DL-MDF-OH²: Optimized Deep Learning-Based Monkeypox Diagnostic Framework Using the Metaheuristic Harris Hawks Optimizer Algorithm

Abstract

At the time the world is attempting to get over the damage caused by the COVID-19 spread, the monkeypox virus threatens to evolve into a global pandemic. Human monkeypox was first recognized in Africa and has recently emerged in 103 countries outside Africa. However, monkeypox diagnosis in an early stage is difficult because of the similarity between it, chickenpox, cowpox and measles. In some cases, computer-assisted detection of monkeypox lesions can be helpful for quick identification of suspected cases. Infected and uninfected cases have added to a growing dataset that is publicly accessible and may be utilized by machine and deep learning to predict the suspected cases at an early stage. Motivated by this, a diagnostic framework to categorize the cases of patients into four categories (i.e., normal, monkeypox, chicken pox and measles) is proposed. The diagnostic framework is a hybridization of pre-trained Convolution Neural Network (CNN) models, machine learning classifiers and a metaheuristic optimization algorithm. The hyperparameters of the five pre-trained models (i.e., VGG19, VGG16, Xception, MobileNet and MobileNetV2) are optimized using a Harris Hawks Optimizer (HHO) metaheuristic algorithm. After that, the features can be extracted from the feature extraction and reduction layers. These features are classified using seven machine learning models (i.e., Random Forest, AdaBoost, Histogram Gradient Boosting, Gradient Boosting, Support Vector Machine, Extra Trees and KNN). For each classifier, 10-fold cross-validation is used to train and test the classifiers on the features and the weighted average performance metrics are reported. The predictions from the pre-trained model and machine learning classifiers are then processed using majority voting. This study conducted the experiments on two datasets (i.e., Monkeypox Skin Images Dataset (MSID) and Monkeypox Images Dataset (MPID)). MSID dataset values 97.67%, 95.19%, 97.96%, 95.11%, 96.58%, 95.10%, 90.93% and 96.65% are achieved concerning accuracy, sensitivity, specificity, PPV, BAC, F1, IoU and ROC, respectively. While for the MPID dataset, values of 97.51%, 94.84%, 94.48%, 94.96%, 96.66%, 94.88%, 90.45% and 96.69% are achieved concerning accuracy, sensitivity, specificity, PPV, BAC, F1, IoU and ROC, respectively.

1. Introduction

Over the past 20 years, there have been numerous viral illness outbreaks, including those caused by H1N1, Chikungunya, Zika, Ebola, Nipah, MERS, SARS, COVID-19 and H7N9 avian flu [1]. Recently, the world has started to recover from the COVID-19 pandemic; however, the recent outbreak of monkeypox has introduced increased concerns in global communities. In non-endemic areas, an unusual wave of monkeypox cases has been documented. Currently, the virus has spread to numerous nations outside of that endemic area (e.g., the UK, Spain, Portugal, the US, Sweden, Canada, Belgium, Australia, Italy, Germany, the Netherlands, France and Mexico) [2]. Less than 1% of the confirmed cases at this time are from endemic nations, making this outbreak distinct in that it is mostly spreading in Europe and the Americas [3]. Figure 1 shows the global and regional outbreaks of monkeypox cases [3,4]. According to the World Health Organization (WHO), a moderate risk to global public health is posed by this outbreak and no public health emergency is declared [5]. However, some organizations (e.g., World Health Network (WHN)) have expressed a raised concern [6] highlighting the necessity for immediate global action toward the disease.

Monkeypox is a zoonotic disease from the genus Orthopoxvirus. It is closely related to cowpox, measles and smallpox in clinical features and belongs to the Poxviridae family (a member of the genus Orthopoxvirus) [7]. Typically, it is transferred by rodents and monkeys, yet human-to-human spread is prevalent [8]. In 1958, the virus was initially identified in a laboratory in Copenhagen, Denmark, inside the body of a monkey [9]. In 1970, the first human case of this disease was detected in the Congo during a campaign to eliminate smallpox (i.e., chickenpox) [10]. In the central and western parts of Africa, many individuals who live close to tropical rainforests are affected by monkeypox. The virus spread when a person contacts an infected animal, person or material. Hence, it is transmitted through animal bites, respiratory droplets, direct body contact or mucous of the mouth, nose or eyes [11]. For monkeypox patients, the main symptom is a rash that is located on or near the genitals, anus and other areas (e.g., the face, mouth, hands, chest or feet). Other early-stage symptoms include fever, chills, body aches, swollen lymph nodes, fatigue, exhaustion and headache, wherein the long-term effect is red bumps on the skin [12]. To date, monkeypox is not extremely contagious, yet the cases continue to rise. In 1990, only 50 cases existed in west and central Africa [13]. Nevertheless, in 2020, the cases rose to 5000 cases. In 2022, the number of cases infected by the virus is 57,527, reported by 103 countries worldwide [4]. Thus, massive fear and anxiety among people are growing, which is reflected in opinions on social media.

It is worth mentioning that there is no suitable treatment for this virus, according to the guidelines provided by the Centers for Disease Control and Prevention (CDC) [14]. Nevertheless, the CDC approved two oral drugs (i.e., Brincidofovir and Tecovirimat). These drugs are mainly used to treat the smallpox virus and now are used to treat the monkeypox virus [15]. The ultimate solution to this virus is vaccination. In the United States, several vaccines are available for the monkeypox virus approved by the FDA (i.e., the Food and Drug Administration), but they have not yet been used for humans. In some countries, vaccines for the chickenpox virus are used to treat the monkeypox virus [16]. Its diagnostic procedure includes initial observations of the unusual characteristics of skin lesions present. Then, the history of exposure is investigated. Regardless, the monkeypox virus can be confirmed using a polymerase chain reaction (PCR) [13]. In addition, testing skin lesions using electron microscopy is the definitive way to diagnose the virus. The small dissimilarities in the skin rash of several diseases (i.e., chickenpox, cowpox and measles) along with the rarity of monkeypox made the early detection of it very challenging.

AI is a rapidly developing field that represents models that may be used in a variety of scientific fields [17,18]. By applying this knowledge to the current approach to decision-making in a particular area, AI can be thought of as a method of learning and recognizing patterns and interactions from a sufficient number of representative models [19,20]. Lately, machine learning (ML) models have shown tremendous potential in image-based diagnoses (e.g., tumor cell identification, cancer detection and COVID-19 detection) [13]. In the last decade, different deep learning (DL) algorithms have been used widely in several medical image tasks (e.g., organ abnormality detection [21], organ localization [22], gene mutation detection [23], cancer staging [24] and grading [25]). Additionally, DL models have been effective in detecting skin lesions automatically, as long as adequate training examples are available [26]. Accordingly, similar applications can be adapted to detect monkeypox cases. As images of the infected human skin can be acquired and then used in detecting this disease.

