FGS-HDNN: Fractional Gravitational Search based Hybrid Deep Neural Network for Glaucoma Detection using Fundus Images

Glaucoma is a retinal disease that damages the eye's optic nerve, frequently causing an irreversible loss of vision. However, the accurate diagnosis of this disease is difficult but early-stage diagnosis may cure this retinal disease. The objective of this research is to diagnose glaucoma disease in the top of the eye's optical nerve. The proposed approach detects glaucoma via four major steps namely Data enhancement phase, segmentation phase, feature extraction phase, and classification phase by the fractional gravitational search-based hybrid deep neural network (FGSA-HDNN) classifier. The proposed classifier is used for the exact classification of glaucoma infected images and normal images. Here, the proposed approach utilizes the statistical, textural, and vessel features from the segmented output. Also, the proposed FGSO algorithm is used for testing the deep neural network. From the experimental results, it is observed that the proposed glaucoma detection has obtained a sensitivity of 99.64%, a specificity of 97.84%, and an accuracy of 98.75% that outperforms other state-of-art methods.


Introduction
In the sophisticated healthcare group, medical imaging has become the very important device.
Since this holds various visual documents and reports of patients and refines data for numerous diseases like glaucoma, diabetic retinopathy, and macular degeneration, it became riskier [1].
Glaucoma is the neurodegenerative disease which encourages the optic nerve damage that results in visual loss. The glaucoma is the asymptotic as well as the chronic disease; vision loss is avoided by the early diagnosis. Regarding the social condition and population, glaucoma is the world's largest reason for the blindness only caused by the cataracts. It is reversible via the surgery, while the glaucoma leads to enduring blindness [2].
Glaucoma has become one of the world's most prevalent causes of blindness. It has been estimated that by 2020 there will be 80 million individuals with glaucoma. It does not reflect symptoms in its early stages and by damaging the optic nerve, progressively causes irreversible loss of vision [3]. Glaucoma is caused mainly due to the rise of intraocular pressure (IOP) inside the eye. This increased pressure often leads to improper balance of fluid in the eye and hence it injures the optical nerves and also affects the eye movement. Further, the majority of Glaucoma cases involve no prior symptoms or pain and so it is often called a 'silent thief of sight' [4] The early stage of glaucoma is asymptomatic, and about 50% of patients are reported to be unaware of the disease. Glaucoma progression without diagnosis and proper therapy gradually lead to irreversible loss of vision. For the proper management of the disease, early diagnosis and treatment of glaucoma are essential [5]. Glaucoma occurs mainly in people over 40 years old, but can also occur in young adults, children, or even children. There is an elevated incidence of glaucoma over the age of 40 years, a history of glaucoma in the family, impaired vision, diabetes, and eye injuries. The signs of glaucoma are vision loss, eye redness, nausea or vomiting, a hazy-looking eye (particularly in children), narrowness of the vision. The study was carried out with fundus images collected from the fundus camera [6].
Usually, some test procedures can be done by Optical Coherence Tomography (OCT) for suspected glaucoma diagnosis, such as tonometry. Tonometry is an intraocular pressure monitoring process. For a fast air pulse with an air-jet effect to calculate IOP [7], air-puff based contactless tonometers are used. Gonioscopy is the method to verify the opening and closing of angle, optical coherence, or fundus imaging to verify the visibility of the retina and the optic nerve. Glaucoma can be diagnosed with help of the retinal fundus image which is used to measure the thickness of the retinal nerve fiber layer (RNFL). Moreover, it is one of the noninvasive techniques mostly used by ophthalmologists, since it can be exploited easily to take images of healthy as well as non-healthy retinas [8].
Several studies have been carried out to investigate the correlation between the visual field test results and structural estimates that are produced by optical coherence tomography (OCT) scanners. Of these, RNFL thickness has been estimated to find the linear relationship of vision loss at an advanced stage of the disease. Moreover, previously the ocular disease can be detected and diagnosed with the help of fundus 2D-color images incorporated with Deep Learning Techniques. This includes retinal vessel segmentation, segmentation of the optic disc and optic cup, glaucoma recognition, and image registration [9]. In order to detect glaucoma, the retinal fundus image is used to calculate the thickness of the retinal nerve fiber layer (RFNL). It is the ratio of the size of the optic nerve (named as a disc) to the size of the excavation that occurs inside the optic nerve during the enlargement of eye pressure (named cup). This parameter is known as the Cup-to-Disc Ratio (CDR) [10]. Several studies have been focusing on automated identification of glaucoma based on color fundus images, but the CDR estimate is the most complex one.
To ease the estimation of CDR precisely this paper proposes a novel method known as a glaucoma detection system that performs the diagnosis of glaucoma by exploiting the prescribed characteristics and proposed a novel classifier for the efficient classification of the glaucoma infected images from the normal images. The main contributions of this paper are • The input fundus images are enhanced before subjecting to the process of segmentation phase.
• The segmentation is performed efficiently to extract the blood vessels and the optic disc image from the input images.
• The fractional gravitation search optimization (FGSO) algorithm based hybrid deep neural network is proposed to classify the glaucoma image from the extracted features.
The rest of the paper is arranged as follows: Section 2 discusses the related works of the various classification algorithms used for glaucoma detection. Section 3 provides details about the proposed methodology to detect glaucoma incorporated with the classifier. The experimental results and the comparative analysis with other state-of-art methods are discussed in Section 4.
The conclusion of the paper is stated in section 5.

