Hybrid Whale and Gray Wolf Deep Learning Optimization Algorithm for Prediction of Alzheimer’s Disease

Abstract

In recent years, finding the optimal solution for image segmentation has become more important in many applications. The whale optimization algorithm (WOA) is a metaheuristic optimization technique that has the advantage of achieving the global optimal solution while also being simple to implement and solving many real-time problems. If the complexity of the problem increases, the WOA may stick to local optima rather than global optima. This could be an issue in obtaining a better optimal solution. For this reason, this paper recommends a hybrid algorithm that is based on a mixture of the WOA and gray wolf optimization (GWO) for segmenting the brain sub regions, such as the gray matter (GM), white matter (WM), ventricle, corpus callosum (CC), and hippocampus (HC). This hybrid mixture consists of two steps, i.e., the WOA and GWO. The proposed method helps in diagnosing Alzheimer’s disease (AD) by segmenting the brain sub regions (SRs) by using a hybrid of the WOA and GWO (H-WOA-GWO, which is represented as HWGO). The segmented region was validated with different measures, and it shows better accuracy results of 92%. Following segmentation, the deep learning classifier was utilized to categorize normal and AD images. The combination of WOA and GWO yields an accuracy of 90%. As a result, it was discovered that the suggested method is a highly successful technique for identifying the ideal solution, and it is paired with a deep learning algorithm for classification.

1. Introduction

Dementia is a type of memory impairment that affects around 50 million individuals globally today. Alzheimer’s disease (AD) is a degenerative illness characterized by the formation of amyloid proteins in brain cells. An increase in amyloid protein levels induces cell death and impairs signal transmission. AD is an increasing public health problem among developed countries [1]. It can be characterized by impaired intellectual functioning, which interferes with day-to-day activities or personal relationships. The impairment comprises social behavior, memory loss, thinking, and language problems. Aging is the major factor causing AD, which mainly affects people above the age of 65 years [2]. CSF-containing ventricle channels are significantly expanded in Alzheimer’s disease patients [1,2]. The GM, CC, WM, HC, and ventricle are the biomarkers to diagnose AD. Because of the location of the medial temporal lobe, variations in the ventricular shape lead to changes in GM (MTL). Ventricle changes are utilized to calculate the rates of hemispheric atrophy. The quantity of neurofibrillary tangles and the degree of senile plaque deposition in nerve cells are significantly connected to the ventricular volume, which causes GM and WM shrinkage [2]. Tissue loss usually starts in the gray matter and moves to the white matter, corpus callosum, and hippocampus. A rapid loss of tissue signals the early stages of Alzheimer’s disease. Unmyelinated and myelinated neurons are both included in WM and GM. GM is composed of two components: neural connections and processing. GM is made up of a large number of cells that are situated in WM. WM serves as an information highway by connecting GM with the rest of the body components. The CC is a brain subregion that links the cerebral hemispheres, specifically the left and right hemispheres, and assists in communication between them. Apart from these entire regions, one of the major biomarkers for AD is the HC since it is completely responsible for creating new memories and recovering old ones. The primary function of the HC is to retain information in the long-term memory bank and to advance spatiotemporal communication during sleep. Tissue loss is often greater in the HC area than in the other brain subregions. Changes in the structure of the brain’s subregions, especially the GM, WM, CC, ventricle, and HC, are diagnostic markers for Alzheimer’s disease. However, it has been demonstrated that the HC area is more important than the other regions in diagnosing AD due to its direct link to memories [3,4,5]. As a result, the segmentation of brain SRs is an essential step in distinguishing between normal and AD brains. Currently, the holding technique is a prominent method for image segmentation. To differentiate between normal and Alzheimer’s disease brain tissues, SRs such as the GM, WM, CSF, ventricle, CC, and HC must be segmented. For image segmentation, thresholding approaches have been classified as bi-level and multilevel. Although bi-level thresholding can be applied to normal images, it is computationally exorbitant [6,7]. Multilevel thresholding, on the other hand, may segment complicated pictures. The optimization approach can then be used to solve the disadvantages of bi-level thresholding strategies. The primary goal of applying optimization techniques is to tackle tough issues in a limited amount of time and with limited resources.