In the current study, a diagnostic framework categorizes patients’ cases into four categories (i.e., normal, monkeypox, chicken pox and measles). The major aim of the framework is to report a precise decision through the combination of pre-trained CNN models and machine learning classifiers. First, the hyperparameters of the pre-trained model are optimized through the usage of a metaheuristic optimization algorithm. After that, the features can be extracted from the feature extraction and reduction layers. Finally, these features are classified using machine learning models. The predictions from the pre-trained model and machine learning classifiers are then processed using majority voting. Currently, five pre-trained CNN models are utilized (i.e., VGG19, VGG16, Xception, MobileNet and MobileNetV2). The Harris Hawks Optimizer (HHO) algorithm is also used as the metaheuristic optimization algorithm. The seven utilized machine learning classifiers are Random Forest (RF), AdaBoost, Histogram Gradient Boosting (HGB), Gradient Boosting (GB), Support Vector Machine (SVM), Extra Trees (ET) and K-Nearest Neighbors (KNN). The current study contributions can be summarized in the following points:

-

Suggesting a hybrid algorithm using pre-trained CNN models and machine learning classifiers to diagnose monkeypox.

-

Utilizing the Harris Hawks Optimizer (HHO) metaheuristic algorithm to optimize the hyperparameters of the pre-trained model.

-

Extracting the features from the feature extraction and reduction layers to be used by machine learning algorithms.

-

Using the majority voting approach to process the predictions from the pre-trained CNN models and machine learning classifiers

The rest of the paper is organized as follows: Section 2 introduces the related work in this field. Section 3 examines the methodology, dataset gathering and preprocessing phase, fine-tuning the pre-trained models and optimizing the hyperparameters, classifying the extracted features using ML and majority voting and predictions. Section 4 analyzes the experimental results. In Section 5, the limitations of the current study, a conclusion concerning the paper and future work are presented.

2. Related Studies

Machine learning is a discipline of Artificial Intelligence (AI) with demonstrated promises in several domains, varying from industrial sectors and decision-making tools to disease diagnosis and medical imaging. Hence, doctors can obtain precise, secure and fast solutions in which the specific qualities of ML have gained acceptance as a practical decision-making tool [27,28]. It is worth mentioning that AI approaches played a powerful role in COVID-19 diagnosis from different medical images (e.g., chest X-ray, computed tomography (CT) and chest ultrasound). Thus, the scientific community is motivated to use AI-based approaches to diagnose monkeypox from medical skin images. Ahsan et al. [13] introduced a novel dataset containing the images of monkeypox infected patients, called “Monkeypox2022”, a publicly available dataset. Their dataset was constructed by gathering the patients’ images from multiple online open-source portals.

Additionally, a modified VGG16 model was suggested to perform monkeypox diagnosis. Their experiments comprised two separate studies (i.e., studies one and two). The two studies differed in the utilized parameters (i.e., batch size, learning rate and the number of epochs). The reported results showed that the modified model identified monkeypox patients with an accuracy of 97 ± 1.8% and 88 ± 0.8% for studies one and two, respectively. Further, The post-prediction and feature extraction results were overlooked and explained using a popular explainable AI technique named Local Interpretable Model-Agnostic Explanations (LIME). LIME aims to obtain deeper insights into distinctive features that represent the onset of this virus. It showed that the suggested models could learn and localize the infected areas. In Ali et al. [29], a dataset named “Monkeypox Skin Lesion Dataset” (MSLD) containing the skin lesion images of chickenpox, measles and monkeypox was constructed. Their dataset was collected from news portals, publicly accessible case reports and websites. Additionally, data augmentation was applied to increase the sample size. Further, three pre-trained DL models (i.e., VGG16, ResNet50 and InceptionV3) were utilized to detect monkeypox and other diseases. An ensemble of the three models was also implemented by utilizing majority voting. The best overall accuracy of 82.96 (± 4.57%) was achieved by ResNet50. On the other hand, the ensemble system and VGG16 model accomplished accuracies of 79.26 (± 1.05%) and 81.48 (± 6.87%), respectively. It is worth mentioning that the ensemble did not show superior results compared to the ResNet50 model based on averaged performance across the three-fold cross-validation experiment. Moreover, they created a web application that can be used for monkeypox screening online.

Islam et al. [30] proposed the first large monkeypox Skin Image Dataset 2022. In addition, their dataset was utilized to test the practicality of employing cutting-edge AI deep models to identify monkeypox on skin scans. Seven pre-trained DL models (i.e., ResNet50 [31], DenseNet121 [32], Inception-V3 [33], SqueezeNet [34], MnasNet-A1 [35], MobileNet-V2 [36] and ShuffleNet-V2 [37]) were applied to classify digital skin images into six categories (i.e., monkeypox, measles, smallpox, chickenpox, cowpox and healthy cases). Their chosen DL models vary in the number of parameters, varying from 1 to 26 million. The reported results showed that the achieved precision was 85% which was relatively low due to the small training samples. Additionally, five-fold cross-validation experiments were conducted. They proved that DL models might distinguish digital skin scans of measles and chicken pox lesions and rashes.

The DNA sequences of HPV (Human Papilloma Virus)-causing warts and MPV (Monkeypox Virus)-causing monkeypox were examined and deep learning was used to classify the samples in Alakus and Baykara [38]. There were four stages of the investigation. Initially, their DNA sequences identified the viruses that cause warts and monkeypox. Then, these sequences were mapped utilizing several DNA-mapping techniques. After that, the mapped sequences were categorized using a deep learning system. In the final phase, accuracy and F1-score were used to compare how well various DNA-mapping techniques performed. An F1-score of 99.83% and an average accuracy of 96.08% were recorded. Eid et al. [39] presented a novel method called BER-LSTM for precisely predicting monkeypox cases based on an LSTM. The Al-Biruni Earth Radius (BER) optimization algorithm was used to adjust the network hyperparameters. Six different ML models were used in the experiments that were carried out to demonstrate the superiority of their suggested approach. Four additional optimization algorithms are taken into account for comparison. The comparison’s findings supported the superiority of the suggested strategy. Additionally, different statistical tests (i.e., regression, Wilcoxon and one-way Analysis of Variance (ANOVA)) were used to assess the stability and importance of the suggested strategy. The outcomes of these tests highlighted the suggested approach’s robustness, importance and effectiveness. Two strategies were suggested in Abdelhamid et al. [40] to increase the categorization accuracy of photos of monkeypox. The transfer learning (i.e., GoogleNet) and meta-heuristic optimization (i.e., the sine cosine, Al-Biruni Earth radius and the particle swarm optimization algorithms) were utilized for feature extraction and feature selection and optimization of the parameters, respectively. A publicly accessible dataset was utilized to assess the proposed algorithms. Ten evaluation criteria were used to analyze the proposed optimization of feature selection for monkeypox classification. A series of statistical tests were also run to assess the suggested algorithms’ efficiency, importance and robustness. The obtained findings proved the superiority of their approach and efficacy over alternative optimization techniques: averaging 98.8% categorization accuracy.

