Classification of Glaucoma Based on Elephant-Herding Optimization Algorithm and Deep Belief Network

Abstract

This study proposes a novel glaucoma identification system from fundus images through the deep belief network (DBN) optimized by the elephant-herding optimization (EHO) algorithm. Initially, the input image undergoes the preprocessing steps of noise removal and enhancement processes, followed by optical disc (OD) and optical cup (OC) segmentation and extraction of structural, intensity, and textural features. Most discriminative features are then selected using the ReliefF algorithm and passed to the DBN for classification into glaucomatous or normal. To enhance the classification rate of the DBN, the DBN parameters are fine-tuned by the EHO algorithm. The model has experimented on public and private datasets with 7280 images, which attained a maximum classification rate of 99.4%, 100% specificity, and 99.89% sensitivity. The 10-fold cross validation reduced the misclassification and attained 98.5% accuracy. Investigations proved the efficacy of the proposed method in avoiding bias, dataset variability, and reducing false positives compared to similar works of glaucoma classification. The proposed system can be tested on diverse datasets, aiding in the improved glaucoma diagnosis.

1. Introduction

Glaucoma is a type of ocular neuropathy which threatens vision if left untreated [1]. It is a gradual condition that develops as the ocular pressure rises [2]. Nearly 80 million people are affected by glaucoma globally [3]. Optic nerves are impaired gradually due to high ocular pressure [4,5]. Approximately 75–80% of the glaucoma cases are identified only at the developed stage, which cannot be cured. One cannot visualize any known symptoms of glaucoma except for narrowing of vision at a later stage [6]. The progression is slow and painful at certain levels [6]. It requires lifelong treatment, and it is impossible to reinstate vision loss. Hence, early detection of glaucoma and treatment stands as the best means of prevention. Concerning the fundamental issue, there is a critical need to build a system that can function well without the need for excessive equipment, qualified medical practitioners, or time. Computer-assisted techniques could help detect the disease at its early stages using advanced machine-learning and deep-learning methods. Trained deep-learning models could take advantage of minor changes, such as retinal layer thinning, that human specialists cannot notice.

Nevertheless, the optic nerve damage precipitated slowly develops, and as symptoms stem from it, the disease advances significantly [7]. Nevertheless, the most up-to-date technology can probably hinder glaucoma development in patients [8,9]. Figure 1 shows sample images of normal and glaucomatous eyes.

To scrutinize various glaucoma traits, ophthalmologists utilize confocal scanning laser ophthalmoscopy (CSLO) [10], Heidelberg retina tomography (HRT), optical coherence tomography (OCT), along with fundus images [11]. For instance, numerous retinal features, optic nerves head (ONH), peripapillary atrophy, and retinal nerve fiber layer, are perceived for glaucoma diagnosis [12]. Assessment of increased IOP, abnormal visual field (V.F.), damaged ONH, etc. are usually investigated for glaucoma [13,14,15]. The OD is split into three disparate areas: the cup (central region), the neuroretinal rim, and parapapillary atrophy [16,17]. The white cup-like structure located in the disc center is the OC. The ratio of OC size to OD size is normally an important measure analyzed in glaucoma diagnosis, denoted as the cup-disc ratio (CDR). The main contributions of this paper can be summarized as follow:

• Developing an optimized model employing a deep belief network classifier (DBN);
• Employing modified Wiener filter (MWF), circular Hough transform (CHT), and Otsu’s thresholding for OD and OC segmentation, respectively;
• Generating a distinct hybrid feature set to aid in diagnosis;
• Selecting relevant features through the ReliefF algorithm based on predictive importance weights;
• Fine-tuning DBN by elephant-herding optimization algorithm (EHO);
• Investigating the model’s robustness to noise such as Gaussian and salt-pepper;
• Analyzing the isolated and combined feature set contribution in glaucoma identification.

The paper is structured as follows: Section 2 outlines various works related to the proposed method. Section 3 presents the adopted architecture of the method proposed. Experimental outcomes along with dataset preparation are explored in Section 4. The conclusion is elucidated in Section 5.