Related works
A great effort has been made, and numerous studies have been presented by many researchers are described in Table 1. Rahouma KH et al. [11] proposed the gray level co-occurrence matrix (GLCM) and gray-level run-length matrix (GLRLM) for the extraction of glaucoma features. with the principal component analysis (PCA) method was proposed by Chan YM et al. [12] for combining the feature vectors. Optical coherence tomography angiogram (OCTA) images were used for the glaucoma detection. This method assists the clinicians to identify the glaucoma at the early stage. The OCTA images consist of four types of images such as Ocular sinister (OS) macular, OS disc, Ocular Dexter (OD) macular, and OD disc. In the pre-processing stage image resizing to 300x300 pixels and converted to grayscale, contrast limited adaptive histogram equalization (CLAHE) employed for the contrast enhancement. The problem of this approach was the tedious feature extraction process, clinically significant because of the usage of more number of images.
Li L et al. [13] proposed the automatic glaucoma detection for fundus mages to increase the reliability and accuracy. Large scale attention-based glaucoma (LAG) database achieved from the Chinese glaucoma study alliance (CGSA) and Beijing Tongren Hospital consists of 11,760 fundus images. The RGB channels of images are transformed into binary glaucoma labels for removing high redundancy in the fundus image. The automated glaucoma CNN (AG-CNN) was used to classify the normal image as well as glaucoma infected image. Glaucoma detection used image channels (ICs) as well as the discrete wavelet transform (DWT) for the fundus image was suggested by Kirar BS et al. [14]. Images were resized to the red channel, blue channel, green channel, and grayscale images, and CLAHE was used on ICs to eradicate the unnecessary lighting effect and improving image visibility. The least-square support vector machine (LS-SVM) employed as the classifier for the classification of glaucoma infected images as well as non-infected images. RIM-1 database consists of 505 images out of which 250 were the glaucoma infected images and 255 were the normal images. The main drawback of this approach was difficult to detect multiclass glaucoma.
The new more exact approach for the detection of computerized glaucoma recognition by quasi-bivariate variational mode decomposition (QB-VMD) was proposed by Agrawal DK et al. [15]. RIM-1 database consists of the MIAG database evaluated on 505 images, in those 255 healthy images and 250 glaucoma infected images. Noise removal, median filter, and CLAHE for the contrast enhancement were the techniques used in the pre-processing stage. LS-SVM classifier employed to categorize the infected and healthy images. The accuracy of the classifier was 85.94% and 86.13% for 3-fold and 10-fold cross-validation databases. The large database multistage glaucoma detection was not possible using this method. The method to detect the digital fundus image for examining the glaucoma image and normal images was proposed by Dey A et al. [16]. PCA was used for feature extraction and the SVM method for image classification. The database contains 100 digital eye fundus images. For the pre-processing techniques such as 2-dimensional Gaussian filter used for noise deduction and adaptive histogram equalization employed for contrast enhancement.
The glaucoma detection with wavelet based method for the real-time screening systems was proposed by Abdel-Hamid L [17]. The integration of statistical features and wavelet textural evaluated from the optic disc region was utilized to split the images to healthy or glaucoma images. There are two public databases were used: Glaucoma DB and HRF databases. In the preprocessing stage, red channel illumination was normalized, unsharp masking, and CLAHE was executed to maximize the contrast. K nearest neighbor (kNN) classifier was used for the classification of normal and infected images and the accuracy obtained was 96.7%. Sharma A et al. [18] proposed the automatic glaucoma diagnosis for the digital fundus image. Drishti [19] [27] proposed enhanced methods to separate disease infected and features. Gaussian filter, histogram equalization, noise reduction, ROI detection were the techniques used for the pre-processing.
SVM and ANN were the classifiers employed for the glaucoma infected images classification.
Automatic diagnosis allows the regular checks can be made for the patients to detect the glaucoma infection.