Thirteen different pre-trained DL models for monkeypox detection were presented in Sitaula and Shahi’s study [41]. Initially, the models were fine-tuned by adding universal custom layers. Four well-known metrics were used to examine the results (i.e., recall, precision, F1-score and accuracy). Next, the top-performing DL models were assembled after identifying which performed the best overall by using majority voting. The experiments were conducted using a publicly available dataset and an average precision, recall, F1-score and accuracy of 85.44%, 85.47%, 85.40% and 87.13%, respectively. Sahin et al. [42] presented a mobile application that uses deep learning to detect monkeypox cases. The deep network is applied to categorize the images as positive or negative for monkeypox detection. Images of skin lesions of people with monkeypox and other skin lesions were utilized in the training. A publicly available dataset and a deep transfer learning strategy were applied for this goal. It is worth mentioning that MATLAB was used to conduct training and testing procedures. The best-performing network was rebuilt and trained using TensorFlow, which was modified to run on mobile devices by turning it into a TensorFlow Lite model. The TensorFlow Lite model was then integrated into the mobile application to identify monkeypox. A CNN-LSTM-based hybrid deep learning strategy was presented in Mohbey et al. [43] to detect sentiment polarities on monkeypox tweets. The main goal was to gain a deeper understanding of the wide range of responses people have to the existence of this illness. They employed a hybrid approach that is based on CNN and LSTM to investigate people’s perceptions of monkeypox diseases. Additionally, the positive, negative and neutral polarity of a user’s tweet were all taken into account. The accuracy of the prediction models was evaluated using an architecture based on CNN and LSTM. The accuracy of the suggested model on the monkeypox twitter dataset was 94%. To evaluate the suggested approach and results in the most efficient use of time and resources, other performance metrics including recall and F1-score were used. The results were then evaluated against more conventional machine learning techniques.

3. Methodology

The current study suggests a diagnostic framework to categorize the cases as normal or with pox (e.g., monkeypox, chicken pox and measles). The major aim of the framework is to report a precise decision through the hybridization of pre-trained CNN and machine learning models. First, the pre-trained models’ hyperparameters are optimized through the Harris Hawks Optimizer (HHO) algorithm. After that, the features can be extracted using the feature extraction and reduction layers. Next, these features are classified using machine learning models. The predictions from the pre-trained model and machine learning classifiers are then processed using majority voting. Figure 2 introduces a graphical presentation of the suggested methodology.

3.1. Dataset Gathering and Preprocessing Phase

Two open datasets are employed in the current study’s investigations. The first dataset is named Monkeypox Skin Images Dataset (MSID) [44] and contains 770 images. There are four categories in this dataset: (1) normal (293 images), (2) chickenpox (107 images), (3) measles (91 images) and (4) monkeypox (279 images). Internet-based sources are used to compile all of the images. The Department of Computer Science and Engineering, Islamic University, Kushtia-7003, Bangladesh, created the full dataset. The second dataset is named Monkeypox Images Dataset [45] and contains 659 images. There are four categories in this dataset: (1) normal (215 images), (2) chickenpox (100 images), (3) measles (80 images) and (4) monkeypox (264 images).

Table 1. The used datasets summary.

Dataset	Classes #	Classes	Images #	Size of Image	Extensions	Source (Link)
Monkeypox Skin Images Dataset (MSID)	4	normal, chickenpox, measles and monkeypox	770	224 × 224	“.png”	https://www.kaggle.com/datasets/dipuiucse/monkeypoxskinimagedataset
Monkeypox Images Dataset	4	normal, chickenpox, measles and monkeypox	659	256 × 256	“.jpg”	https://www.kaggle.com/datasets/sachinkumar413/monkeypox-images-dataset

summarizes the used datasets and Figure 3 shows samples from the used datasets. It is worth mentioning that the labels are encoded using one-hot encoding (e.g., 3 will be $[0,0,1,0]$ if four classes are in the dataset).

3.2. Fine-Tuning the Pretrained Models and Optimizing the Hyperparameters

In this block, the pre-trained models are fine-tuned, while the hyperparameters are optimized using Harris Hawks Optimizer (HHO). The targets are to find (1) the most promising fine-tuned CNN model and (2) the best hyperparameter combination that reports effective metrics. For all pre-trained CNN models: (1) the “ImageNet” pre-trained weights are used as the weight initializers [46], (2) a global average pooling layer follows the pre-trained model, (3) a dropout layer follows the global average pooling layer and (4) the output layer consists of four units and “SoftMax” is used as an activation function. The utilized pre-trained CNN models in the current study are (1) VGG16, (2) VGG19, (3) MobileNet, (4) MobileNetV2 and (5) Xception.

3.2.1. Hyperparameter Optimization Steps

As mentioned before, the HHO is utilized to optimize the hyperparameters. A mapping technique is used to engage the HHO with the hyperparameters. Metaheuristic optimizers initiate the population with random values between lower and upper bounds. The mapping occurs by converting the solution cell value from the floating value to a hyperparameter. For example, when registering 14 cells in a solution, there will be 14 available hyperparameters. Then, the pre-trained CNN models are used to validate each solution and the best combination is obtained and reported. The algorithm steps can be summarized as follows:

(1)

Initiate the HHO population.

(2)

Validate each solution using the pre-trained CNN model after fine-tuning it.

(3)

Sort the reported scores and find the best of them.

(4)

Repeat steps (2) and (3) for a set of iterations.

(5)

Report the best-so-far solution at the end.

