fundus images, in which they removed redundant regions in order to regularise the input images. However, previous works did not succeed in achieving high sensitivity and specificity because there was a lack of training data and the networks themselves were overly simplistic.

The optic cup in retinal fundus images was segmented using the U-Net segmentation method, which was developed by [18]. Better results in detecting glaucoma can be attained through segmentation of the optic cup and optic disc. In this case, the optic disc image served as the basis for the ROI, which was then cropped and segmented using the U-Net algorithm [18]. To better detect glaucoma proposed [17], that is attention-based convolutional neural network (AG-CNN) model, which they are tested on the large-scale attention-based glaucoma database (LAG). The accuracy and reliability of glaucoma identification may suffer if large amounts of redundancy are removed from fundus images. The AG-CNN model thought about this, and then made a call. This model is the result of merging subnets for attention prediction, localization of pathological regions, and classification. The model has a 96.2% sensitivity and a 0.983 AUC for identifying glaucoma. In a few instances, only a portion of the ROI was highlighted, making it impossible to pinpoint the precise location of the problem [19].

The convolutional neural network model for retinal image classification was proposed by [20]. Rather than using the images themselves as input, this system uses the features extracted from the images to feed a CNN model, which then classifies the images as normal or abnormal. In order to categorise retinal images practically verified [21] in hybrid graph convolutional network (HGCN). With this network, features from a convolutional neural network (CNN) are integrated into a design for graph learning based on modularity, yielding a graph convolutional network. A proposed ensemble learning based model for retinal image classification presented [22]. In this proposed deep learning-based model for retinal image classification using U-Net and Inception V3 network [23]. Transfer learning-based mode with saliency maps are proposed in [24]. A proposed glaucoma detection model using U-Net and Efficient net presented [25]. There’s no need to manually separate the fuzzy optical cups because a CNN can figure out how to classify things based on their many features. There are various methods of U-Net has been proposed for medical image segmentation namely U-Net [26], 3D U-Net [27], Attention U-Net [28], CE-Net [29], U-Net++ [30] and Trans U-Net [31 –38]. However, most of these approaches suffer from insufficient training data, which in turn causes the overfitting issue [39 –46]. When there isn’t enough data, we turn to transfer learning. This work shifts the focus from categorising nature images to detecting glaucoma in fundus images. The fundus images provide no useful information for detecting glaucoma because they have many similar areas compared to nature images. Fundus images’ black background and the eyeball’s periphery are two such examples. The redundancy regions may mislead CNN into prioritising irrelevant data. In order to resolve this issue, we proposed a stacked attention based U-Net segmentation and automatic Glaucoma disease classification using MobileNet-v2.

3 Proposed model

In medical imaging, target structures often vary in size, shape, and texture from one class to the next. The local receptive field of conventional CNNs used for segmentation yields local feature representations. In order to overcome this problem, we proposed an automatic detection of glaucoma. We proposed this model in two stages (Fig. 1). The first stage is based on Optic Disc Segmentation, which is responsible for locating and extracting the optic disc from an image of the retinal fundus by using a Sacked Attention-based U-Net architecture. This stage is based on the optic disc segmentation algorithm. The extracted disc is classified as healthy or glaucomatous in the second stage using MobileNetV2.

3.1 Stacked attention based U-Net

U-Net can be broken down into two sections: The first is the conventional neural network (CNN)-based contracting route presented in Fig. 2(a b). In Fig. 2(c) p and s represents strides and padding respectively, this structure shows that proposed U-Net model internal process. Each shrinking path block is made up of two 3 x 3 convolutions, an activation unit using the ReLU function, and a max-pooling layer. There are multiple occurrences of this pattern. The second section of U-Net, known as the expansive path, is where the algorithm’s novelty lies. In this section, each stage uses a 2 x 2 up-convolution to upsample the feature map. After that, the up-sampled feature map gets cropped, and then it gets joined with the feature map from the layer that corresponds to it in the reducing path. Next, we employ ReLU activation followed by two consecutive 3x3 convolutions. Finally, the feature map is convolutionally reduced to produce the segmented image using only one channel. Pixel features near the image’s borders provide the least amount of context and must be removed before the image can be used.