Proposed methodology
The proposed methodology of glaucoma detection is depicted in Fig 1. Here deep neural network is used as an efficient method to detect glaucoma infected and normal images. The four modules used in the proposed method include pre-processing phase, segmentation phase, feature extraction phase, and classification phase.

3.1.
Image pre-processing phase The pre-processing method considers the most important phase in the image processing that is established to assure all the images and is normalized with each other before the authentic examination. It is utilized to improve the data images by conquering unnecessary features like speckles, blind spots, noise, low contrast, irrelevant variations, etc. The pre-processing scheme improves the important aspects for further processing. It includes the schemes like image resizing, channel extraction, noise removal, image enhancement is employed [29] and the preprocessed image is shown in Fig 2. (i) (ii) (iii) (iv)

Image resizing
The input fundus and OCT images are obtained from the database and resized to 300 x 300 pixels to create all the original images with same resolution, dimensions, and similar scale for better analysis purposes.

Image extraction
The resized images are employed to extract the green channel, red channel, blue channel and grayscale images [30]. The fundus image consists of RGB color as shown in Fig 2. The Pseudocode color representation is utilized to distinguish all the information channels more accurately. Several operations cannot be established that consist of many colors, hence only one channel is extracted to optimize the result. Human eye is most responsive to green channel because the green channel carries more information for processing and to differentiate the healthy and the glaucoma infected images. Hence the Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to enhance the green channel image.

Median filter
Errors in the image processing occur due to the occurrence of various types of noises like salt and pepper, Gaussian, Gamma, uniform, Rayleigh, Erlang, exponential, etc in the input images.
Hence, median filters are employed to remove these noises from the input images for improved and precise results.

Contrast limited adaptive histogram equalization (CLAHE)
CLAHE method is employed in the entire information channels for the image quality enhancement, image visibility and also to remove the unnecessary lighting consequences. The green channel images are improved by utilizing the CLAHE scheme [31]. The input image quality shall be affected because of noises and artifacts that results in low contrast. The image contrast is improved by CLAHE and is employed to process the image quality and also improves the differences among the pixel of image intensities inside the nearby distance. The speed of identifying the diseases using the input OCTA and fundus images depends upon the image quality. Histogram equalization changes the information channel in the image due to the probability distribution and extends the distribution limit for improving the contrast and the visual effects. This equalization establishes the transformation utilities to generate the output image that has a uniform histogram. The CLAHE specifically targets the image entropy and obtains better equalization employing the maximum entropy. This equalization removes the common problem in the adaptive histogram equalization to overcome the reliable regions inside the image because the histograms of such regions are highly concentrated and normally dense.
CLAHE permits the user to set the pre-described value to control and restrict the amplification of the contrast, avoiding the problem of over amplifying the near reliable regions inside the image.