3.2.2. Population Initialization

Each solution begins at a random position in the $[0,1]$ range, where 0 is the lower bound and 1 is the upper bound. $(N,D)$ represents the population size, where N is the total number of solutions and D is the width (i.e., the total number of cells) in each solution. The lower bound and upper bound are $LB$ and $UB$, respectively, and their values are 0 and 1. Moreover, $Random$ is a random value in the range $[0,1]$, i is the solution index and j is the cell index. Equation (1) shows how to initiate each cell.

$${Population}_i^j=Rando{m}_{\in [0,1]}\times ({UB}^j-{LB}^j)+{LB}^j$$

(1)

3.2.3. Learning and Solution Validation

Each solution is mapped to a set of hyperparameters, training, validating and testing processes are performed, and a fitness score is assigned to each solution. The transfer learning method is utilized in this step due to its benefits. The main benefits include resource savings and improved effectiveness while creating new models. The models are trained and tested on the training, validation and testing subsets using an 85% split ratio. In this work, five pre-trained CNN models are applied (i.e., VGG16, VGG19, MobileNet, MobileNetV2 and Xception). The different hyperparameters used in the current investigation and their various values are presented in

Table 2. The different hyperparameters used in the current investigation and their various values.

Training Loss Function
Categorical Crossentropy	Categorical Hinge	KLDivergence	Poisson	Squared Hinge	Hinge
Weights or Parameters Optimizer
SGD	Ftrl	AdaMax	AdaGrad	RMSProp	RMSProp Centered
AdaDelta	SGD Nesterov	NAdam	Adam	Adam AMSGrad
Images Scaling Techniques
L2-Normalization	Max-Absolute	Standardization	Min-Max	Robust	No Scaler
Batch Size		$4\to 48(\mathrm{step}=4)$
Dense Layer Dropout Ratio		$[0\to 0.6]$
Freezing Ratio in the Pretrained Model		$1\to 100(\mathrm{step}=1)$
Data Augmentation Different Configurations
Apply?		Yes		No
Flip Images Horizontally		Yes		No
Flip Images Vertically
Rotate Images		$0^{\circ }\to {45}^{\circ }(\mathrm{step}=1^{\circ })$
Shift Images in Width		$0\%\to 25\%$
Shift Images in Height
Shear Images
Zoom Images
Brightness Change the Images		$[0.5\to 2.0]$

.

The performance metrics reported in this step are (1) accuracy, (2) specificity, (3) sensitivity, (4) precision, (5) balanced accuracy, (6) F1-score, (7) intersection over union (IoU) and (8) receiver operating characteristic curve (ROC). The corresponding equations are presented in Equations (2) to (7).

The input image size is set to $(128\times 128\times 3)$. The percentage of correct predictions made out of all possible predictions is known as the model’s accuracy. The precision is calculated to overcome the constraints resulting from accuracy. This shows the percentage of reliable positive predictions. Recall tries to measure the proportion of true positives that are correctly detected. The goal of specificity (i.e., sensitivity) is to measure the percentage of true negatives that are accurately detected. The F1-score is the harmonic mean of precision and recall. When working with imbalanced data (i.e., when one of the target classes appears more frequently than the other(s)), it is feasible to use balanced accuracy, which is the arithmetic mean of sensitivity and specificity. When all N solutions have been completed, they will be arranged according to fitness scores in descending order. Hence, the highest-performing solution will be sorted at the top, while the worst will be at the bottom.

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

(2)

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(3)

$$\mathrm{Sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(4)

$$\mathrm{Specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$

(5)

$$\mathrm{F}1-\mathrm{Score}=\frac{2\times \mathrm{Precision}\times \mathrm{Sensitivity}}{\mathrm{Recall}+\mathrm{Precision}}$$

(6)

$$\mathrm{Balanced}\mathrm{Accuracy}=0.5\times (\mathrm{Sensitivity}+\mathrm{Specificity})$$

(7)

3.2.4. Population Updating Using HHO

Heidari et al. presented Harris Hawks Optimization (HHO) as a population-based metaheuristic optimization algorithm [47]. This optimizer mathematically imitates the cooperative behavior of Harris hawks to investigate, hunt, surprise and chase prey (e.g., rabbits). HHO includes both the exploration and exploitation stages.

During the exploration phase, after perching on a few random spots, the hawks attentively look and explore for the targeted prey. The hawks adopt two alternative techniques, each having a probability p of 50% for selection. When $p<0.5$, the first technique is conducted where the hawks may view the prey using the positions of other hunters in the hunting swarm. In the second technique, when $p\ge 0.5$, the hawks sporadically perch on trees and scan the whole search area. Equation (8) can express these two states where $X(t+1)$ denotes the hawks’ updated location vector in the following iteration $t+1$. $X\left(t\right)$ represents the location vector of the hawks in the current iteration t. $X_r\left(t\right)$ represents the location of a random hawk. $X_{prey}\left(t\right)$ represents the location of the prey. $r_1$, $r_2$, $r_3$, $r_4$ and p represent random values $\in [0,1]$. $X_m\left(t\right)$ is the average of the positions of the hawks calculated using Equation (9) where $X_j\left(t\right)$ is the t-iteration position of hawk j and N is the total number of hawks.

The shift from exploration to exploitation is the second phase, where the hawks calculate the energy of the prey. The victim attempts to flee from the hawks’ strikes during this stage, which is halfway between exploration and exploitation. Equation (10) is used to calculate the prey’s escaping energy where E decreases as it flees. The escaping energy’s value falls inside the range $\in [-1,1]$. Outside of this range, the values of E show that the exploration phase has not yet ended. When $E>1$, exploration takes place and when $E<1$, exploitation occurs. During the exploitation stage, the calculated energy is utilized by the hawks to encircle the rabbit from various angles. The positions of hawks are mapped to the desired potential solutions and the hawk in the closest position to the prey has the best position. The two fundamental actions are present during this period (i.e., fighting hawks and escaping prey). The hawks use an assault technique known as surprise pounce on their prey. Depending on the situation, there are four ways to attack: soft besiege, soft besiege with progressive fast dives, hard besiege and hard besiege with progressive rapid dives. The likelihood that the prey will escape (r) and the prey’s fleeing energy (E) determine which of these approaches is better where r is probability $\in [0,1]$ and E$\in [-1,1]$. When both $E\ge 0.5$ and $r\ge 0.5$, the soft besiege approach is used. The hawks use a soft besiege in this situation because the prey cannot flee, as shown in Equation (11) where $r_5$ is a random number $\in [0,1]$. When $r\ge 0.5$ and $E<0.5$, the hard besiege approach is used. Equation (12) demonstrates that in this situation, the prey is so exhausted that the hawks do not need to use any energy to grab them. When $r<0.5$ and $E\ge 0.5$, the soft besiege with progressive rapid dives approach is used. In this approach, the victim still has the energy to flee; thus, the hawks employ a gentle besiege. The hawks select the most appropriate path to attack the prey in this circumstance. They weigh the implications of their potential next move toward the prey. They update their location using Equation (14). Otherwise, they use the Levy flight (LF) technique to approach the prey in quick dives based on Equation (15), where D is the search space dimensions, S is a random vector of size $1\times D$ and Levy is the Levy flight function where $\nu$ Equation (16) is used to determine $\delta$. Equation (13), where Fitness is a fitness function, is the final equation for the position update in the scenario of a soft besiege with increasing quick dives. When $r<0.5$ and $E<0.5$, the hard besiege with progressive rapid dives strategy is used. In this approach, the hawks utilize hard besiege since the victim lacks the energy to flee. The hawks employ the same tactic as in a case of soft besiege with increasing quick dives. They try to close the distance between their specific position and the prey’s location. Equation (17) is used to update the position. $Y^{\prime }$ and $Z^{\prime }$ are determined using Equations (18) and (19) accordingly.

