Ant Colony and Whale Optimization Algorithms Aided by Neural Networks for Optimum Skin Lesion Diagnosis: A Thorough Review

Abstract

The adoption of deep learning (DL) and machine learning (ML) has surged in recent years because of their imperative practicalities in different disciplines. Among these feasible workabilities are the noteworthy contributions of ML and DL, especially ant colony optimization (ACO) and whale optimization algorithm (WOA) ameliorated with neural networks (NNs) to identify specific categories of skin lesion disorders (SLD) precisely, supporting even high-experienced healthcare providers (HCPs) in performing flexible medical diagnoses, since historical patient databases would not necessarily help diagnose other patient situations. Unfortunately, there is a shortage of rich investigations respecting the contributory influences of ACO and WOA in the SLD classification, owing to the recent adoption of ML and DL in the medical field. Accordingly, a comprehensive review is conducted to shed light on relevant ACO and WOA functionalities for enhanced SLD identification. It is hoped, relying on the overview findings, that clinical practitioners and low-experienced or talented HCPs could benefit in categorizing the most proper therapeutical procedures for their patients by referring to a collection of abundant practicalities of those two models in the medical context, particularly (a) time, cost, and effort savings, and (b) upgraded accuracy, reliability, and performance compared with manual medical inspection mechanisms that repeatedly fail to correctly diagnose all patients.

1. Introduction

The utilization of artificial intelligence (AI), machine learning (ML), neural networks (NNs), deep learning (DL), and other practical numerical schemes has recently increased owing to the diversified contributions and robust identification tasks they can provide in numerous disciplines and scientific areas, such as medicine, biology, computer science, and industrial engineering. Complex mathematical problems that are challenging to resolve and manage can be flexibly controlled and addressed with remarkable performance and accuracy with those feasible numerical models. The critical advantages of AI, DL, and ML algorithms can be summarized as follows [1]:

(A)

Promotes the potential for handling big data and the analysis of complicated algebraic problems.

(B)

Provides a flexible investigation of nonlinear correlations.

(C)

Elevates rates of detection, recognition, and classification efficiency.

In recent years, significant breakthroughs have improved the efficiency and effectiveness of managing different clinical tasks, leading to enhanced treatments and therapies for patients. Nurses, physicians, doctors, therapists, and other healthcare providers gave significant levels of confidence and reliability to AI, DL, and NNs to support them in executing fast healthcare decision-making and the precise recognition of vague data or difficult-to-determine information correlated to clinical problems and medical challenges [2,3]. For example, doctors [4,5] can utilize numerical algorithms and intelligent ML models to detect or classify unclear databases in Magnetic Resonance Imaging (MRI) [6,7] Computed Tomography (CT) scans, X-rays, and ultrasound visual data related to skin lesion disorders (SLDs). These conditions may be challenging for less-experienced therapists to detect accurately. Simultaneously, some identification obstacles in this clinical area could be treated by ML and AI models, as reflected in overcoming expensive and time-consuming aspects, low nurse skills, and erroneous and faulty diagnoses in cases where conventional medical recognition methods are implemented [8,9]. Identifying SLDs is not considered an easy task, since new patient data of medical problems do not necessarily correlate to other historical patient databases because of diverse ailment characteristics and iatric situations. Simultaneously, obstacles pertaining to the identification of multiple disorders in this clinical context could be handled by ML and DL models because of their remarkable workability in classifying precisely the kind of ailment or the problem in visual databases in spite of numerous historical patient data that need costly and time-consuming labeling. However, even if all data are annotated, precise identification of the SLD pertaining to the patient is considered not an easy task. For this reason, erroneous medical prescriptions, from faulty recognition and incorrect diagnoses, can be prevented when ML and DL models are considered.

Various scholars [10,11] have addressed a series of practical significances of ML and DL models in supporting high-experienced healthcare providers (HCPs) in conducting accurate classification takes of SLDs, as shown in Figure 1.