Image segmentation
The process next to the pre-processing is the segmentation. It is the procedure of separating the image into numerous segments based on the characteristics of the image pixels. As the glaucoma disease concerns the retinal optic nerve, the region of interest (ROI) for the detection of glaucoma that focuses on the optical disc and blood vessels [32]. The pre-processed image is given as the input to the segmentation section. The segmentation process split the image into two segments namely the blood vessel and optic disc. The segmented image IS is numerically expressed as, From Eqn. (1), the factor V I is used to denote the blood vessel segments and the factor O I is used to represent the optical disc segments in the image I . The process in the segmentation of the image I is described as follows.

Segmentation phase by blood vessels by thresholding method
In the eye, the blood vessels are used to carry the oxygenated blood. However, sudden rise in the interocular pressure (IOP) is able to modify the eyes blood vessels. The blood vessel segmentation is obtained from the pre-processed image enhances glaucoma detection. It is one of the most important steps because the important attributes are used for the glaucoma identification. In the segmentation process, the fundus image utilizes the green channel. And in the blood vessel segmentation, first the optic disc has to be eradicated and then the morphological opening is employed. The CLAHE output deducted for separating the vessels to obtain the final image. The noise from the image is removed and the image is transformed into black and white. From this, a new threshold value is evaluated to separate the blood vessels [33].
If all the pixels are less, then the threshold level is set to 0, and if all the pixels are greater, then the threshold level is set to 1. There are four regions involved in the blood vessel segmentation such as inferior (I), superior (S), nasal (N) and temporal (T) and revealed in Fig 3[8]. The region which has additional blood vessels for glaucoma infected person than the healthy person is the nasal region.

Segmentation phase by optic disc
The adaptive threshold-based scheme is used for optic disc segmentation and the glaucoma detection is by optic disc cup. Due to the principal orange setting of the fundus image retina, the disc segmentation is achieved by red channel. In the initial process of disc segmentation, the red channel provides better results to attach the preliminary threshold for disc localization within the retina. In some cases, the green channel provides enhanced performance than the red channel [34]. The threshold is established from the histogram equalization that is explained in the previous section and employed to the green channel. The image achieved is the binary image and provided to the morphological closing. The morphological dilation requirement occurs due to the optic nerve head (ONH) and wrapped by blood vessels. To eradicate the gaps, the morphological closing is made. The dimension of the structural element is equal to the preliminary blood vessel width on the selected optic nerve head for the closing process. The preliminary blood vessel is established at the optic disc boundary gap due to its highest value equivalent to the primary blood vessel width [35]. The preprocessed green channel generates the ONH region only.
The average amount of pixels is evaluated according to the subsequent parameter in [36] and considered as the disc.
In eqn (2) d denotes the diameter of the retina which is converted into pixels. This average pixel amount is obtained to evaluate the threshold for the segmentation. If the pixel between the two grayscale tones is greater than p A , then the biggest tone is chosen as the threshold. There are three conditions for the segmentation evaluation, they are the ratio of height versus block width which has the disc equal to the elliptical shape, pixel amount within the blob to be less than p A , and the location of the block. The segmentation procedure repeats iteratively and evaluates the threshold accuracy until the three conditions satisfied. The steps in the optical disc segmentation are as follows: First, the image is partitioned into various blocks. The bright pixel count is performed for every image block, and then establishes the bright pixel count and the total white pixels are represented by the bright pixels in the specified region. The image optical disc from the center point is described by fixing the circle. The optic disc segmented image is denoted as O I .