The learning and scoring layer operates once the updating procedure has finished. It is repeated for T iterations, where T is 10 in this work. The scoring layer then operates, delivering the best values for the hyperparameters along with their achieved performance metrics.

$$X(t+1)=\left\{\begin{array}{cc}{X}_r\left(t\right)-r_1\times X_r\left(t\right)-2\times r_2\times X\left(t\right)\hfill & ,\mathrm{if}p\ge 0.5.\hfill \\ \left(X_{prey}\left(t\right)-X_m\left(t\right)\right)-r_3\times \left(UB+r_4\times \left(UB-LB\right)\right)\hfill & ,\mathrm{Otherwise}.\hfill \end{array}\right.$$

(8)

$$X_m\left(t\right)=\frac{1}{N}\times \sum _{j=1}^N\left(X_j\left(t\right)\right)$$

(9)

$$E=2\times E_{e0}\times (1-\frac{t}{T})$$

(10)

$$X(t+1)=\Delta X\left(t\right)-E\times \left((2\times (1-r_5))\times X_{prey}\left(t\right)-X\left(t\right)\right)$$

(11)

$$X(t+1)=X_{prey}\left(t\right)-E\times \mathrm{\Delta }X\left(t\right)$$

(12)

$$X(t+1)=\left\{\begin{array}{cc}Y,\hfill & \mathrm{if}\mathrm{Fitness}\left(Y\right)<\mathrm{Fitness}\left(X\right(t\left)\right).\hfill \\ Z,\hfill & \mathrm{if}\mathrm{Fitness}\left(Z\right)<\mathrm{Fitness}\left(X\right(t\left)\right).\hfill \end{array}\right.$$

(13)

$$Y=X_{prey}\left(t\right)-E\times \left((2\times (1-r_5))\times X_{prey}\left(t\right)-X\left(t\right)\right)$$

(14)

$$Z=Y+S\times \mathrm{Levy}\left(D\right)=Y+S\times \left(0.01\times \frac{\nu \times \delta }{{\iota }^{\frac{1}{\beta }}}\right)$$

(15)

$$\delta ={\left(\frac{\mathrm{\Gamma }(1+\beta )\times sin(0.5\times \pi \times \beta )}{\mathrm{\Gamma }(0.5\times (1+\beta ))\times \beta \times 2^{\left(\frac{\beta -1}{2}\right)}}\right)}^{\left(\frac{1}{\beta }\right)}$$

(16)

$$X(t+1)=\left\{\begin{array}{cc}{Y}^{\prime },\hfill & \mathrm{if}\mathrm{Fitness}\left(Y^{\prime }\right)<\mathrm{Fitness}\left(X\left(t\right)\right).\hfill \\ Z^{\prime },\hfill & \mathrm{if}\mathrm{Fitness}\left(Z^{\prime }\right)<\mathrm{Fitness}\left(X\left(t\right)\right).\hfill \end{array}\right.$$

(17)

$$Y^{\prime }=X_{prey}\left(t\right)-E\times \left((2\times (1-r_5))\times X_{prey}\left(t\right)-X_m\left(t\right)\right)$$

(18)

$$Z^{\prime }=Y^{\prime }+S\times \mathrm{Levy}\left(D\right)$$

(19)

3.3. Classifying the Extracted Features Using Machine Learning

After fine-tuning the pre-trained models, the best models are used. Next, the models are loaded and the feature extraction and reduction layers are used to extract the features from the images. After that, these features are used for classification using machine learning algorithms. In this study, seven machine learning classifiers are used. They are Random Forest (RF), AdaBoost, Histogram Gradient Boosting (HGB), Gradient Boosting (GB), Support Vector Machine (SVM), Extra Trees (ET) and K-Nearest Neighbors (KNN). The selected hyperparameters for these algorithms are defined in

Table 3. The selected hyperparameters of the used machine learning algorithms.

RF	AdaBoost	HGB	GB	SVM	ET	KNN
→ C = Gini	→ # of Est. = 100	→ Loss = Log Loss	→ Loss = Log Loss	→ Reg. = 1	→ C = Gini	→ Weights = Uniform
→ # of Est. = 100	→ LR = 1	→ LR = 0.1	→ LR = 0.1	→ Kernel = RBF	→ S = Random	→ # of Nb. = 5

C: Criterion, Est.: Estimators, LR: Learning Rate, Reg.: Regularization, S: Splitter and Nb.: Neighbours.

. Finally, each classifier is trained and tested on the features using 10-fold cross-validation and then the weighted average performance metrics are reported.

3.4. Majority Voting and Predictions

In majority voting (i.e., hard voting), every classifier votes for a class and the majority wins. Usually, it is a method used to improve the performance to conduct better performance than any single model used in the ensemble. In this work, majority voting is applied between the best pre-trained CNN model and top-q ML classifiers where q is an even number to generate k predictions where k is an odd number equal to $q+1$. k is set to 3, 5 and 7 in the current study to be compared and find the best k.

4. Experiments and Discussions

In this study, the experiments are conducted using two datasets. For each dataset, the following items are discussed: (1) fine-tuning and hyperparameter optimization, (2) feature extraction and reduction, (3) optimization of machine learning classifiers and (4) the majority voting process. The first subsection handles the Monkeypox Skin Images Dataset (MSID)while the second subsection handles the Monkeypox Images Dataset (MPID). The programming language used is python and the working environment is Google Colab.

4.1. Monkeypox Skin Images Dataset (MSID) Experiments

The experiments relating to hyperparameters and parameter optimization are presented at the beginning of this subsection. At the same time, the rest of this subsection discusses applying the top-q selection and majority voting approach. While pre-trained models based on transfer learning are used to optimize the parameters (i.e., weights), HHO is used to optimize the hyperparameters. The image size is set to $(128\times 128\times 3)$ and the number of epochs is set to 7. 85% is the train-to-test split ratio. Ten is chosen as the population size and the number of HHO iterations. The first pre-trained weights are taken from the “ImageNet” database. Because this is a multi-class classification issue (i.e., as there are four classes in the MSID dataset), the output activation function is set to “SoftMax”.