There are various problems characterized as optimization problems in real time. If complexity increases, the need for optimization methods also increases. In the initial stage, mathematical techniques were used to optimize these problems, and then heuristic optimization methods such as genetic algorithms (GAs) were proposed to simulate Darwinian evolution [8]. Next, bird behavior was mimicked by using Particle Swarm Optimization (PSO) [9]. These two algorithms have been upgraded and applied in different areas [10,11]. Additionally, swarm intelligence methods such as ant lion optimization (ALO) [12], the whale optimization algorithm (WOA) [13], moth–flame optimization (MFO) [14], the dragonfly algorithm (DA) [15], etc., have been developed to solve many different problems. Similarly, various authors have introduced different algorithms, such as the monarch butterfly optimization algorithm (MBO) [16], the Enhanced Slime Mold Algorithm (SMA) [17], Runge–Kutta threshold segmentation [18], etc., for image segmentation, and they have produced better results. All of the above swarm intelligence algorithms mimic nature. Sometimes, metaheuristic algorithms will have limitations in obtaining optimal solutions due to poor exploration and exploitation phases [19]. In this situation, there is a need to apply hybrid methods to achieve the highest performance. Classification is crucial in distinguishing between normal control (NC) and AD pictures. Various categorization approaches have been presented by various authors, including Self-Organizing Maps (SOMs), neural networks (NNs), K-Nearest Neighbors (KNN), Support Vector Machines (SVMs), and others. These algorithms use feature extraction and feature reduction approaches to operate with previously selected features. Deep learning, on the other hand, does not require the preselection of features for categorization because it contains a collection of layers to perform all of these processes during the categorization process [20].

Initially, the presented work started with the acquisition of images, which were collected from Chettinad Health City. These images are real-time images, and they may have some distortion. Therefore, a preprocessing stage was implemented to lessen the complexity and improve the quality of MR images. Disease identification and brain alterations were then investigated by introducing the WOA for segmenting brain subregions. Following that, in order to increase the performance of the WOA exploration and exploitation phases, attain targeted global optimal solutions, and solve real-time, large-scale problems, this system adds the HWGO, which is a hybrid technique. With various metrics, qualitative and quantitative analyses have been used. Finally, clinical score validation was performed for clinical analysis. These overall processes will help to clinically detect normal and AD brains.