Fig. 2(a)

Proposed Model.

Fig. 2(b)

Internal Architecture.

Fig. 2(c)

Internal process present U-Net model.

The energy function for the network is given by: $E = \sum u (x) \log (p_{k} (x))$ (1) where p_k represents the final feature map with the SoftMax function applied to each individual pixel: $p_{k} = \frac{exp (b_{k} (x))}{\sum_{i = 1}^{k} exp (b_{k} {(x)}^{'})}$ (2)

And b_k denotes the activation function in channel k. The expression for the loss function can be written as follows: $Loss = \sum_{i} y_{i} {logx}_{i} + (1 - y_{i}) \log (1 - x_{i})$ (3)

Where y_i denotes the model output and the corresponding true label represented as x_i.

Unless the relevant context is encoded, there is a chance that features representing adjacent pixels with the same label will differ from one another in their local feature representations. The result could be inconsistency within the class, which would have a negative impact on recognition accuracy. In order to solve this problem, we investigate the cognitive processes that lay the groundwork for associating different features with one another. In order to incorporate features learned at varying levels of complexity, a stack of spatial and channel self-attention modules was developed.

In this context, features that are manifested across multiple scales are referred to as P_n, where n stands for the level in the architecture. Because features are available at varying resolutions for each level n. This results in expanded feature maps $P_{n}^{'}$ . The $P_{n}^{'}$ from each of the scales are then concatenated together to form a tensor, which is then convolved to produce a multiscale feature map with a common name, $FM = conv ([P_{0}^{'}, P_{1}^{'}, P_{2}^{'}, P_{3}^{'}])$ . As a result, it stores not only the high-level semantics learned in the lower layers, but also the low-level details learned in the upper layers. The layers are structured in a hierarchical structure. The stacked attention modules are then given the new stacked feature map and each of the scale-specific feature maps to generate the stacked attention features, S_n: $S_{n} = attenMod (conv ([P_{n}^{'}, FM]))$ (4) where attenMod refers to each individual attention module that is stacked. The combination of FM at each layer allows for the joint encoding of parallel low-level information and high-level semantics from FM, leading to a more stable representation.

Let input feature map to the attention module is denoted by P ∈ R^cxlxb., where c, landb. represents channel, width and height respectively. Input feature map is passed through the convolutional block resulting in a map $P_{0}^{s}$ , follows the same operation before transpose resulting in $P_{1}^{s}$ . To create the spatial attention map SA^s, we multiply both maps and then apply softmax to the resulting matrix. ${SA}_{i, j}^{s} = \frac{exp (P_{0, i}^{s}, P_{0, j}^{s})}{\sum_{i = 1}^{lxb} exp (P_{0, i}^{s}, P_{1, j}^{s})}$ (5)

In the next step, the input is fed into the other layer resulting in $P_{2}^{s}$ . As a result, the SA_p can be formulated as follows, which is the resultant feature vector corresponding to the spatial attention module. ${SA}_{p} = d \sum_{i = 1}^{lxb} {SA}_{i, j}^{s} P_{2}^{s} + P_{j}$ (6) where, d is set to 0 to indicate no preference for the spatial attention map and is subsequently trained to prefer it. For this reason, the spatial attention module uses the spatial attention map as a guide to selectively aggregate global context to the acquired features.

Channel attention modules have three branches like spatial attention modules. $P_{o}^{c}$ and $P_{1}^{c}$ are the input feature maps. These two maps are multiplied and resulting in a channel attention map $({CA}_{i, j}^{c})$ , ${CA}_{i, j}^{c} = \frac{exp (P_{0, i}^{c}, P_{0, j}^{c})}{\sum_{i = 1}^{lxb} exp (P_{0, i}^{c}, P_{1, j}^{c})}$ (7)

The transposed version of the input feature is multiplied and the final channel attention map is generated as given in below equation, ${CA}_{p} = e \sum_{i = 1}^{lxb} {CA}_{i, j}^{c} P_{2}^{c} + P_{j}$ (8) where c regulates how much weight is placed on the channel attention map as opposed to the input feature map F. Like d, e starts at 0 and is learned over time which adds weighted versions of the features from all the channels.