2. Related Works

A brief literature review based on feature extraction and neural network classifier for glaucoma detection is elucidated in this section. Raja et al. [18] described a statistical feature extraction method based on the hyper analytic wavelet transformation (HWT) in which statistical characteristics were extracted and passed to a support vector machine (SVM). The particle swarm algorithm is used to adjust the HWT and SVM-RB simultaneously to get the optimum fit. Issac, A. et al. [19] and Koh J.E. et al. [20] presented a similar feature extraction strategy for glaucoma classification. Haralick’s features-centered categorization of glaucoma using back propagation neural networks (BPNN) was suggested by Samanta et al. [21]. The results of the experiment revealed good accuracy, sensitivity, and specificity.

Acharya, U. et al. [22] proposed an effective technique that included preprocessing, picture convolving with filter banks, and the chosen features fed into the KNN classifier. Jun et al. [23] demonstrated a super-pixel-based OD and OC segmentation model for diagnosing glaucoma. The unreliable results exhibited scope for improvement. Gift and Nirmal [24] suggested gray wolf optimized NN produce enhanced accuracy for glaucoma detection through a sequence of steps. Preprocessing, image normalization and feature extraction were done, and the features were given to the GWO-NN classifier. Anushikha et al. [25] presented an automated diagnostic system utilizing wavelet features from the segmented OD, which was extorted for analysis in addition to classification. Experimental outcomes signified an accuracy of 0.947. Ajesh et al. [26] reported a new multi-feature extraction approach for glaucoma identification and classification by integrating discrete wavelet transform (DWT) and ML algorithms that produced an accuracy of 95% in glaucoma identification.

For fine-imaging analysis, DWT was computationally intensive. Diaz-Pinto et al. [27] provided an automatic technique using retinal structural features and Luv color space for OD and OC segmentation, obtaining 81% specificity and 87% sensitivity. Studies have reported numerous methodologies for recognizing glaucoma through CAD. They use either conventional machine learning, deep learning, or both. Several works related to the diagnosis of glaucoma using CNN have been demonstrated in the literature [28,29,30,31,32]. The CNN employed was constructed from scratch, and different datasets were used to investigate the models. Data augmentation via rotation, random flip, image translation, etc. was performed to increase the dataset size artificially. OCT images were also used to segment the retinal vasculature apart from the fundus images. The deep networks were employed to extract and learn the layer properties of the retina using a pretrained backbone network and reinforcement learning. A multiscale feature generated could then be used on modules, such as the encoder–decoder, to retrieve retinal information and capture finer retinal boundaries. Many optimization algorithms are also used to solve the various optimization problems. A blended approach is adopted in our method, where the features extracted are selected and passed to an optimized deep network for classification.

3. Proposed Method

This paper reports using the DBN classifier, a robust, novel, and efficient glaucoma detection method on retinal fundus images (RFI). Input is acquired from fundus databases and preprocessed for noise removal using MWF. Utilizing CHT, the OD is then segmented from the noise-removed image. The OC is cropped from the OD image using Otsu’s thresholding algorithm. Structural and functional features are extorted and fed to the DBN classifier. The EHO algorithm optimizes DBN’s parameters to obtain a minimum cost function and the best solution for healthy and glaucomatous classification. Figure 2 illustrates the schematic diagram of the system.

3.1. Preprocessing

The fundus image collected must be preprocessed to highlight the vessel and other morphological features. Given the benefits of grayscale images, RBG is converted to grayscale first [33], then filtered to remove background noise [34]. A modified Wiener filter is employed to prevent impulse noise while preserving the edges [35]. The noise reduction aims to protect crucial structural content for disease detection, besides determining noisy pixels [36].

3.2. OD and OC Segmentation

OD segmentation is done on the image after noise removal. First, the OD location is identified. Subsequently, edges are calculated through the Canny edge filter. As the OD in the retina is a circular object, the circle detection Hough transform (CHT) is applied to identify the area. The boundary positions along with the region of the OD are obtained. The ROI termed OD is cropped from the original image for further OC segmentation. The OC is segmented from pre-segmented OD using Otsu’s thresholding [37].

3.3. Feature Extraction

Structural, textural, and intensity-based features contributing to the disease diagnosis are extorted in this phase [38].