2. Literature Review

AD is a memory disorder that causes cognitive decline. According to a statistics report, nearly fifty million people are affected by AD across the world [1]. Various authors have investigated AD pathologies and biomarkers related to AD [1,2,3,4,5,21]. AD destroys the cells and inhibits neuronal connectivity [22]. Relevant genes have been identified in all regions to diagnose the AD region [23]. Linear tests have been conducted to measure brain atrophy in the early stages [24]. For this reason, many authors have concentrated on brain subregions, namely, the GM, WM, HC, ventricle, and CC, to diagnose AD. The relationship between HC atrophy, WM disruption, and GM hypometabolism for AD was examined by Villan et al. [25]. According to the above literature, the CC, GM, ventricle, WM, and HC brain regions are major hallmarks of AD. Hence, it is essential to segment these entire regions to analyze AD. Several methods, including clustering-based, region-based, thresholding-based, and graph-based approaches, have been used in attempts to segment the BSRs. Thresholding is a widely used approach for picture segmentation nowadays and is derived from these methodologies. In image segmentation, thresholding methods have been divided into bi-level (just two classes, such as the background and object, with the same threshold) and multilevel (pixels with several classes). Normal photos can employ bi-level thresholding, although it is computationally expensive. Multilevel thresholding, however, has the ability to segment complex images [6]. In order to handle challenging problems within time and resource constraints, multilevel thresholding has been created to overcome the drawbacks of bi-level thresholding techniques. The metaheuristic optimization technique is the most popular method for segmentation and has been used by many researchers in the past few decades due to its flexibility, simplicity, robustness, and so on [26]. The most famous metaheuristic algorithms include Bacterial Foraging Optimization (BFO) [27], GA [28], Artificial Bee Colony Optimization (ABC) [29], moth–flame optimizer (MFO) [30], Ant Colony Optimization (ACO) [31], Particle Swarm Optimization (PSO) [9], Cuckoo Search (CS) [32], and so on. Most of the above-mentioned techniques become stuck in the local optimal solution, and this reduces the quality of segmentation. Hence, this system introduced a hybrid method with the combination of the WOA and GWO to increase the exploration and exploitation stages. GWO is a metaheuristic optimization algorithm that is inspired by the standard leadership characteristics of hunting. Similarly, the WOA replicates humpback whales’ natural behavior. The WOA is a nature-inspired optimization technique that mimics the biological behavior of humpback whales. These whales are mostly found in groups, and they are experienced in special hunting methods, such as the bubble-net feeding strategy [33]. The main difference between the WOA and GWO is that the WOA employs a spiral to replicate a bubble net to hunt/chase the prey, whereas GWO simulates hunting by selecting a random search agent to find/chase the prey. As a result, the hybrid WOA and GWO optimization approach has been developed to increase the WOA performance and locate global optima. Korashy et al. applied a hybrid WOA/GWO solution to the coordination challenge of guiding overcurrent relays [34]. Using the strengths of the WOA and GWO, improved hybrid optimization algorithms have been implemented, and global optimal solutions have been obtained [35]. The primary goal of hybridization is to increase the accuracy of separate algorithms by combining their advantages and avoiding their disadvantages. Similarly, the WOA and GWO complemented each other to improve segmentation quality. In general, analyzing AD based on structural variation necessitates segmentation and classification processes. For the past few decades, the most widely used classification methods have been Support Vector Machines (SVMs), Radial Basis Function Networks (RBFNs), neural networks (NNs), etc., for classifying normal and AD images [35]. Generally, the classification includes four steps: (1) feature extraction, (2) feature selection, (3) feature reduction, (4) feature-based classification. It consists of multiple processes and requires more time to execute. Hence, there is a need to overcome these difficulties using the current emerging techniques for classification, such as deep learning (DL) methods, which have been used. DL is a subset of machine learning (ML), and it has a deep architecture to provide efficient classification. According to a systematic review, DL is the best classifier for neuroimaging data to classify normal and AD images [36,37]. The presented work started with image acquisition from real-time hospital images, which were preprocessed to remove non-target areas, such as the skull, skin, and muscle. This step will help to improve the quality of the images. Then, the brain’s internal regions, namely, the ventricle, GM, HC, WM, and CC, are segmented using WOA and HWGO techniques. The performance was tested using several measures to compare the ground truth and segmented areas. Finally, the SR images were input to the CNN classifier, which classifies the NC and AD.

3. Proposed Framework Methodologies

This study’s workflow is drawn in Figure 1, which contains five steps: data acquisition, preprocessing, segmentation, ground-truth validation, and classification.

3.1. Dataset Availability

The suggested method uses hospital brain scans to partition the brain areas into the ventricle, GM, HC, WM, and CC. T2-weighted brain MR images with axial, coronal, and sagittal slices were used. The parameters are shown in

Table 1. Proposed dataset details collected from Chettinad Hospital.

Particulars	NC	AD
Number of datasets	100	100
Sex (M/F)	73/27	68/32
Age group	60.5 ± 5.5	67.5 ± 7.5
Clinical Dementia Rate (CDR)	0	1 and 2
Mini-Mental Score Examination value	26 ± 3.5	16 ± 2.5

. The suggested approach segments the ventricle, HC, CC, and GM regions brain using hospital MR brain images as the input. Chettinad Hospital provided T2-weighted slices of sagittal, coronal, and axial brain MR images for the present investigation. The scanner has a 1.5-tesla standard and can produce images with 1 mm spacing, 5 mm thickness, and 256 × 256 resolution. A total of 100 NC and 100 AD photos were gathered, with a 60–75 age range. The ventricle and GM were segmented using axial slices. Sagittal slices were chosen to segment the CC. Similar to the previous example, coronal slices were used to segment the HC, which aids in clinical and epidemiological investigations and allows for examining both sides (the left and right) of the HC. Sagittal and axial slices cannot clearly show the HC region on both sides.