In this stacked attention model, we propose incorporating sequential refinement modules that progressively improve attentive features. The theory is that by making small adjustments one after another, noise can be suppressed while the significance of various local regions is enhanced. In the first step of the process, which is the generation of self-attention features, F is used by the spatial and channel attention modules. In addition, we incorporate an encoder-decoder network that takes the input features P and outputs a representation in the latent space. To incorporate the class data into the next set of stacked attention modules. This can be done by making sure that the representation of encoder-decoders is as close as possible, which can be formulated as follows: $LR = \sum_{i}^{m} e_{i} (P_{A}^{i - 1}) - e_{i + 1} {(P_{A}^{i})}_{2}^{2}$ (9)

$P_{A}^{i}$ stands for the attention features produced by the i-th dual attention module, and m is the number of iterations; e_i is the encoded representation of the i^th encoder-decoder network and $P_{A}^{i - 1}$ represents the features fed as input into the stacked model. The first attention module generates self-attention features and then combines them with the reconstructed feature maps from the first encoder-decoder to produce stacked attention. The reconstructed feature map is given by equation (10), $P_{rec} = \sum_{i}^{m} P_{i} - {\hat{P_{i}}}_{2}^{2}$ (10)

Where $\hat{P_{i}}$ represents reconstructed feature map. Loss for both spatial and channel attention modules LR^s as given by, $L = \sum_{s} {LR}^{s}$ (11)

Loss for reconstructed feature map is given by, $L = \sum_{s} P_{rec}^{s}$ (12)

3.2 Pre-trained MobileNetV2

Transfer learning uses a model to learn a new task. This method uses a pre-trained model on a large dataset to complete tasks on other datasets. Transfer learning is popular because it accurately classifies small datasets. A deep learning model trained from scratch on a small dataset lacks data variation information, making it difficult to achieve high accuracy. The MobileNetV2 network is the base of this methodology, having been pre-trained on the ImageNet dataset. On top of the MobileNetV2’s convolutional layers, we implement a new set of computations known as the head model. The output of the base model is fed into the first layer of the head model, a global pooling layer, which produces a feature map with 7x7x1280 pixels. In order to significantly reduce the dimensionality of the data, the global pooling layer performs a pooling operation to produce a one-dimensional feature vector. The proposed method generates a 1x1x1280 pixel output feature map using an average pooling operation with a 7x7 pixel kernel size in the global pooling layer [47 –53]. Two additional layers, both of which are fully connected to one another, follow the global pooling layer. ReLU activation function activated 128 and 64 nodes in the fully connected layers. Because we use one-hot encoding and the dataset has two classes-glaucoma and non-glaucoma output layer of the two nodes. Softmax activates the output layer. The 1x1 convolutional layer is the only one without a batch normalisation layer and an activation function (ReLU6). As the output of the convolutional layer is low dimensional, it only has batch normalisation. Figure 3 displayed that sample image datasets described in this work.

Fig. 3

Sample dataset images [24].

4 Experiemental analysis

Several experiments were performed to calculate the glaucoma identification and the results of these experiments are discussed in detail here. We have used a publicly available database (ORIGA) to test the dimensional stability of proposed method for detecting and classifying glaucoma cases and implemented in Matlab. The ORIGA database contains 650 samples, 168 of which show the glaucoma-affected regions of human eyes, and the remaining 650 show healthy human eyes. The ORIGA dataset is notoriously difficult to use for glaucoma classification due to the prevalence of artefacts within its samples. These include, but are not limited to, large variations in the size, colour, position, and texture of OD and OC. In addition, there are numerous types of distortion in images, including noise, blurring, colour and intensity variations. Fig. 3 displays some examples from the used data set. Training and testing are in the ratio of 80 % and 20% respectively.