Table 1. List of features extracted.

Structural Features	Cup to Disc Ratio (CDR), Neuro Retinal Rim (NRR), Cup Shape
Textural features	Wavelet-based features, gray level co-occurrence matrix (GLCM) features—energy, correlation, homogeneity, contrast, and entropy gray-level run length—low gray level run emphasis, gray level non-uniformity, segmentation-based fractal texture analysis (SFTA)
Intensity features	Brightness, color moments, super pixels, enhanced local binary pattern (ELBP), speeded-up robust feature (SURF), pyramid histogram of oriented gradients (PHOG), local energy-based shape histogram (LESH)

lists the set of features generated through this process.

Using the procedure described above, the 111 features extracted are then selected using the relief F algorithm before being fed to DBN for classification.

3.4. Feature Selection

This stage is critical because rarely discernible characteristics are deleted, putting the modeling process under more computational strain. In this work, the ReliefF algorithm is preferred for dimensionality reduction due to its promising results, as explained in [38]. ReliefF ranks features according to their weight participation, with the most active features being placed first. As illustrated in Figure 3, other features contribute far less to the last features. As a result, we can choose the 15 most potent features based on their weight and exclude characteristics that add to the model’s computational cost.

ReliefF chooses T, some instances at random, but subsequently, k searches for the closest same-class neighbors, which are referred to as the nearest hit values of H. The k-nearest neighbors are the one–one scores among the multiple classes, also known as the M(T) nearest misses. The number of nearest neighbors is set at three in our study. The 111 features are reduced to 15 optimal features based on their participation at the top of the weighted list for the highest accuracy using the procedure described above. The remaining ones are ignored because there is not much difference in output, which increases the computing weight of a model. Figure 4 depicts cumulative accuracy vs. a number of features, and 15 contribute more than the total variations.

3.5. Classification

3.5.1. Deep Belief Networks (DBN)

The DBN class of neural network (NN) can be considered a generative model that uses a set of Boltzmann machines as basic building elements [39]. Each layer of the DBNs has a restricted Boltzmann machine (RBM). DBN extracts H.L. features from the data slated for training to improve the between-classes separation power. The training is performed on all the layers through supervised mode, and the backward propagation mode modifies the weight in the network to reduce over-fitting. This work develops a DBN model trained using greedy layer-wise learning [40] by stacking up RBMs, as shown in Figure 5.

RBM concentrates on a particular layer during its learning procedure and ignores others. We assume we have a DBN with L levels, with Wi being the RBM’s weight matrix at layer i. In addition, the hidden units at the ith layer form the layer (i + 1) input unit. In the model proposed, the set of weight matrices is assigned as W = {W1, W2, W3} and the set of hidden layers as h = {h1, h2, h3}. The weight matrix between layers ith and (i + 1), which is denoted by Wi, while the jth hidden layer is denoted by hj. The following energy function was used to compute the combined distribution of the hidden and visible layers:

$$p\left(v,h\right)=\frac{e^{-E\left(v,h\right)}}{{{\displaystyle \sum }}_{v,h}{e}^{-E\left(v,h\right)}}$$

(1)

where $E\left(v,h\right)$ denotes RBM’s energy function,

$$E\left(v,h\right)=-{\displaystyle \sum }_{i=1}{a}_i{v}_i-{\displaystyle \sum }_{j=1}{b}_j{h}_j-{\displaystyle \sum }_{i,j}{v}_i{h}_j{W}_{ij}$$

(2)

$W_{ij}$ denotes the weight between the visible and the hidden layer, $a_i$ and $b_j$ describe the visible and hidden layer coefficients. This study uses the stochastic gradient descent (SGD) method following log-likelihood (L.L.) to accomplish optimal training. This is accomplished by optimizing the RBM’s parameters a, b, and wij. The derivatives of the $\log p\left(v,h\right)$ w.r.t $W_{ij}$, $a_i$ and $b_j$ must be computed to update the weights and biases. The resulting equations are

$$W^{t+1}=W^t+\eta \left(p\left(h\mid v\right)v^T-p\left(h\mid v\right)v^T\right)-\lambda W^t+\alpha \mathsf{\Delta }{W}^{t-1}$$