3.2. Preprocessing

Generally, hospital images are not of good quality due to their homogeneous intensity values of both the object and background region. Since the segmentation of the brain’s inner regions, especially the GM, CC, HC, ventricle, and WM, is a challenging task due to their similar intensity values, there is a need to apply preprocessing for image quality improvements. Initially, the adaptive histogram equalization approach was employed to enhance the contrast and extract distinct regions, and Otsu’s thresholding method was utilized to remove undesired areas, such as the skull, skin, and muscle. Adaptive histogram equalization was used to convert each pixel using a transformation function derived from a neighboring region. Similarly, Otsu’s method produces satisfactory outcomes in removing unwanted regions due to its technique of maximizing the between-class variance for extracting a better optimal solution. This approach aids in the extraction of non-brain and brain tissues with weak boundaries and equal intensities. These processes are helpful in numerous clinical applications to obtain very good accuracy, and they are key resources for brain region segmentation. The methodology of the study is presented in Figure 1.

3.3. Segmentation

The process of splitting an image into various items is known as segmentation, and it is a time-consuming procedure owing to unsatisfactory results, computational overhead, and a low generalization capability. To overcome these drawbacks, optimization procedures are designed to attain the best optimal value. Therefore, optimization using the WOA has been proposed to segment the brain’s inner regions. However, it has drawbacks in identifying search agents (SAs) for obtaining the optimal solution. Hence, the presented work concentrated on a hybrid of the WOA and GWO to achieve an accurate solution.

3.3.1. Whale Optimization Algorithm (WOA)

The WOA is a well-known optimization approach for solving complicated issues. It contains three operators: searching for prey, enclosing prey, and the bubble-net foraging behavior of humpback whales. Bubble-net foraging, which uses separate bubbles, is a unique method of hunting. The WOA’s hunting process is outlined as follows. Encircling prey: Over multiple cycles, humpback whales encircle their prey and update their positions until they achieve the optimal solution [34]. The main aim of bubble-net foraging is to identify the subregions of the brain (this will be called the hunting of subregions) in terms of obtaining the optimized solution.

The prey can be attacked by humpback whales using the spiral bubble-net approach, which is described in the following. The shrinkage of the encircling mechanism is used to decrease the coefficient vector $A_1$:

$$A_1=2a_1.r1-a_1,$$

where r1 is a random number between 0 and 1.

Updating the spiral position: The motion of humpbacks is mathematically stated as follows:

$$X_1\left(n+1\right)={D^{\prime }}_1*e^{bv}*\cos \left(2\pi v\right)+X{\prime }_1\left(n\right),\hspace{1em}\hspace{1em}\hspace{1em}D{\prime }_1=\left|X_1\left(n\right)-{X^{\prime }}_1\left(n\right)\right|,$$

where v is a random number, b is a constant (it determines the spiral’s form), ${X^{\prime }}_1$ is the vector of the prey’s location, and$X_1$ is the humpback whales’ location vector. Generally, humpback whales roam within the circle of prey in their spiral path. Then, the probability of selecting the spiral model or encircling mechanism is updated with the following equation:

$$X_1\left(n+1\right)=\{\begin{array}{c}{X^{\prime }}_1\left(n\right)-A_1*D_1 if p_1<0.5\hfill \\ {D^{\prime }}_1*e^{bv}*\cos \left(2\pi v\right)+{X^{\prime }}_1\left(n\right),\hspace{1em}if p_1<0.5\hfill \end{array},$$

where $p_1$ is a random value within [0, 1].

Searching for prey: Prey is searched for randomly by humpback whales. The location of the whale is updated by using a random search agent and then the best search agent, which is described below:

$$D_1=\left|C_1*X_{rand}-X_1\right|,\hspace{1em}\hspace{1em}\hspace{1em} X_1\left(n+1\right)=X_{rand}-A_1*D_1, X_{rand}:$$

which is the random position vector selected from the current position [26,34,35].

3.3.2. Gray Wolf Optimization (GWO)

GWO is a metaheuristic algorithm that applies a special leadership quality to the hunting mechanism [34]. The mathematical process of GWO involves the societal hierarchy, encircling prey, searching for prey, and hunting and attacking prey, which is discussed in the following.

Societal hierarchy: The leader wolves are alpha (α), which is the first level. Similarly, the second- and third-level wolves are beta (β) and delta (δ).