Multiple evaluation metrics, including Intersection over Union (IoU), accuracy, precision, recall, and mean average precision (mAP), are used to evaluate in this method’s localization and categorization results in this work. Equation is used to determine accuracy is given by Equation (13), $Accuracy = \frac{True Positive + True Negative}{True Positive + True Negative + False negative + Fale Positive}$ (13)

The formula for determining the mAP score is given in equation (14) as follows: where AP represents the mean precision across all classes and q stands for the sample size. The letter Q also stands for the entire number of subjects in a test. $mAP = \sum_{i = 1}^{q} AP (q_{i}) / Q$ (14)

Other metrics are gives as follows, $Precision = \frac{True Positive}{True Positive + False Positive}$ (15) $Recall = \frac{True Positive}{True Positive + False Negative}$ (16) $IoU = \frac{True Positive}{True Positive + False Positive + False Negative}$ (17)

For an efficient computer-aided approach to identifying and classifying glaucoma affected regions, timely and accurate identification of the optical disk head lesions is essential. Because of this, we decided to design an experiment to evaluate the localization capability of stacked attention-based U-Net by testing against the ORIGA database. The results of this experiment are depicted in Fig. 4, and they were obtained from the experiment. Experimental results state clearly that the proposed solution, which is known as stacked attention-based U-Net, is able to correctly diagnose OD and OC lesions of a wide range of sizes and positions. In addition to that, present work is able to deal with a wide variety of sample distortions, such as blurring, colour variations, and variations in brightness.

Fig. 4

Segmented image samples.

The proposed method is able to precisely identify lesions with fewer signs because of its localization capabilities. Since mAP and IoU are the most widely used evaluation measures among researchers, we rely on them to quantitatively assess the efficacy of our approach in localization and segmentation. We find that proposed method yields mean IoU values of 0.979 and mAP values of 0.98 on average. The visual and numerical results demonstrate the effectiveness of our framework in identifying and classifying glaucoma hotspots. Figures 5 6 depicted as accuracy and loss function of proposed work.

Fig. 5

Accuracy.

Fig. 6

Loss.

We have conducted experiments to compare proposed framework’s glaucoma recognition results to those obtained by using alternative object detection methods, such as U-Net [26], 3D U-Net [27], Attention U-Net [28], CE-Net [29] U-Net++ [30], and Trans U-Net [31 –36]. The results of this analysis is shown in Table 1. We have used the mean absolute percentage (mAP) and IoU evaluation metrics to make comparisons with other segmentation methods. It is clear from Table 1 that proposed stacked attention-based U-Net achieves the highest mAP value of 0.981 and the highest IoU of 0.979. The conventional U-Net method also achieves the lowest mAP value (0.923) and the lowest IoU (0.929). In addition, the TransNet yields findings that are consistent described. The computational advantage provided by proposed stacked attention mechanism makes presemt work clearly superior to the alternatives in terms of detecting and classifying glaucoma lesions. Further, the framework is able to precisely localise the region of interest and achieve the highest mAP value relative to its competitors because of the reliable feature detection provided by the proposed method.

Table 1

Comparison with other segmentation methods

Method	mAP	IoU
U-Net[26]	0.923	0.929
3D U-Net[27]	0.933	0.934
Attention U-Net [28]	0.935	0.943
CE-Net [29]	0.943	0.955
U-Net++[30]	0.965	0.965
TransNet [31]	0.975	0.966
Proposed Method	0.981	0.979

After segmentation of optical disc, pre-trained MobileNetV2 is used for classification. Accuracy and loss model is given in Figs. 5 6 respectively. In addition, as shown in Table 2, we compared with proposed segmentation to various traditional methods. A stacked attention-based U-Net segmentation method produces superior results with a pre-trained model, as shown by the analysis.

Table 2

Proposed segmentation with various methods

Methods	Precision	Recall	Accuracy	Sensitivity	Specificity
AlexNet	0.902	0.899	0.900	0.898	0.911
ResNet - 18	0.920	0.919	0.912	0.910	0.902
ResNet - 34	0.934	0.929	0.923	0.921	0.913
VGGNet	0.943	0.933	0.941	0.953	0.928
MobileNet V1	0.959	0.949	0.959	0.960	0.951

Further, we have conducted a comparative analysis, selecting the most up-to-date methods using the same dataset, to verify this approach glaucoma identification and classification performance. In order to be objective, we have compared the outcomes of present method with those of the techniques detailed in [20 –25]. Table 3 displays the quantitative comparisons made by using evaluation metrics with different methods.