Table 4. The reported results concerning the “Monkeypox Skin Images Dataset” (MSID) after fine-tuning and hyperparameter optimization using HHO.

Model	VGG19	VGG16	Xception	MobileNet	MobileNetV2
The Optimized Hyperparameters using HHO
TLLR	60%	14%	59%	0%	3%
Loss	Poisson	Categorical Crossentropy	KLDivergence	KLDivergence	KLDivergence
BS	40	8	16	12	36
Dropout	40%	5%	33%	21%	53%
Optimizer	SGD Nesterov	NAdam	NAdam	NAdam	AdaMax
Scaler	Robust	Max. Normalization	Min-Max	Min-Max	Max. Normalization
DA Appliance?	Yes	Yes	Yes	Yes	Yes
RR	19	5	21	35	9
WSR	0.1	0.03	0.05	0.09	0.25
HSR	0.18	0.02	0.15	0.03	0.18
SR	0.06	0.04	0.2	0.17	0.14
ZR	0.01	0.02	0.07	0	0.17
HF	No	Yes	Yes	Yes	Yes
VR	No	Yes	Yes	No	Yes
BR	0.59–1.01	0.67–0.7	1.89–1.98	0.5–1.18	0.62–1.03
Confusion Metrics
TP	728	710	705	702	714
TN	2254	2270	2250	2264	2229
FP	26	34	54	40	39
FN	32	58	63	66	42
Losses and Performance Metrics
Loss	0.291	0.252	0.275	0.237 √	0.378
ACC	98.09% √	97.01%	96.19%	96.55%	97.32%
Sensitivity	95.79% √	92.45%	91.80%	91.41%	94.44%
Specificity	98.86% √	98.52%	97.66%	98.26%	98.28%
PPV	96.55% √	95.43%	92.89%	94.61%	94.82%
BAC	97.32% √	95.49%	94.73%	94.84%	96.36%
F1	96.17% √	93.92%	92.34%	92.98%	94.63%
IoU	92.62% √	88.53%	85.77%	86.88%	89.81%
ROC	97.34% √	95.53%	94.77%	94.90%	96.38%

TLLR: Transfer Learning Learn Ratio, BS: Batch Size, DA: Data Augmentation, RR: Rotation Range, WSR:Width Shift Range, HSR: Height Shift Range, SR: Shear Range, ZR: Zoom Range, HF: Horizontal Flip, VR: Vertical Flip, BR: Brightness Range, TP: True Positive, TN: True Negative, FP: False Positive, FN: False Negative, ACC: Accuracy, PPV: Precision, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union, ROC: Receiver Operating Characteristic Curve and √: Best Loss and Performance Metrics.

presents the reported results concerning the MSID dataset after the fine-tuning and hyperparameter optimization using HHO. From it, the best ACC, sensitivity, specificity, PPV, BAC, F1, IoU and ROC are 98.09%, 95.79%, 98.86%, 96.55%, 97.32%, 96.17%, 92.62% and 97.34%, respectively, as the VGG19 model reports. Applying data augmentation (DA) is recommended in the five models, where four of them applied horizontal flipping and only three applied vertical flipping. Three models suggest the NAdam optimizer and KLDivergence training loss function. From that, the VGG19 is selected to continue the rest of the steps. The best model is loaded and the VGG19’s feature extraction and reduction layers are extracted. The VGG19’s optimized hyperparameters are set and the original dataset is scaled using the robust scaler. The predictions are obtained as they will be used later in the majority voting process. The size of the extracted features is $(4\times 4\times 512)$ and they are flattened to be used by the machine learning classifiers; hence, the size became 8192.

Table 5. The reported results concerning the “Monkeypox Skin Images Dataset” (MSID) after optimizing three machine learning classifiers on the extracted and processed features beside the results of the VGG19 model.

Model	RF	AdaBoost	HGB	GB	SVM	ET	KNN	VGG19
ACC	96.69%	92.50%	96.70%	96.38%	97.64%	88.67%	95.75%	97.25%
Sensitivity	93.38%	84.42%	93.64%	92.60%	95.06%	78.18%	91.43%	94.29%
Specificity	96.82%	95.47%	96.98%	96.95%	97.84%	91.92%	96.73%	97.91%
Precision	93.29%	86.61%	93.59%	92.45%	94.98%	78.36%	91.34%	94.25%
BAC	95.10%	89.94%	95.31%	94.77%	96.45%	85.05%	94.08%	96.10%
F1	93.13%	85.20%	93.48%	92.48%	94.96%	78.21%	91.31%	94.23%
IoU	87.60%	75.45%	88.04%	86.45%	90.69%	65.55%	84.62%	89.44%
ROC	95.27%	90.14%	95.43%	94.89%	96.53%	85.60%	94.23%	96.17%

ACC: Accuracy, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union and ROC: Receiver Operating Characteristic Curve.

presents the reported results concerning the MSID dataset after optimizing the machine learning classifiers using the extracted and evaluated features along with the results of the VGG19 model. The results of the VGG19 model are from the evaluation of the original images (i.e., no augmentation is applied).

Table 6. The reported results concerning the “Monkeypox Skin Images Dataset” (MSID) from VGG19 and after applying the majority voting using $k\in [3,5,7]$.

Approach	VGG19 Only	Majority (Top-3)	Majority (Top-5)	Majority (Top-7)
ACC	97.25%	97.67%	97.35%	97.44%
Sensitivity	94.29%	95.19%	94.68%	94.81%
Specificity	97.91%	97.96%	97.54%	97.67%
Precision	94.25%	95.11%	94.61%	94.72%
BAC	96.10%	96.58%	96.11%	96.24%
F1	94.23%	95.10%	94.54%	94.67%
IoU	89.44%	90.93%	89.92%	90.19%
ROC	96.17%	96.65%	96.20%	96.33%

ACC: Accuracy, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union and ROC: Receiver Operating Characteristic Curve.

presents the reported results concerning the MSID dataset from VGG19 after applying the majority voting. All values of k are better than the VGG19-only approach. However, when $k=3$, it reported the highest metrics. By utilizing it, the performance metrics improved from 97.25%, 94.29%, 97.91%, 94.25%, 96.10%, 94.23%, 89.44% and 96.17% to 97.67%, 95.19%, 97.96%, 95.11%, 96.58%, 95.10%, 90.93% and 96.65% concerning the ACC, sensitivity, specificity, PPV, BAC, F1, IoU and ROC metrics. Figure 4 shows a graphical comparison between the results.