Encircling prey: Prey is encircled by gray wolves during hunting ($D_1$), which is mathematically represented by the following formulas [34]:

$$D_1=\left|C_1*X_p\left(n\right)-X_1\left(n\right)\right|,\hspace{1em}\hspace{1em}{X}_1\left(n+1\right)=X_p\left(n\right)-A_1{D}_1$$

where $X_p$ is the prey position vector, n is the current iteration, $X_1$ is the gray wolf position vector, and $A_1$ and $C_1$ are coefficient vectors [34].

$$A_1=2a_1.r1-a_1,\hspace{1em}\hspace{1em}{C}_1=2* r2,$$

where r1 and r2 are random numbers between 0 and 1.

Hunting: The mathematical process of hunting is defined below [34]:

$$D{1}_{\alpha }=\left|C_1*X_{\alpha }-X\right|, D{1}_{\beta }=\left|C_2*X_{\beta }-X\right|,\hspace{1em}\hspace{1em}D{1}_{\delta }=\left|C_3*X_{\delta }-X\right|,$$

$$X_1=X_{\alpha }-A_3*d_{\alpha },\hspace{1em}\hspace{1em}\hspace{1em}{X}_2=X_{\beta }-A_3*d_{\beta },\hspace{1em}\hspace{1em}\hspace{1em}{X}_3=X_{\delta }-A_3*d_{\delta },$$

where $X_{\alpha }$, $X_{\beta }$, and $X_{\delta }$ are the α (alpha), β (beta), and δ (delta) gray wolves’ positions [38].

Finally, the positions of the SAs can be updated by using the following formula [35,39]:

$$X\left(n+1\right)=\frac{X_1+X_2+X_3}{3}$$

Attacking the prey: If |A| < 1, the wolves move toward the prey to attack. Searching for prey: If |A| > 1, the wolves move away from the prey for filtering [34].

3.3.3. Hybrid WOA and GWO

The presented work implemented a hybrid of the WOA and GWO to improve the performance of the WOA. The improvements were achieved by the leadership characteristic of GWO and were applied to the WOA bubble-net strategy. Because of its natural aptitude for hunting, GWO picks the three best options, such as alpha, beta, and delta, which are utilized in the WOA exploitation stage to adjust the SA’s location and improve the WOA’s performance [34]. Bubble-net attacking process: The proposed hybrid method updates the positions of the whales using the natural leadership behavior of GWO by using a mathematical representation, which is given below [34,35]. Shrinking encircling mechanism: The positions of humpback whales are updated with the following: $X\left(n+1\right)=\frac{X_1+X_2+X_3}{3}$ [26,34]. Updating the spiral position: The humpback whales’ positions are updated along with a spiral-shaped path, which is given below [26,34]:

$$D1{\prime }_{\alpha }=\left|C_1*X_{\alpha }-X\right|, D1{\prime }_{\beta }=\left|C_2*X_{\beta }-X\right|, D1{\prime }_{\delta }=\left|C_3*X_{\delta }-X\right|$$

$$X_1\left(n\right)=X_{\alpha }\left(n\right)+{D^{\prime }}_{\alpha }*e^{bv}*\cos \left(2\pi v\right),$$

$$X_2\left(n\right)=X_{\beta }\left(n\right)+D{\prime }_{\beta }*e^{bv}*\cos \left(2\pi v\right)$$

$$X_3\left(n\right)=X_{\delta }\left(n\right)+D{\prime }_{\delta }*e^{bv}*\cos \left(2\pi v\right)$$

$$X\left(n+1\right)=\frac{X_1+X_2+X_3}{3}$$

Pseudocode for the proposed hybrid algorithm (Algorithm 1):

Algorithm 1. Pseudocode

S1: Initial population of the SA has been generated  S2: Determine the fitness function for each search agent.  S3: $X_{\alpha }$: first best solution  S4: $X_{\beta }$: second best solution  S5: $X_{\delta }$: third best solution  S6: if1 n < max no. of iterations  S7: for i = 1 to no. of search agent  S8: update the parameters of A, C, a, l, and p  S9: if2 (p1 < 0.5)  S10: if3 (|A| < 1)  S11: search agent position has been updated.  S12: else if3 (|A| ≥ 1)  S13: random agent selection ($X_{rand}$) and update the search agent position.  S:14: end if3  S15: else if2 (p1 ≥ 0.5) and update the position of the search agent.  S16: end if2  S17: checks the search agent position whether it has gone outside of the assigned search space.  S18: positions of $X_{\alpha }$, $X_{\beta }$, $X_{\delta }$ are updated.  S19: update the number of iterations  S20: end if  S21: R$\mathrm{eturn} X_{\alpha }$

3.3.4. Ground-Truth Validation

Quality evaluation methodologies, including Feature Similarity Index Metrics and Structure Similarity Index Metrics, have been utilized to determine the validity of the similarity between GT and SR [39,40].