(3)

$$a^{t+1}=a^t+\eta \left(v-\tilde{v}\right)+\alpha \mathsf{\Delta }{a}^{t-1}$$

(4)

$$b^{t+1}=b^t+\eta \left(p\left(h\mid v\right)-p\left(\tilde{h}\mid \tilde{v}\right)\right)+\alpha \mathsf{\Delta }{b}^{t-1}$$

(5)

where $p\left(h_j=1\mid v\right)=\sigma \left({{\displaystyle \sum }}_{i=1}^m{w}_{ij}{v}_i+b_j\right),p\left(v_i=1\mid h\right)=\sigma ({{\displaystyle \sum }}_{j=1}^n{w}_{ij}{h}_j+a_i)$, and σ(·) represents the logistic sigmoid function. $\tilde{v}$ and $\tilde{h}$ denote the reconstructed v and h, respectively. N is the number of hidden nodes, η, the learning ratio, α, the momentum weight, and λ, the weight decay. The weight matrix and accompanying bias vectors of the visible and hidden nodes are learned using contrastive divergence (CD) and persistent contrastive divergence (PCD). This optimization process uses the BP with the conventional gradient ascent algorithm to tune the weight matrices to optimal values. The optimization algorithm considers the outcome of an extra layer built over the DBN after its previous greedy training to minimize some error metrics. Softmax, or logistic units, are frequently used in this layer.

3.5.2. Elephant Herd Optimization (EHO) Algorithm

The EHO algorithm was introduced by Wang et al. in 2015 [41]. Elephants behave socially and encompass a complex structure of calves and females. An elephant group comprises numerous clans with a matriarch as the leader and her calves or other related females. A female forms a clan. EHO concerns the succeeding assumptions.

• The elephant group is classified into clans, and each such clan comprises specific elephants.
• A specific number of male elephants (ME) depart their clan to live independently.
• Each clan has a leader termed the matriarch.

The matriarch group keeps the best solution in the elephant herd. The entire elephant population is divided into $j$ clans. Matriarch ci influences the new position of each elephant ci. The elephant $j$ in clan ci can be calculated using

$$x_{new,c_{i,j}}=x_{c_{ij}}+a\times \left(x_{best,c_i}-x_{c_{i,j}}\right)\times r$$

(6)

where $x_{new,c_{i,j}}$ indicated the new position and $x_{c_{ij}}$ denotes the old position for elephant j in the clan ci. $x_{best,c_i}$ represents matriarch $c_i$, which denotes the best elephant. a ∈ [0,1] shows a scaling factor, $r\in$ [0,1]. The best elephant is computed for each clan using

$$x_{new,ci,j}=\beta \times x_{\mathrm{center},},i$$

(7)

Here, $\beta \in \left[0,1\right]$ indicates the second parameter that guides the impact of the $x_{center,ci,d}$ delineated in

$$x_{center,ci,d}=\frac{1}{n_{ci}}\times {\displaystyle \sum }_{j=1}^{n_{ci}}{x}_{ci,j,d}$$

(8)

where $1\le d\le D$, and nci represent the number of elephants in clan $x_{ci,j,d}$ is the d_th dimension of individual elephant $x_{ci,j,d}$ the center of clan ci ($x_{center,ci,d}$) can be updated (Equation (8)). The separating process could be modelled as a separation operator when tackling optimization issues. In each clan, the worst valued elephants are moved to the next position indicated by

$$x_{worst,d}=x_{min}+\left(x_{max}-x_{min}+1\right)\times rand$$

(9)

Here the lower and upper bands of the search space are indicated by $x_{min}$ and $x_{max}$, respectively. $rand\in$ [0,1] signifies the random value picked from the uniform distribution.

The EHO algorithm was examined for various benchmark set functions and in medical diagnosis [42,43,44,45,46], showing better results. This study employs the EHO algorithm for DBN parameter optimization. The output of the DBN model is grounded in weights and the biases of preceding layers in the network. EHO does not employ the previous individuals in the later updating process as other optimization algorithms. EHO is a swarm-inspired algorithm that deals with global optimization missions characterized by clan updating and searching operations. EHO does not resort to relaxation and is less vulnerable to noise. They perform better in constrained, optimized environments. High convergence rate and low localization errors with less execution time are the important characteristics of EHO. The algorithm can tackle non-convex ML problems directly.