Table 3

Comparison with the State-of-art

Methods	Precision	Recall	Accuracy	Sensitivity	Specificity
Deep CNN [20]	0.932	0.922	0.934	0.901	0.912
Graph CNN [21]	0.939	0.921	0.938	0.919	0.922
Ensembling [22]	0.940	0.934	0.935	0.920	0.933
U-Net+Inception V3 [23]	0.959	0.949	0.946	0.954	0.943
ODGNet [24]	0.960	0.954	0.965	0.965	0.956
U-Net+Efficient Net [25]	0.963	0.965	0.978	0.970	0.961
Proposed Model	0.974	0.973	0.989	0.979	0.987

Figures 7 , 10 and 11 shows that proposed glaucoma detection model outperforms state-of-the-art methods. We also compared with and without attention mechanism presented in Table 4 and Fig. 12. An opposed to proposed method with others to use extremely complex and deep networks towards compute features, which leads to the model over-fitting problem and the loss of spatial features. Whereas, in comparison, proposed approach employs stacked attention-based U-Net as a base segmentation and localization network, which is capable of localize optical disc head and optical cup. It follows that proposed architecture offers a practical and efficient method for optical disk and optical cup recognition, which can aid physicians in making a prompt diagnosis of glaucoma-affected areas.

Fig. 7

Accuracy.

Fig. 8

Precision.

Fig. 9

Recall.

Fig. 10

Sensitivity.

Fig. 11

Specificity.

Table 4

Stacked attention model analysis

Metrics	With attention mechanism	Without attention mechanism
Precision	0.974	0.952
Recall	0.973	0.969
Sensitivity	0.979	0.952
Specificity	0.987	0.975
Accuracy	0.989	0.964
mAP	0.981	0.966
IoU	0.979	0.959

Fig. 12

Comparison with attention mechanism.

5 Conclusion

A Deep CNN is proposed as a revolutionary strategy for glaucoma diagnosis and prediction in this research work. The glaucoma dataset was used to train the U-Net and DCNN model for glaucoma image analysis. Less than a quarter of the data was used for this research, with the rest of U-Net classification being used in this proposed segmentation.

To segregate the data was used to extract the features. Then, it will pre-trained U-Net transfer learning model linked with DCNN. A deep convolutional neural network was used to classify the images in order to diagnose glaucoma. These retinal fundus photographs were used to determine whether or not the patient had glaucoma. Identifying the areas of the fundus sample that have been affected by glaucoma requires trained human experts who can recognise subtle visual differences and classify the images into the appropriate categories. However, due to the complexity of glaucomatous regions and the difficulty of gaining access to domain experts, a fully automated system is required. In this paper, we proposed an automatic detection of glaucoma using stacked attention-based U-Net and pre-trained MobileNet2. The stacked attention-based U-Net is used for optic cup and optic head segmentation, and the segmented image is then fed to the pre-trained MobileNet2 model for classification. With an accuracy of 98.9%, the proposed approach outperforms other methods. In the future, Generative adversarial networks may be utilised to overcome the limited datasets that are currently available, and proposed method may also be applied to other publicly accessible datasets.

References

Abdullah Sarhan , Jon Rokne and Reda Alhajj , Glaucoma detection using image processing techniques: A literature review, Computerized Medical Imaging and Graphics 78 (2019), 101657.

Anindita

and Agus

, Automatic glaucoma detection based on the type of features used: a review, Journal of Theoretical and Applied Information Technology 72(3) (2015), 366–375.

Qaisar Abbas , Glaucoma-deep: detection of glaucoma eye disease on retinal fundus images using deep learning, International Journal of Advanced Computer Science and Applications 8(6) (2017).