3.4. Classification

Deep learning (DL) is used for deep classification, which is also part of machine learning (ML). The input and output layers of DL have a standard topology, with hidden layers oriented toward the network architecture. The basic purpose of choosing a DL classifier is to obtain extremely high classification performance. One of the basic DL architectures is the Convolution Neural Network (CNN), which consists of many layers for visual feature extraction and classification [20]. The Convolution Neural Network (CNN), which has many layers and is the main DL network, is used to classify and reduce the number of features in images. In order to complete feature extraction and classification, this work focused on the CNN. This network has a large number of convolutional filters that extract image features. This work used AlexNet for feature extraction and classification, which aids in the categorization of NC and AD patients because of its various convolution filters that extract visual characteristics. Sensitivity, accuracy, and specificity are calculated as outcomes of the classification, as are statistical measures such as true positive (TP), false negative (FN), true negative (TN), and false positive (FP) rates. These measurements also compute the recall, accuracy, and F-score in order to assess the performance of the proposed approach in Alzheimer’s disease prediction [38].

3.5. Clinical Score Validation

The MMSE is a type of test used to assess neurobiological and physical disabilities. The MMSE score and patient history report are required for the appropriate medical evaluation to classify NC and AD patients [41]. It is used for cognitive screening, which offers grading based on impairment, and it has become the standard test for monitoring the phases of Alzheimer’s disease. The MMSE is made up of eleven questions that cover six cognitive functions: short-term recollection, language, immediate recall, attention, and the ability to track basic verbal and written directions. The suggested approach for determining the association between the SR and cognitive functions involves comparing the MMSE to the SR intensity value. The MMSE score typically ranges from 0 to 30 points and declines as the condition progresses. NC is defined as a score between 26 and 30, whereas AD is defined as a score less than or equal to 18 [42]. Mild cognitive impairment is indicated by a score between 18 and 24 (MCI).

4. Results and Discussion

Initially, images for both NC and AD patients were obtained from Chettinad Hospital. Based on the doctor’s contribution, this study took into account around 200 images from real-time public databases. AD is a neurological/neuropathological illness that affects brain structures and produces a variety of cognitive impairments, including increased confusion, loss of knowledge, impairment in memory skills, and difficulties comprehending and performing activities. As a result, there is a need to focus on the brain’s interior areas, such as the ventricle, CC, GM, HC, and WM, when analyzing AD. The suggested system considers all three slices (axial, coronal, and sagittal) of the encephalon/brain to segment all of the above-mentioned areas. Figure 2 shows the raw pictures of all three slices of NC and AD images. Because of anatomical differences, the encephalon’s inner tissues are very small and difficult to distinguish. Because the interior parts of the brain have undefined borders and have relatively low contrast, they must be preprocessed.

To identify the target region of the brain, preprocessing steps such as skull stripping and contrast enhancement were conducted to produce good-quality images. Otsu’s thresholding was used in skull stripping to eliminate non-brain tissues from input images, such as the fat, scalp, muscle, skull, and skin. Similarly, histogram equalization was applied in order to preserve information with high contrast. Figure 3 shows the preprocessed images.

Implementation of Hybrid Technique

Generally, the segmentation of brain regions is a problematic process since the brain has a compound structure. Hence, the presented work started with the implementation of the WOA method for segmenting the brain’s inner regions, which are the ventricle, GM, CC, WM, and HC. The WOA has two phases of hunting: exploitation and exploration. The WOA employs the bubble-net attack tactic for hunting during the exploitation stage. Though the conventional whale optimization algorithm produces the best results in some cases, this technique may rely on the local optimal solution. As a result, in order to address this issue, a hybrid of the WOA and GWO (HWGO) was developed for the segmentation process in order to enhance the hunting exploitation process. In the suggested system, GWO’s management and leadership qualities may be exploited in the WOA’s bubble-net hunting approach. Throughout the WOA exploitation stage, the proposed algorithm selects the three best solutions from all (combination of the WOA and GWO) search agents, and other SAs adjust their locations based on their positions.