4.2. Monkeypox Images Dataset (MPID) Experiments

The experiments relating to hyperparameters and parameter optimization are presented at the beginning of this subsection. At the same time, the process of applying the top-q selection and majority voting process is discussed later. While pre-trained models based on transfer learning are used to optimize the parameters (i.e., weights), HHO is used to optimize the hyperparameters. The image size is set to $(128\times 128\times 3)$ and the number of epochs is set to 7.85% as the train-to-test split ratio. Ten is chosen as the population size and the number of HHO iterations. The first pre-trained weights are taken from the “ImageNet” database. Because this is a multi-class classification issue (i.e., there are four classes in the MPID dataset), the output activation function is set to “SoftMax”.

Table 7. The reported results concerning the “Monkeypox Images Dataset (MPID)” after fine-tuning and hyperparameter optimization using HHO.

Model	VGG19	VGG16	Xception	MobileNet	MobileNetV2
The Optimized Hyperparameters using HHO
TLLR	37%	29%	26%	0%	48%
Loss	Poisson	KLDivergence	Squared Hinge	Categorical Crossentropy	Categorical Crossentropy
BS	8	28	12	4	8
Dropout	23%	19%	9%	0%	51%
Optimizer	AdaGrad	AdaGrad	AdaMax	Adam	SGD
Scaler	Robust	Robust	Min-Max	None	Standardization
DA Appliance?	Yes	Yes	No	Yes	No
RR	0	37	29	0	33
WSR	0.02	0.05	0.13	0	0.11
HSR	0	0.16	0.03	0	0.1
SR	0.02	0.25	0.05	0	0.04
ZR	0.2	0.2	0.16	0	0.07
HF	Yes	No	Yes	Yes	Yes
VR	No	Yes	Yes	Yes	Yes
BR	0.5–0.66	0.69–0.96	0.72–1.4	0.5–0.5	1.31–1.4
Confusion Metrics
TP	619	613	574	548	544
TN	1939	1905	1894	1874	1857
FP	29	27	50	94	111
FN	37	31	74	108	112
Losses and Performance Metrics
Loss	0.302	0.166 √	0.851	0.547	1.690
ACC	97.48%	97.75% √	95.22%	92.30%	91.50%
Sensitivity	94.36%	95.19% √	88.58%	83.54%	82.93%
Specificity	98.53%	98.60% √	97.43%	95.22%	94.36%
PPV	95.52%	95.78% √	91.99%	85.36%	83.05%
BAC	96.44%	96.89% √	93.00%	89.38%	88.64%
F1	94.94%	95.48% √	90.25%	84.44%	82.99%
IoU	90.36%	91.36% √	82.23%	73.07%	70.93%
ROC	96.47%	96.91% √	93.11%	89.57%	88.83%

TLLR: Transfer Learning Learn Ratio, BS: Batch Size, DA: Data Augmentation, RR: Rotation Range, WSR:Width Shift Range, HSR: Height Shift Range, SR: Shear Range, ZR: Zoom Range, HF: Horizontal Flip, VR: Vertical Flip, BR: Brightness Range, TP: True Positive, TN: True Negative, FP: False Positive, FN: False Negative, ACC: Accuracy, PPV: Precision, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union, ROC: Receiver Operating Characteristic Curve and √: Best Loss and Performance Metrics.

presents the reported results concerning the MPID dataset after the fine-tuning and hyperparameter optimization using HHO. From it, the best ACC, sensitivity, specificity, PPV, BAC, F1, IoU and ROC are 97.75%, 95.19%, 98.60%, 95.78%, 96.89%, 95.48%, 91.36% and 96.91%, respectively, as reported by the VGG16 fine-tuned CNN model. Applying data augmentation (DA) is recommended by three models and four recommended applying horizontal and vertical flipping. From that, the VGG16 is selected to continue the rest of the steps. The best model is loaded and the VGG16’s feature extraction and reduction layers are extracted. Then, the VGG16’s optimized hyperparameters are set and the original dataset is scaled using the Robust scaler. By evaluating the original dataset, without DA, the TP, TN, FP and FN are 622, 1947, 30 and 37, respectively. The predictions are obtained as they will be used later in the majority voting process. The size of the extracted features is $(4\times 4\times 512)$ and they are flattened to be used with the machine learning classifiers; hence, the size became 8192.

Table 8. The reported results concerning the “Monkeypox Images Dataset (MPID)” after optimizing three machine learning classifiers on the extracted and processed features beside the results of the VGG16 model.

Model	RF	AdaBoost	HGB	GB	SVM	ET	KNN	VGG16
ACC	96.39%	92.16%	96.92%	96.13%	97.36%	88.16%	96.27%	97.47%
Sensitivity	92.56%	84.07%	93.78%	91.96%	94.54%	76.93%	92.56%	94.84%
Specificity	97.10%	95.75%	97.63%	97.00%	98.23%	91.32%	97.42%	98.28%
Precision	92.50%	87.25%	93.74%	91.89%	94.59%	76.97%	92.60%	94.87%
BAC	94.83%	89.91%	95.70%	94.48%	96.39%	84.13%	94.99%	96.56%
F1	92.47%	84.90%	93.75%	91.90%	94.55%	76.87%	92.54%	94.85%
IoU	86.35%	74.49%	88.47%	85.47%	89.88%	63.68%	86.46%	90.40%
ROC	94.94%	90.15%	95.75%	94.57%	96.42%	84.74%	95.07%	96.59%

ACC: Accuracy, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union and ROC: Receiver Operating Characteristic Curve.

presents the reported results concerning the MPID dataset after optimizing the machine learning classifiers using the extracted features along with the results of the VGG16 model. The results of the VGG16 model are from the evaluation of the original images (i.e., no augmentation is applied).

Table 9. The reported results concerning the “Monkeypox Images Dataset (MPID)” from VGG16 and after applying majority voting using $k\in [3,5,7]$.