Feature extraction phase
After completing the segmentation process, the extraction of features is established. Several

Cup-to-Disc Ratio (CDR)
The cup to disc ratio is the factor efficiently utilized for glaucoma identification. The sudden enhancement in the intraocular pressure causes the cupping of the disc. For the usual disc, the value of CDR is less than 0.3 and the value of CDR is larger than 0.3 for glaucoma infected image.

Neuro-retinal rim (NRR)
It is described as the region among the optic cup edge and the optic disc edge. In temporal regions, the proportion of the area is absorbed by NRR and the NRR is engaged in inferior as well as the superior regions because the nasal regions become thicker than all the other regions.
The usage of AND operation among the optic disc and optic cup offers the area of NRR.

Retinal nerve fiber layer (RNFL)
It is created by the optic nerve fiber expansion which is the thickest layer near the optic disc. For the normal eye, the RNFL is frequently observed in the inferior temporal region pursued through the superior temporal regions, inferior nasal and temporal nasal.

INST features
This feature indicates the proportion of the blood vessels area present in the regions of superior and inferior to the regions of the temporal and nasal for the optic disc image segmentation. inferior for the optic disc image segmentation.

Mean
Mean is used to determine the average number of white pixels. The mean of the input fundus image is described as follows: Where ) , ( j i represents the pixel position in the image, N denotes the total amount of total amount of pixels in the image and X X N  = .

Variance
Variance separates the image contrast information and it depends upon the mean value. It is represented as follows: From Eqn. (6), V represents the variance, the total intensity levels in the image is indicated as M and the histogram probability of the image is indicated as ) (x P .

Textural features
The features that are extensively implemented in the analysis of the medical field are textural features that provide useful information regarding the particular objects characteristics or ROI inside the image. The gray-level co-occurrence matrix (GLCM) is the frequently employed scheme for the evaluations of texture feature. It explains the image texture by displaying how frequently particular pixel pairs occurred in the image hence obtain the information regarding the pixel arrangement in that particular image. Various textural features are implemented for the optic disc features arrangement such as energy, contrast, homogeneity, entropy, correlation, dissimilarity, cluster prominence, cluster shade, information correlation measure, and difference variance [37]. Difference variance and contrast are utilized to calculate the contrast of the image where the cluster head imitates the human perception. Moreover, the energy, homogeneity, correlation, entropy, and information correlation measures are employed to characterize the homogeneity of the image. The GLCM is described in g (x, y) whereas x and y describe the row and column inside the particular matrix. The textural features were evaluated as per the below equations.

Statistical features
There are various statistical features described in this statistical feature extraction like energy, mean, skewness, variance, and kurtosis. The skewness and mean describes the entire intensity of the identified optic disc area while kurtosis and variance explains the illumination discrepancy and the energy provides the information present in the optic disc field [17] [38]. The value of the mean provides the average gray level of every area and it is represented as The value of the variance evaluates the number of gray level fluctuations commencing from the mean and it is described as The term skewness evaluates the distribution asymmetry in the region of the sample mean.
The minus value represents the data spread more towards the left side rather than right side and vice versa and it is evaluated as Kurtosis is utilized to calculate the distribution tailedness relative to the normal distribution and it is described as follows:

Image classification
Once the feature extraction process is completed, the selected features are fed to the classification phase input [39].

Deep belief networks
The architecture of the feed-forward neural network is the significant element of deep belief network and the network consists of several hidden layers. In that input layer is the visible unit, N represents the number of layers in the hidden layers along with the output layers. The weights among the DBN factors are represented as wi and the biases bi of layer I and between the unit layers are i-1.