The major parameters of the WOA and HWGO are the number of search agents, number of iterations, and number of threshold values, which are listed in

Table 2. The optimal values for the proposed algorithms.

Parameters	WOA					HWGO
	Ven	GM	WM	CC	HC	Ven	GM	WM	CC	HC
No. of SAs	100	100	100	100	100	100	100	100	100	100
I	100	100	100	100	100	100	100	100	100	100
T1	8	5	5	10	13	8	5	5	10	13

. The process of fitness function evaluation is repeated until it reaches the best solution. The ending criteria were set to the number of iterations (I) and the number of thresholds (T1). These parameters were experimentally tuned quantitatively and qualitatively to achieve the best solution for segmenting all of the above-mentioned regions. The qualitative measures specify that if the value of T is small, then the SR also includes neighborhood pixels. Similarly, when the value of T is large, the SR regions are blurred. Finally, we obtained the T values for the GM, CC, WM, ventricle, and HC, which were 5, 10, 5, 8, and 13, respectively. Then, the next parameter designated for selecting the optimal value is the number of SAs, and it was fixed at 100 for segmentation. The segmented brain’s inner regions, especially the GM, ventricle, CC, HC, and WM, are presented in Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8 for the different optimization techniques. The segmented output clearly shows the qualitative performance of the WOA and HWGO for various regions. According to Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8, the regions can be segmented more accurately with HWGO than with the WOA due to its quality of searching for the global solution. Based on the qualitative analysis, it is demonstrated that the WOA displays very low performance compared to HWGO due to its naturally poor encircling (bubble-net attack) mechanism during hunting.

The parameter values are finalized with the quantitative analysis of GT and SR images. The GT images were collected from Chettinad Hospital with the approval of the doctors. After the segmentation operation, the SR results were validated using GT pictures and numerous performance indicators, including FSIM and SSIM. These metrics serve to establish the comparability of GT and SR based on image properties and structure. In the same way, these measures reveal the accuracy of the segmentation results for NC and AD. The measures were calculated for the brain’s five inner regions with the WOA and the hybrid of the WOA and GWO. These measures clearly illustrate the improvement of the WOA in hybrid segmentation due to the implementation of GWO in the WOA. GWO provides better support in identifying the best solution in terms of finding the three best solutions during the encircling mechanism. Similarly, HWGO improves the performance of the WOA due to its combination with GWO in finding the similarity between the SR and GT of the brain’s inner regions. Finally, the accuracy of HWGO is 94% in both the FSIM and SSIM for the HC, GM, ventricle, WM, and CC. This proves that the segmentation performance of the WOA is improved since GWO enhances the reliability and quantitative performance of the WOA in HWGO (Figure 9). The suggested technique (HWGO) outperforms the WOA in the segmentation of the GM, ventricle, WM, CC, and HC because it features an encircling mechanism that finds the best results. In comparison, the suggested method shows increased performance because it shows the highest similarity between the SR and GT regions of the GM, ventricle, CC, WM, and HC. It has a significantly higher FSIM and SSIM accuracy of 94 percent compared to the WOA (Figure 9).

Table 3. Statistical measures using GT and segmented hippocampus.

Statistical Measures	WOA	HWGO	WOA	HWGO	WOA	HWGO	WOA	HWGO	WOA	HWGO
	GM		WM		Ventricle		CC		HC
Accuracy	0.88	0.91	0.89	0.91	0.89	0.91	0.912	0.934	0.92	0.94
Sensitivity	0.89	0.90	0.89	0.90	0.89	0.90	0.90	0.915	0.90	0.93
Specificity	0.88	0.90	0.88	0.90	0.88	0.90	0.91	0.91	0.91	0.91

displays the accuracy, sensitivity, and specificity of GT and SR, which are utilized to categorize normal control and AD pictures using the DL classifier. Based on these measures, it is observed that both segmentation and classification showed better optimal solutions in the hybrid method, and it improves the exploration and exploitation levels of the individual method. Finally, the presented work applied the Mini-Mental Score Evaluation to compare the results with the clinical status.