3.5.3. Fine Tuning of DBM

The learning rate, hidden units, momentum weight, and weight decay are the four basic parameters set up in most RBMs. The use of traditional methods for computing the error function is an NP-hard problem due to its complexity and differentiation. Meta-heuristics have been employed to solve this issue. The EHO algorithm is used to optimize the DBN training by fine tuning the parameters in this work. Here, the parameters set are

$$n \in \left[5,100\right], \eta \in \left[0.1,0.9\right], \lambda \in \left[0.1,0.9\right] and \alpha \in \left[0.00001,0.01\right]$$

A fitness function must be designed to steer the searching process to attain the best answers to meet the objectives. Mean squared error (MSE) is adopted as the fitness function. It measures the error between the output and the desired value and is given by

$$MSE=\frac{1}{T}{\displaystyle \sum }_{j=1}^N{\displaystyle \sum }_{i=1}^T{(D_j\left(i\right)-Y_j\left(i\right))}^2$$

(10)

where T—data number, N—number of the output layers.

The EHO looks for a collection of DBN parameters that minimizes MSE. $D_j\left(i\right)$ denotes the value from the jth unit in the DBN’s output layer at the time ‘t’, $Y_j\left(i\right)$ represents the jth factor of the desired value. The process is repeated until the halting criteria is met.

The optimization steps of EHO are as follows:

• Set the EHO parameters and initialize the population.
• Evaluate the individual fitness value (RMSE) of the DBN, as per the learning rate and the number of batch learning. Identify the optimal individual.
• Check if the termination condition is reached; if so, end the iteration and output the result; or else, go to the next step.
• Update each individual position. Reinitialize the individuals beyond the lower and upper limits.
• Start a new iteration by updating the optimal individual.

4. Results and Discussion

Evaluation of the proposed model’s performance is presented in this section. The experiment is executed in MATLAB with the following specifications: the Intel Core i7 Processor, Windows 10, 3.20 GHz CPU speed, and 4GB RAM.

4.1. Dataset Preparation

This work uses DRISHTI–GS1, ACRIMA, ORIGA-Light, and LAG datasets for evaluation. The images were captured utilizing a Canon CR-1 fundus camera at 2336 × 3504 resolution with a 45° FOV and a disparate acquisition setting. The dataset used in the methodology is from public and private datasets annotated by an ophthalmologist who has over 15 years of experience in the field. The list of databases is depicted in

Table 2. Dataset labeling.

Database/Images	Normal	Glaucoma	Total	Type
DRISHIT-GSI [44]	12	89	101	Public
ACRIMA [29,45]	309	396	705	Public
OTIHS-lihjy [46]	482	368	650	Public
LAG [47]	3432	2392	5824	Private
Total	4235	3045	7280

. A total of 7280 images obtained from various public and private databases are used for investigating the proposed system’s performance after eliminating a few irrelevant images. Images are trained and tested in the ratio of 70:30, respectively.

Images that are preprocessed are subjected to OD and OC segmentation. Sample images and results of the segmentation are given below in Figure 6.

4.2. Performance Analysis

Experiments are done to show the efficacy of the DBN–EHO on different fundus datasets. Initially, the investigations are performed to show the efficacy of EHO in optimizing the DBN. EHO is compared with meta-heuristic algorithms, such as artificial bee colony (ABC), firefly algorithm (FA), harmony search (HS), cuckoo search (CS), particle swarm optimization algorithms (PSO), and differential evolution algorithms (DE). A hold-out technique with 20 training and test sets generated at random, subjected to 10 iterations for each RBM learning procedure, and a mini-batch of 20 are performed for consistent comparisons with other works. Five agents of over 50 iterations are employed to achieve convergence using all strategies in tests. Control parameters for all the algorithms are outlined in

Table 3. Control parameters of all algorithms.