➢ Pre-training stage
The parameter initialization is the major problems in the deep neural network training. The reduced local minimum for the fault identification is established by the random initialization.
The Restricted Boltzmann Machine (RBM) for the training series issues are employed in [40]. The energy function is signified in below equation.
Hence, the bias vectors of the visible and hidden layers are represented as x, y, and the weight is represented as w by the conditional distributions are represented as follows: The logistic function turns out to be ) tend in the interval of (0, 1).
The RBM training processes are described in the following section. Initially, the training process is unsupervised and disregards the class label. RBM on the other hand express the socialized vector and it collides with the identical input data. Lastly, the deliberation in the RBM is to circulate the hidden unit's criteria. Consequently, this process is persistent and the factors attained are scheduled as pursues: DNN pre-training schemes • Initially, initialization of training vector by the visible unit b.
• Next, update the hidden units in parallel as per equations (38) and (42).
• Also, update the visible units in parallel as per Eqn. (24).
• Depends upon the examined reconstruction of a similar equation in the above utilized in Step 2, the hidden units are again updated in parallel.
• Update the weights by utilising When the process of training is by RBM, the multilayer is produced and at the top another RBM shall be heaped. The stacked RBM among the units in the previously RBM trained layers and the initialized vector is established through the present weights and biases. The novel RBM input is by the previously trained layers and this process is continued until the preferred terminating criteria is satisfied. In this, there are 253 layers of input, three hidden layers, and only one layer of output. The stage of fine-tuning is through the achieved weights of deep neural networks.

Fine-tuning process
The value of the weight is adjusted and the error is minimized by the fractional gravitational search optimization (FGSO) algorithm. Initially, the values of the weights are initialized randomly ij w . The solution demonstration is a significant method for resolving the issues in the whole optimization algorithm. Next, the fitness function for every solution is computed. The classification accuracy is regarded as the fitness function and the value depends on Eqn. (31).
Hence, update the fitness computation and FGSO solutions. In the next step, the procedure for the update is provided. The layer of the output is offered at the deep neural network top which divides the images. The dataset for the training T d is employed for weight optimization in the training process. First, the fresh features are provided to the DNN after that the weights are provided. Eventually, the database T d is tested to classify the image that depends upon the best weight (w). Mostly, the decreased aspects are provided to the DNN, whereas the weight is randomly modified. Through the selection of the optimal weights, the FGSO algorithm is employed by the proposed approach and the procedures are described below.

➢ Fractional gravitational search optimization (FGSO) algorithm
FGSO algorithm is the integration of the GSA [41] and the fractional theory [42]. This search algorithm is mainly depends on the gravity law. The problem of convergence is neglected in the GSA algorithm by enhancing the search space through the fractional calculus theory adaptation.
The proposed fractional gravitational search algorithm is explained in the following section.
In the FGSO algorithm, the data entities are regarded as the agents and the masses are used to calculate the performance. The agent location represents the solution for the issue. The steps for the FGSO are explained below.
Step 1: Initialization stage Initially, the location of the objects and the agents are generated randomly. The location of the qth objects is described for J objects.
.... .... , Step 2: Fitness Evaluation stage In the search space the fitness function for every particle is evaluated by using Eqn. (31). The fitness employed for the two agent's location is described in this stage. At first the distance between the individual object from the centre point is computed and then acquire the optimal centre point is computed. The fitness function for the FGSO algorithm is formulated as below. in [43].
Step 3: Inertial mass computation When the inertial mass is high the particle attraction is also high. The optimal solution is achieved from the higher particle mass. The agent's inertial mass is computed using the subsequent equation.
From the Eqn. (32), q x value is described as The value of the best as well as the worst function is formulated as Step 4: Gravitational force stage The gravitational forces among the entity q and g is computed by the below Eqn. (37) ( ) From Eqn. (37),  represents the small constant, ag X indicates the dynamic mass of the gravitation linked with the entity g , qh X represents the reactive mass of the gravitation linked with the agent q , the gravitational constant for the particular time is represented as ) (t H .
Step 5: Velocity and acceleration computation The agent velocity is computed using the fractional present velocity inserted to its acceleration and it is formulated as From Eqn. (38), l q u represents the agent acceleration at the particular time t. The agent acceleration is computed by the update depends upon the inertial mass and this mass gives the best solution. The acceleration is computed by the below formula.
Step 6: Fractional location updation The agent location at the subsequent repetition is calculated by the fractional theory adaptation in the GSA algorithm. The agent location is updating by using Eqn. (40) If the fitness value of the agent of the next repetition and the earlier iterations are identical, the fractional calculus took over its region.
Then the fraction calculus is represented as  L .  is the constant value and normalized to the actual number among the range 0 and 1.
The left hand side of the equation (41) is rearranged as Substituting the Eqn. (42) in (41) ) Eqn. (43) represents the updated fractional agent location.
Step 7: Termination stage The location updation is repeated until the optimal fitness function is achieved. The value of the global fitness function is the optimal fitness function and the location respective to the consequent agents is regarded as the optimal solution to the issue.