Muthu Rama Krishnan Mookiah , Rajendra Acharya

, Choo Min Lim , Andrea Petznick and Jasjit Suri

, Data mining technique for automated diagnosis of glaucoma using higher order spectra and wavelet energy features, Knowledge-Based Systems 33 (2012), 73–82.

Yoshua Bengio , Aaron Courville and Pascal Vincent , Representation learning: A review and new perspectives, IEEE Transactions on Pattern Analysis and Machine Intelligence 35(8) (2013), 1798–1828.

Shishir Maheshwari , Ram Bilas Pachori and Rajendra Acharya

, Automated diagnosis of glaucoma using empirical wavelet transform and corr-entropy features extracted from fundus images, IEEE Journal of Biomedical and Health Informatics 21(3) (2016), 803–813.

Abhishek Dey and Samir Bandyopadhyay

, Automated glaucoma detection using support vector machine classification method, British Journal of Medicine and Medical Research 11(12) (2016).

Maíla Claro , Leonardo Santos , Wallinson Silva , Flávio Araújo , Nayara Moura and André Macedo , Automatic glaucoma detection based on optic disc segmentation and texture feature extraction, Clei Electronic Journal 19(2) (2016), 5–5.

Anju Soman and Deepthy Mathew , Glaucoma detection and segmentation using retinal images, International Journal of Science, Engineering and Technology Research 5(5) (2016), 1346–1349.

10.

Odstrcilik , Jan , Radim Kolar , Ralf-Peter Tornow , Jiri Jan , Attila Budai , Markus Mayer , Martina Vodakova , et al., Thickness related textural properties of retinal nerve fiber layer in color fundus images, Computerized Medical Imaging and Graphics 38(6) (2014), 508–516.

11.

Rajendra Acharya

, Sumeet Dua , Xian Du and Chua Kuang Chua , Automated diagnosis of glaucoma using texture and higher order spectra features, IEEE Transactions on Information Technology in Biomedicine 15(3) (2011), 449–455.

12.

Sumeet Dua , Rajendra Acharya , Pradeep Chowriappa

and Vinitha Sree

, Wavelet-based energy features for glaucomatous image classification, IEEE Transactions on Information Technology in Biomedicine 16(1) (2011), 80–87.

13.

Jun Cheng , Jiang Liu , Yanwu Xu , Fengshou Yin , Damon Wing Kee Wong , Ngan-Meng Tan , Dacheng Tao , Ching-Yu Cheng , Tin Aung and Tien Yin Wong , Superpixel classification based optic disc and optic cup segmentation for glaucoma screening, IEEE Transactions on Medical Imaging 32(6) (2013), 1019–1032.

14.

Julian Zilly , Joachim Buhmann

and Dwarikanath Mahapatra , Glaucoma detection using entropy sampling and ensemble learning for automatic optic cup and disc segmentation, Computerized Medical Imaging and Graphics 55 (2017), 28–41.

15.

Huazhu Fu , Jun Cheng , Yanwu Xu , Changqing Zhang , Damon Wing Kee Wong , Jiang Liu and Xiaochun Cao , Disc-aware ensemble network for glaucoma screening from fundus image, IEEE Transactions on Medical Imaging 37(11) (2018), 2493–2501.

16.

Xiangyu Chen , Yanwu Xu , Damon Wing Kee Wong , Tien Yin Wong and Jiang Liu , Glaucoma detection based on deep convolutional neural network. In 2015 37th annual international conference of the IEEE engineering in medicine and biology society (EMBC), pp. 715–718. IEEE, 2015.

17.

Annan Li , Jun Cheng , Damon Wing Kee Wong and Jiang Liu , Integrating holistic and local deep features for glaucoma classification, In 2016 38th annual international conference of the IEEE engineering in medicine and biology society (EMBC), pp. 1328–1331. IEEE, 2016.

18.

Prastyo

P.H.

, Sumi

A.S.

and Nuraini

, Optic cup segmentation using U-net architecture on retinal fundus image, Journal of Information Technology and Computer Engineering 4(2) (2020), 105–109.

19.