Algorithm	Parameters
ABC	N = 30, MCN = 100, limit = 20
HS	$\mathrm{HMCR}=0.7,\mathrm{PAR}=0.7,\eta =1$
FA	$\gamma =1,{\beta }_0=1,\alpha =0.2,MCN=100$
CS	$\alpha =0.1,$ pa = 0.25
PSO	Wmax = 0.9, Wmin = 0.2, C1 = 2, C2 = 2
DE	F = 0.8, C = 0.5
EHO	$\mathrm{nClan}=5,\alpha =0.25,\beta =0.05$

.

In

Table 4. Average MSE over the LAG dataset.

Algorithm	Layer-1		Layer-2		Layer-3
	CD	PCD	CD	PCD	CD	PCD
ABC	0.0891	0.8940	0.0881	0.0884	0.0880	0.0878
HS	0.1259	0.1345	0.1256	0.1169	0.1158	0.1156
FA	0.0864	0.0864	0.864	0.0860	0.0864	0.0862
CS	0.1146	0.1146	0.1176	0.1175	0.1164	0.1162
PSO	0.1086	0.1086	0.0988	0.0992	0.1045	0.1046
DE	0.1250	0.1254	0.1254	0.1254	0.1158	0.1156
EHO	0.0756	0.0756	0.0778	0.0778	0.0776	0.774

, the MSE values of each algorithm on the original LAG dataset considering DBN. A Wilcoxon signed-rank test with a 0.05 significance level is utilized to analyze the method statistically. It is observed from Table 4 that EHO outperformed all the other algorithms in terms of the lowest MSE employing fewer layers. ABC and FA algorithms have also performed well next to EHO.

Taking the performance of EHO, the performance of the proposed work is assessed by finding ‘true positive’ (TP), ‘true negative’ (TN), ‘false positive’ (FP), and ‘false negative’ (FN) values. TP denotes the instances wherein glaucoma is detected correctly. TN shows the condition wherein a person with no glaucoma is classified correctly. FP indicates the number of negative instances recognized as positive. Positive instances recognized as negative are indicated by FN. Precision, accuracy, F-score, specificity, recall, sensitivity, and MCC (Mathew’s correlation coefficient) are used in this work, as in

Table 5. Performance Metrics.

Parameters	Expression
Sensitivity (%)	$\frac{TP}{TP+FN}\times 100$
Specificity (%)	$\frac{TN}{TN+FP }\times$ 100
Accuracy (%)	$\frac{TP+TN}{TP+FN+TN+FP}\times$ 100
Precision (%)	$\frac{TP}{TP+FP}\times$ 100
Recall (%)	$\frac{TP}{TP+FN}\times$ 100
F-score (%)	2 × $\frac{\left(Precision\right)\left(Recall\right)}{Precision+Recall}\times$ 100
Mathew’s correlation coefficient (MCC) (%)	$\frac{\left(TP\times TN\right)-\left(FP\times FN\right)}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}\times$ 100

.

The performance of the proposed work on individual datasets is provided in

Table 6. Performance of the proposed method employing DBN and EHO.

Dataset	Acc (%)	Sens (%)	Spec (%)	Prec (%)	Recall (%)	F-Score (%)	MCC
Drishti-GS1	96.95	98.56	97.44	97.69	96.86	97.68	0.749
ACRIMA	99.34	97.1	98.2	88.92	95.3	93.5	0.772
ORIGA	98.51	94.73	98.7	98.55	97.92	95.32	0.784
LAG	99.31	99.89	100	96.73	94.56	95.64	0.789

. It is inferred that a maximum accuracy of 99.34% on the ACRIMA dataset is attained, followed by 99.31% on the LAG dataset. The classification rate is between 96.95% and 99.34%, ensuring that the DBN–EHO performed better in all the datasets. This indicates that the images in the set are captured under different illuminations. Specificity of 100% on the LAG dataset shows that the model can reduce false positive rates.

The impact of isolated and combined feature sets on the diagnosis of glaucoma in the LAG dataset is depicted in

Table 7. Impact of features in detecting glaucoma from LAG dataset.