Experimental results and discussions
This section describes the proposed early glaucoma detection using the segmentation of retinal blood vessel and implemented in the MATLAB 2014 platform with the system configurations by the i5 processor in 4GB RAM. This illustration computes the factors like pressure drop, stress, blood velocity, and strain for the arteries as well as vascular networks. Additionally, the proposed approach to the classification of images from the OCT images and fundus images is evaluated using the performance measures such as accuracy, specificity, positive predictive value, and sensitivity.

Database
The fundus images have been extensively utilized for the primary assessment of ophthalmic abnormalities. The optic coherence tomography (OCT) is a moderate imaging method that establishes the quantitative examination of retinal layers [44]. These OCT images employed to examine the morphological alterations in the retinal layers that offer the specified image of the ocular disease. The database consists of glaucoma and controlled fundus images as well as the OCT images. The database obtained from 27 subjects scanned on the system of TOPCONS 3D OCT. It contains eye data for 24 subjects and one eye data for 3 subjects. Thus the observed data consists of 50 Fundus and OCT images that contain 22 controlled and 28 glaucoma infected patients..

Performance measures
Specificity, Sensitivity, and accuracy are the metrics used for the measurement of success of the classifiers in classifying images to normal or glaucoma affected images. To make clear the metrics, the following metrics are required.
True positive p t : The glaucoma image numbers that are exactly categorized.
True negative n t : Usual images numbers that are exactly categorized.
False positive p f : Usual images that are wrongly classified as glaucoma infected images.
False negative n f : Number of glaucoma images that are wrongly classified as usual.
Specificity: It is defined as the measure that evaluates the probability of the consequences which are true negative that indicates the non-vessel numbers identified accurately.

Experimental evaluations and discussions:
This section describes the results and discussion of the glaucoma detection. The proposed FGSA-HDNN approach described in this article separates several features from the pre-processing stage, blood vessel and optic disc segmentation. It represents the output image obtained from the proposed approach for the glaucoma recognition system by the healthy image processing. It is observed that the scheme has distinguished the normal as well as glaucoma affected image. The differentiation in the blood vessel segmentation and optic disc segmentation for the normal as well as the glaucoma affected images are shown in Figure 4. disc for the images sets as described in the Table 2. The performance of the proposed approach is computed by comparing the segmentation results with the resultant truth images.
The proposed approach for the optic disc segmentation gives 99.18% accuracy, 70.85% sensitivity and 99.24% specificity. Table 3  (2019) obtains an average accuracy of 95.34% for the optical disc segmentation. The reason for the proposed approach for the optic disc segmentation is that it identifies and segments the inner and outer boundary of the optic disc region more perfectly when compared to other traditional methods for the optic disc segmentation. The performance measures like specificity, sensitivity and accuracy for the various state-ofart methods are compared with our proposed hybridized deep neural network with a fractional gravitational search algorithm (HDNN-FGSA) for glaucoma detection. The performance comparison of the proposed method with the other state-of-art methods is demonstrated in Table   4.

Conclusion
The vision illness strength due to glaucoma is decreased when it is detected at the early stages. In this paper, the glaucoma detection scheme is based on the proposed novel classifier that detects the glaucoma image more accurately. The segmentation process is done by the blood vessels and the optic disc in the input fundus images and the output of the segmentation are used for the