Figure 10 depicts the association between the pixel intensity and MMSE score in the HC area. The normal pixel intensity of the HC region is between 91 and 94, whereas that of Alzheimer’s disease is between 81 and 85. From the results of this validation, the structural anomalies of normal and Alzheimer’s disease patients are assessed using the MMSE score. The graph clearly shows that the MMSE score in both normal and AD patients is highly related to the number of intensity values. As a result, the data collected are therapeutically useful, and they might be utilized to screen Alzheimer’s patients. According to the segmentation and classification results, the HC area is a responsible/key region for identifying an abnormal image of a patient. Hence, the MMSE score was examined for the HC region to differentiate between AD and normal controls, the result of which is shown in Figure 10.

Computational complexity was calculated with the number of iterations and the fitness value. The number of iterations is taken as a major factor in attaining the optimal solution. The convergence characteristic graph for segmenting internal brain regions in AD and normal images using the WOA and the HWGO algorithm. The graph was created by plotting the number of iterations against the fitness score. For the optimization techniques described above, the maximum number of iterations was set at 100.

Table 4. Computational complexity of the proposed optimization algorithm.

Optimization Algorithm	WOA	GWO	HWGO
Number of iterations to attain optimal solutions/maximum number of iterations	82/100	62/100	75/100
Time complexity	30.425	23.782	26.782

demonstrates that the HWGO method is more stable in reaching global optimal solutions with fewer iterations.

Generally, region extraction from the brain is more sensitive, and it is a more challenging task in identifying/classifying disease progression. However, it is very difficult. It is necessary to identify and extract individual regions to analyze major diseases, such as dementia, brain cancer, AD, epilepsy, and others. As a result, various authors have proposed various methods, such as Fuzzy C means, GWO, BAT, ABC, WDO, and so on, for brain disease progression and have shown better outcomes [43,44,45,46,47,48,49,50]. According to

Table 5. Comparison between existing and proposed techniques.

Authors	Selected Regions	Accuracy %
Pang et al. [47]	HC	87.7
Singh and Bala [48]	WM, GM, CSF	91.7
Pham et al. [10]	WM, GM, CSF	92.9
Proposed techniques	GM, CC, ventricle, WM and HC	94

, the highest performance was shown by HWGO compared to other methods. According to the findings, the HC provides more signals for diagnosing atrophy in all situations. The atrophy of the HC has a substantial influence on memory. According to the study, the hippocampal region is the most impacted region in the early stages of Alzheimer’s disease compared to other areas since it is located deep inside the brain; it is important for memory assessments, and it is used as a predictive marker for studying the disease. Similarly, the proposed strategy emphasizes the significance of the HC sector.

5. Conclusions

The most difficult process in diagnosing illness progression is the segmentation of brain areas such as the ventricle, GM, CC, WM, and HC. This research focuses on brain subregion segmentation in order to identify Alzheimer’s disease and normal controls. Initially, this work started with data collection from Chettinad Hospital. Then, it moved on to preprocessing for contrast enhancement and skull stripping. Next, the segmentation and classification process was performed using the WOA and HWGO. Quantitative and qualitative analyses such as similarity measures, statistical measures, and clinical score validation were conducted by calculating the FSIM, SSIM, accuracy, sensitivity, and specificity for the optimal solution. Initially, this work was performed for five regions, specifically the GM, ventricle, CC, HC, and WM. As per clinical and experimental outcomes, the HC region is more suitable and responsible for the diagnosis. Thus, the HC area was quantified using two optimization approaches (WOA and HWGO) and DL, which were utilized to diagnose AD and NC. In general, the hospital images were of low quality. As a result, an attempt was made to improve image quality by adopting preprocessing techniques. This work combined the WOA with GWO to enhance the outcomes of the WOA. Based on this validation, HWGO delivered a maximum accuracy of 92 percent in segmentation when compared to other approaches since it has the capacity of a leader to pick the optimal solution. Because GWO accurately isolates genuine areas of brain subregions, the GWO approach was used with the WOA to assess the performance of the WOA. GWO demonstrated improvements and provided a superior solution in this hybridization. The results show that 94 percent of the data match GT. After classification validation, MMSE validation was also conducted to clinically prove the method’s performance using experimental values. According to the data, the proposed task successfully showed the best outcome. The suggested technique is beneficial for screening systems and as a research tool for neurodegenerative diseases, namely, Alzheimer’s disease.