Approach	VGG16 Only	Majority (Top-3)	Majority (Top-5)	Majority (Top-7)
ACC	97.47%	97.51%	97.34%	97.28%
Sensitivity	94.84%	94.84%	94.69%	94.54%
Specificity	98.28%	98.48%	98.17%	98.10%
Precision	94.87%	94.96%	94.71%	94.56%
BAC	96.56%	96.66%	96.43%	96.32%
F1	94.85%	94.88%	94.69%	94.55%
IoU	90.40%	90.45%	90.12%	89.83%
ROC	96.59%	96.69%	96.46%	96.35%

ACC: Accuracy, BAC: Balanced Accuracy, F1: F1 Score, IoU: Intersection over Union and ROC: Receiver Operating Characteristic Curve.

presents the reported results concerning the MPID dataset from VGG16 after applying the majority voting. By applying majority voting using $k=3$, the performance metrics improved from 97.47%, 98.28%, 94.87%, 96.56%, 94.85%, 90.40% and 96.59% to 97.51%, 94.48%, 94.96%, 96.66%, 94.88%, 90.45% and 96.69% concerning the ACC, specificity, PPV, BAC, F1, IoU and ROC metrics. The sensitivity for both is equal at 94.84%. Figure 5 shows a graphical comparison between the results.

4.3. Related Studies Comparison

As mentioned before, the “Monkeypox2022” [48] dataset was introduced by Ahsan et al. [13] and is available for free download from their GitHub repository. The dataset was created by compiling images from various open-source and Internet sources that do not impose restrictions on use, even for commercial reasons. The “Monkeypox Skin Lesion Dataset (MSLD)” [49], which contains images of skin lesions brought on by measles, chickenpox and monkeypox, was developed by Ali et al. [29]. The vast bulk of the images originates from public case reports, blogs and news websites. The largest dataset of its sort up to that date, the “Monkeypox Skin Image Dataset 2022,” was revealed by Islam et al. [30].

Table 10. Comparison between the suggested approach and related studies.

Study	Datasets	Machine Learning	Deep Learning	Metaheuristic Optimizer	Majority Voting	Post-Image Analysis	Performance Metrics
Ahsan et al. [13]	Monkeypox2022 [48]		√			√	97 ± 1.8% ($AUC=97.2$) and 88 ± 0.8% ($AUC=0.867$)
Ali et al. [29]	Monkeypox Skin Lesion Dataset (MSLD) [49]		√		√		82.96% (±4.57%), 81.48% (±6.87%) and 79.26% (±1.05%)
Islam et al. [30]	Monkeypox Skin Image Dataset 2022		√		√		Precision of 85%
Sitaula and Shahi [41]	Monkeypox2022 [48]		√		√	√	Precision of 85.44%, Recall of 85.47%, F1-score of 85.40% and Accuracy of 87.13%
Alakus and Baykara [38]	Their own DNA sequences of MPV and HPV dataset		√				An F1-score of 99.83% and an average accuracy of 96.08%
Eid et al. [39]	Monkeypox Dataset (Daily Updated) [50]		√	√		√	MSE, RMSE, MAE, R2, RRMSE, r, MBE and NSE values of 480.53, 20.82, 15.25, 0.73, 1.36, 0.83, 0.06 and 0.61, respectively
Abdelhamid et al. [40]	Monkeypox Skin Images Dataset (MSID) [51]		√	√		√	An average classification accuracy was 98.8%
Sahin et al. [42]	Monkeypox Skin Images Dataset (MSID) [51]		√				Accuracy of 91.11%
Current Study	Monkeypox Skin Images Dataset (MSID) [44]	√	√	√	√		97.67%, 95.19%, 97.96%, 95.11%, 96.58%, 95.10%, 90.93% and 96.65% concerning the ACC, Sensitivity, Specificity, PPV, BAC, F1, IoU and ROC metrics
	Monkeypox Images Dataset [45]	√	√	√	√		97.51%, 94.84%, 94.48%, 94.96%, 96.66%, 94.88%, 90.45% and 96.69% concerning the ACC, Sensitivity, Specificity, PPV, BAC, F1, IoU and ROC metrics

compares the related studies and the suggested approach.

5. Limitations, Conclusions and Future Work

This study proposes a deep learning-based diagnostic framework to diagnose monkeypox. However, the study has several limitations. The used datasets were mainly collected from several open sources rather than clinical facilities or hospitals. Additionally, these datasets had limited samples and only three cases (i.e., monkeypox, chickenpox and measles) were included. The current number of unique patients in the dataset is limited. Thus, the generalization capability of the proposed approach is reduced. Only five CNN architectures and seven machine learning models are used. The lack of related work affects the possibility of conducting a fair comparison to assess the performance of the proposed approach.

Monkeypox diagnosis in an early stage is difficult because of the similarities between it, chickenpox, cowpox and measles. Thus, computer-assisted detection of monkeypox lesions can be helpful for the fast identification of suspected cases. This research uses a hybrid approach for classifying different skin images into their corresponding category with the help of pre-trained CNN models and machine learning classifiers. The images are categorized into four categories (i.e., normal, monkeypox, chicken pox and measles). First, the hyperparameters of the five pre-trained models (i.e., VGG19, VGG16, Xception, MobileNet and MobileNetV2) are optimized using a Harris Hawks Optimizer (HHO) metaheuristic algorithm. After that, the features extracted from the feature extraction and reduction layers are classified using seven machine learning models (i.e., Random Forest, AdaBoost, Histogram Gradient Boosting, Gradient Boosting, Support Vector Machine, Extra Trees and KNN). For each classifier, 10-fold cross-validation is used and the weighted average performance metrics are reported. The predictions from the pre-trained model and machine learning classifiers are then processed using majority voting. The experiments were conducted on two publicly available datasets (i.e., Monkeypox Skin Images Dataset and Monkeypox Images Dataset). To evaluate the effectiveness of the suggested approach, eight performance metrics are taken into account. For the Monkeypox Skin Images Dataset (MSID) dataset, the VGG19 CNN model performed best. After applying the majority voting, values of 97.67%, 95.19%, 97.96%, 95.11%, 96.58%, 95.10%, 90.93% and 96.65% are achieved concerning the accuracy, sensitivity, specificity, PPV, BAC, F1, IoU and ROC, respectively. While for the Monkeypox Images Dataset (MPID), the VGG16 fine-tuned CNN model had the best performance. After applying the majority voting, values of 97.51%, 94.84%, 94.48%, 94.96%, 96.66%, 94.88%, 90.45% and 96.69% are achieved concerning accuracy, sensitivity, specificity, PPV, BAC, F1, IoU and ROC, respectively. The reported values for these metrics demonstrate the effectiveness of the suggested strategy. Finally, to demonstrate the superiority of the suggested procedure, the outcomes are compared to previous works. The comparison showed that the proposed approach outperformed the prior works.

Due to the lack of suitable datasets in this area, obtaining acceptable performance was a big challenge. In future works, the plan is to (1) expand the research by experimenting with different hybrid algorithms, (2) apply the proposed approach with forthcoming datasets, (3) consider other metaheuristic optimization techniques to find the best optimizer and (4) design lightweight models to be suitable to work on limited resources.