, Xu

, Liu

, et al., A large-scale database and a CNN model for attention-based glaucoma detection, IEEE Transactions on Medical Imaging 39(2) (2020), 413–424.

20.

Noha El-Hag

, Ahmed Sedik , Walid El-Shafai , Heba El-Hoseny

, Ashraf Khalaf

A.M.

, Adel El-Fishawy

, Waleed Al-Nuaimy , Fathi Abd El-Samie

and Ghada El-Banby

, Classification of retinal images based on convolutional neural network, Microscopy Research and Technique 84(3) (2021), 394–414.

21.

Guanghua Zhang , Jing Pan , Zhaoxia Zhang , Heng Zhang , Changyuan Xing , Bin Sun and Ming Li , Hybrid graph convolutional network for semi-supervised retinal image classification, IEEE Access 9 (2021), 35778–35789.

22.

Niloy Sikder , Mehedi Masud , Anupam Kumar Bairagi , Abu Shamim Mohammad Arif , Abdullah-Al Nahid and Hesham Alhumyani

, Severity classification of diabetic retinopathy using an ensemble learning algorithm through analyzing retinal images, Symmetry 13(4), 670, 2021.

23.

Anas Bilal , Liucun Zhu , Anan Deng , Huihui Lu and Ning Wu , AI-Based Automatic Detection and Classification of Diabetic Retinopathy Using U-Net and Deep Learning, Symmetry 14(7) (2022), 1427.

24.

Jahanzaib Latif , Shanshan Tu , Chuangbai Xiao , Sadaqat Ur Rehman , Azhar Imran and Yousaf Latif , ODGNet: a deep learning model for automated optic disc localization and glaucoma classification using fundus images, SN Applied Sciences 4(4) (2022), 98.

25.

Mir Tanvir Islam , Shafin Mashfu

, Abrar Faisal , Sadman Chowdhury Siam , Intisar Tahmid Naheen and Riasat Khan , Deep learning-based glaucoma detection with cropped optic cup and disc and blood vessel segmentation, IEEE Access 10 (2021), 2828–2841.

26.

Ronneberger

, Fischer

and Brox

, U-net: convolutional networks for biomedical image segmentation, in Proceedings of the Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. Wells, and A. Frangi, Eds., Springer, Munich, Germany, October 2015.

27.

Ciçek

Ö.

, Abdulkadir

, Lienkamp

S.S.

, Brox

and Ronneberger

, 3D U-net: learning dense volumetric segmentation from sparse annotation, in Proceedings of the Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, S. Ourselin, L. Joskowicz, M. Sabuncu, G. Unal, and W. Wells, Eds., Springer, Athens, Greece, October, 2016.

28.

Oktay

, Schlemper

, Folgoc

, et al., Attention U-Net: Learning where to Look for the Pancreas, in Proceedings of the 1st Conference on Medical Imaging with Deep Learning, Amsterdam, The Netherlands, July 2018.

29.

Long

, Shelhamer

and Darrell

, Fully Convolutional Networks for Semantic Segmentation, in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, Boston, MA, USA, June 2015.

30.

Zhou

, Siddiquee

M.M.R.

, Tajbakhsh

and Liang

, UNet++: a nested U-net architecture for medical image segmentation, Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support 11045 (2018), 3–11.

31.

Chen

, Lu

, Yu

, et al., TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation, 2021.

32.

Malathi Murugesan , Madonna Arieth

, Shankarlal Balraj and Nirmala

, Colon cancer stage detection in colonoscopy images using YOLOv3 MSF deep learning architecture, Biomedical Signal Processing and Control 80(Part 2) (2023), 104283. https://doi.org/10.1016/j.bspc.2022.104283

33.

Malathi

, Kalaiselvi

, Kokila

, Sankarlal

and Thenmalar

, A Hybrid Deep Learning Model for Effective Segmentation and Classification of Lung Nodules from CT Images, Journal of Intelligent & Fuzzy Systems 42(3) (2022), 2667–2679.

34.