Features	Accuracy (%)	Sensitivity (%)	Specificity (%)
Structural (SF)	94.87	95.32	93.52
Intensity (IF)	95.98	89.23	92.41
Textural (TF)	96.21	97.28	99.33
SF + IF	95.86	90.74	95.21
SF + TF	96.78	94.23	97.56
IF + TF	90.88	95.79	94.35
Selected features	99.31	99.89	100

. The extraction of isolated features yielded an enhanced result, increasing the algorithm’s efficiency. When the features are combined, an accuracy of 99.3% is obtained. When the features are combined, better accuracy of 99.34% is obtained. The feature contribution indicates that the DBN optimization also contributes to appreciable performance as the accuracy ranges from 95% to 99% on different kinds of features.

The feature contribution indicates that the DBN optimization also contributes to appreciable performance as the accuracy ranges from 95% to 99% on different kinds of features. A 10-fold cross validation done to reduce bias during testing enhances the algorithm’s robustness and reduces the classifier’s misclassification rate. From

Table 8. 10-fold cross validation results.

Dataset	Accuracy (%)
Drishti-GS1	97.1
ACRIMA	98.5
ORIGA	96.2
LAG	97.8

, it is seen that the cross-validation accuracy across the datasets is appreciably high, ensuring that the model is free from bias. The model applies to a wide range of datasets, compensating for any imbalance. Figure 7 highlights the proposed work against various performance metrics.

A 10-fold CV is carried out to reduce bias during testing, and the results are shown in Table 8.

From Table 8, it is seen that the cross-validation accuracy across the datasets is appreciably high, ensuring that the model is free from bias. The model applies to a wide range of datasets, compensating for any imbalance.

Table 9. Performance analysis with dissimilar classifiers.

Dataset	Classifier	Accuracy (%)	Sensitivity (%)	Specificity (%)
Drishti-GS	KNN	95.34	90.47	93.08
	RF	94.50	91.34	92.33
	SVM	95.86	96.87	96.87
	DBN	96.23	97.56	96.62
	DBN–EHO	96.95	98.56	97.44
ACRIMA	KNN	95.66	90.86	93.78
	RF	94.32	91.24	90.84
	SVM	97.06	96.64	96.12
	DBN	97.26	98.16	97.06
	DBN–EHO	99.34	97.1	98.2
ORIGA-Light	KNN	94.22	96.86	97.08
	RF	91.34	88.56	89.75
	SVM	94.88	95.56	96.69
	DBN	96.06	97.65	97.81
	DBN–EHO	98.51	94.73	98.7
LAG	KNN	94.24	95.56	95.85
	RF	92.78	90.89	91.48
	SVM	95.60	95.68	96.45
	DBN	97.54	95.67	97.43
	DBN–EHO	99.31	100	99.89

reports the outcomes of the classifier when compared with similar conventional ML classifiers, such as K-nearest neighbor (KNN), random forest (RF), support vector machine (SVM), and DBN without optimization. The DBN attains a more appreciable performance than the conventional ML classifiers. The DBN, when optimized for weight by EHO, still achieves better performance across all the datasets.

From Table 9, it is seen that the DBN classifier attains a more appreciable performance than the conventional ML classifiers. The DBN, when optimized for weight by EHO, still achieves better performance across all the datasets. Furthermore, to assess the robustness of the model, salt-pepper and Gaussian noise are added to the LAG dataset (original image set). Gaussian noise is predominant if the images were captured under low illumination. Salt-pepper noise is an impulse noise occurring owing to intense and sparse disturbances. Figure 8 depicts the original and noise-added image (sample). Experimentation is performed with Gaussian noise and salt-pepper noise, with the variance (σ) and noise density (d) varying from 0.1 to 0.5, respectively, and the result is reported in Figure 9.

It is seen from Figure 9 that the accuracy remains fairly the same. This demonstrates that the suggested model is extremely robust in both original and degraded datasets.

Table 10. Comparison of DBN-EHO with CNN models employing transfer learning.

Dataset	Classifier	Accuracy (%)	Sensitivity (%)	Specificity (%)
Drishti-GS	AlexNet	93.84	91.57	92.88
	GoogLeNet	95.46	90.34	93.36
	VGG16	94.12	95.77	96.42
	DBN-EHO	96.95	98.56	97.44

compares the proposed model with well-known CNN models employing transfer learning. Transfer learning is applied to the DRISHTI-GS1 dataset, as the number of images is small. It is inferred that our model can work well compared to well-known, pre-trained models using transfer learning.

Table 11. Comparison of proposed work.

References	Features/Methods	Classifier Used	Images	Performance (%)
Karthikeyan and Rengarajan [65]	GLCM	BPN	Local dataset	Accuracy—95
Issac, A. et al. [19]	CDR, NRR, blood vessel features	SVM and ANN	67	Accuracy—94.11
				Sensitivity—100
Mookiah et al. [66]	Discrete wavelet and HOS	SVM	60	Accuracy—95
				Sensitivity—93.3
				Specificity—96.67
Gifta [24]	GLCM, HOG, SURF	Gray Wolf Optimized NN	N.A.	Accuracy—93.1
				Sensitivity—91.6
				Specificity—94.1
Acharya, U.R. et al. [22]	6 features from LM filter bank	KNN	NA	Accuracy—95.8
Koh, J.E. et al. [20]	PHOG, SURF features	KNN	910	Accuracy—96.21
				Sensitivity—97.42
Samanta et al. [21]	Haralick features	BPN	60	Accuracy—96.26
				Sensitivity—90.43
				Specificity—99.5
Acharya et al. [67]	Texture and HOF	RF	60	Accuracy—91
Acharya et al. [68]	Gabor transformation and principal component analysis	SVM	510	Accuracy—93.10; sensitivity—89.75; specificity—96.20
Yadav et al. [69]	Homogeneity, Contrast, energy, correlation, entropy	N.N.	20	Accuracy—72
Maheshwari et al. [70]	Entropy and fractal	SVM	488	Accuracy—95.19
Bajwa M. N et al. [71]	ROI, Scaling	2-Stage CNN	ORIGA	AUC—0.87
Raghavendra et al. [72]	-	20 layer CNN	1426	Accuracy—98.13
Chen et al. [73]	-	16 layer CNN	SECS, ORIGA	AUC—0.881
Proposed Work	Structural, intensity, and texture features	DBN and EHO	7280	Accuracy—99.34
				Sensitivity—100
				Specificity—99.89

illustrates a comparison of similar other techniques and our technique in diagnosing glaucoma. Many optimization algorithms are also used to solve the other various optimization problems [48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64].

5. Conclusions

Glaucoma is a class of ocular neuropathy wherein the optic nerve gets vandalized, resulting in permanent vision loss. Glaucoma detection in RFI using the DBN classifier is proposed in this work. Input images from different public databases are enhanced through the preprocessing phase. The OD and OC are then segmented using CHT and Otsu’s thresholding. The various structural, intensity, and textural features are extorted from the segmented OC and OD images and fed to the DBN classifier optimized by the EHO algorithm. To investigate the performance, the technique is compared with classifier techniques, such as R.F., KNN, SVM, pre-trained CNNs, etc. Experimental outcomes exhibit the performance of the DBN classifier in recognizing the absence or presence of glaucoma accurately compared to other approaches. The proposed work concentrates on some of the important features contributing to glaucoma disease. The ReliefF algorithm selects the features before feeding them to the classifier for classifying them into either healthy or glaucomatous. The classifier performance is improved through an optimization technique where EHO does the weight update process. Dataset imbalance is also minimized as the model showed better training accuracy when images were selected randomly from each set using 10-fold cross validation. One potential drawback of this method is that it is unclear whether the attained specificities and sensitivities will be generalizable to real-world patient populations with common comorbidities, such as cataracts and surface ocular disease, which can degrade the input image quality. The performance of the computational hardware needs to be improved, along with network structure refinement and data dimension reduction, to attain competitively better computational speed. In the future, a universal domain adaptation method for various datasets (both public and private) is needed to be developed using hybrid weighted deep adversarial learning [74] and adaptive on-line validation [75]. In the future, the detection of glaucoma systems will be enhanced using the feature selection and ranking phase with different hybrid optimization algorithms. Besides, granular computing will be embedded in deep neural networks (granulated CNN) to enhance the computation speed significantly.