Selvaganesh

and Ganesan

, A Hybrid segmentation andclassification techniques for detecting the neurodegenerative disorder from brain Magnetic Resonance Images, Springer Nature 81 (2022), 28801–28822.

35.

Ramachandran Lakshmanan and Mohan Veerasamy , A novel neural network model for shrimp segmentation to detect white spot syndrome, Journal of Intelligent & Fuzzy Systems 43(1) (2022), 1453–1466.

36.

NuthalSrinivasan

and Chinnadurai

, An Efficient video inpainting approach using deep belief network, Computer System Science and Engineering 43(2) (2022), 515–529.

37.

Albert

J.R.

and 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

38.

Johny Renoald Albert , 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.

39.

Vanchinathan

, Valluvan

K.R.

, Gnanavel

, Gokul

and 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.

40.

Satish Kumar Ramaraju , 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), 5939–5956. DOI: 10.3233/JIFS-212583

41.

Logeswaran Thangamuthu , Johny Renoald Albert , Kalaivanan Chinnanan and Banu 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

42.

Rajarathinam Palanisamy , Vijayakumar Govindaraj , Saravanan Siddhan and Johny Renoald Albert , 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), 1163–1176. DOI: 10.3233/JIFS-212668.

43.

Albert

J.R.

, Design and Investigation of Solar PV Fed Single-Source Voltage-Lift Multilevel Inverter Using Intelligent Controllers, J Control Autom Electr Syst 33 (2022), 1537–1562. https://doi.org/10.1007/s40313-021-00892-w

44.

Gnanavel.

, Muruganatham.

, Vanchinathan

and 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

45.

Renoald

A.J.

, Stonier Albert Alexander and Vanchinathan Kumarasamy, Testing and Performance Evaluation of Water Pump Irrigation System using Voltage-Lift Multilevel Inverter, International Journal of Ambient Energy, 2022, pp. 1–14, DOI: 10.1080/01430750.2022.2092773

46.

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. DOI: 10.3233/JIFS-220089.

47.

Albert

J.R.

, Kannan.

, Karthick.

, Selvan.

, Sivakumar.

and Gnanavel.

, An Experimental and Investigation on Asymmetric Modular Multilevel Inverter an Approach with Reduced Number of Semiconductor Devices, J Electrical Systems 18(3) (2022), 318–330.

48.

Babypriya

, Renoald

A.J.

, Shyamalagowri

and Kannan

, An Experimental Simulation Testing of Single-diode PV Integrated MPPT Grid-tied Optimized Control Using Grey Wolf Algorithm, 1 Jan. 2022:1–20, DOI: 10.3233/JIFS-213259

49.

Madhumathi Periasamy , Thenmalar Kaliannan , Shobana Selvaraj , Veerasundaram Manickam , Sheela Androse Joseph and Albert

J.R.

, Various PSO methods investigation in renewable and nonrenewable sources, International Journal of Power Electronics and Drive Systems 13(4) (2022), 2498–2505. DOI: 10.11591/ijpeds.v13.i4.pp2498-2505

50.

Albert

J.R.

and Dishore Shunmugham Vanaja , Solar Energy Assessment in Various Regions of Indian Sub-continent, Solar Cells – Theory, Materials and Recent Advances, Intech Open, (December 10th 2020). DOI: 10.5772/intechopen.95118

51.

Albert

J.R.

, et al., A Remote Diagnosis Using Variable Fractional Order with Reinforcement Controller for Solar-MPPT Intelligent System, Photovoltaic Systems pp. 45–64, Publisher: CRC press.

52.

Kaliannan

, Albert

J.R.

, Begam

D.M.

and Madhumathi

, Power Quality Improvement in Modular Multilevel Inverter Using for Different Multicarrier PWM, European Journal of Electrical Engineering and Computer Science 5(2) (2021), 19–27. https://doi.org/10.24018/ejece.2021.5.2.315

53.

Renoald

A.J.

, Muhamadha Begam

and Nishapriya

, Micro grid connected solar PV employment using for battery energy storage system, Journal of Xidian University 15(3) (2021), 85–97. https://doi.org/10.37896/jxu15.3/010