Healthcare providers encounter several challenges when it comes to accurately detecting and classifying skin lesions. These challenges include subjectivity in human assessment, time constraints, limited access to expertise, data availability, and the complexity of lesions. However, there are promising technological solutions that can help overcome these challenges. Algorithms like Ant Colony Optimization (ACO), Whale Optimization Algorithm (WOA), and Neural Networks (NNs) provide effective methods for optimizing parameters and enhancing accuracy and efficiency in lesion classification. NNs, specifically deep learning models, can adjust to variations in lesion characteristics and generalize learning to new images. Integrating these techniques offers several benefits, such as automating the diagnostic process, reducing the time and effort required by healthcare providers, optimizing resource utilization, and providing consistent and reliable classification across different settings. Numerous studies have demonstrated the effectiveness of ACO and its variants in enhancing the efficiency, accuracy, and effectiveness of skin lesion categorization.

These algorithms have been successful in identifying various skin lesion disorders and defects in medical images. The numerical results from these studies show high classification accuracy percentages, with some reaching approximately 95.9%. ACO-optimized edge-detection methods outperform other optimization algorithms, demonstrating superior performance in skin lesion classification.

The contributions of this review can be summarized as follows:

• The study conducts a comprehensive review of the recent literature to highlight the use of machine learning (ML) and deep learning (DL) techniques in skin cancer detection in combination with optimization algorithms for these techniques, which include the ant colony (ACO) and whale optimization (WOA) algorithms.
• The study presents a comparative analysis of ML and DL methods, including ACO, WOA, and NNs, for the classification of skin lesions.
• The study lists several limitations of these techniques, and demonstrates the importance of improving these algorithms, especially in the field of skin diseases. This will lead to accurate diagnosis, assist dermatologists in decision-making, and improve patient outcomes by facilitating early detection and treatment of skin lesions.
• The study provides recommendations and proposals to increase the level of performance of detection and classification models in this field by integrating the two algorithms, (ACO) and (WOA), with NNS.

Overall, this study contributes to advancing the field of skin lesion detection and classification by leveraging modern AI techniques, and provides insights into their effectiveness and applicability in clinical practice. Within this framework of SLD background, it is essential to explain, for the benefit of scholars, the sequence of this paper to facilitate a better understanding of the thorough review. The layout of this paper was developed and organized according to the following sequence:

• Section 2 explains the main research method adopted to execute the thorough overview. It clarifies the considered criteria of the secondary data collection procedure to cover essential ideas and prominent aspects of ACO and WOA in identifying different types of SLDs precisely when supported with NNs.
• Section 3 illustrates the involvement of ACO and WOA in recognizing various SLDs accurately, shedding light on variant breakthroughs and state of the arts that corresponded to more elaborations on the performance of ACO and WOA in conducting their identification tasks of SLDs.
• Section 4 indicates an overview of numerous outputs correlated with the utilization of ACO, WOA, or both in carrying out the necessary classification of SLDs from different research articles.
• Section 5 explores how healthcare providers can overcome challenges in skin lesion detection by leveraging advanced techniques such as ACO, WOA, and NNs. We will discuss the importance of accurate diagnosis and classification, particularly in cases where early detection is crucial for improved patient outcomes. Additionally, we will examine how traditional methods can be complemented or replaced by these intelligent algorithms to enhance accuracy and efficiency in skin lesion detection.
• Section 6 classifies the research’s main future work and recommendation opportunities that can offer areas for improving the robustness of this work.
• Section 7 presents concise conclusions drawn from our study. We summarize key findings regarding the application of AI, DL, ML, NNs, ACO, and WOA in skin lesion detection. Our conclusions highlight the effectiveness of integrating these techniques for improving diagnostic accuracy and patient outcomes in dermatology. We also identify potential avenues for future research to further advance the field.
• Finally, Section 8 shows the major research limitations encountered in completing this overview. Finally, all abbreviations are presented in Appendix A.

2. Materials and Methods

To fulfill the primary research objective of categorizing the major significance and crucial applications in medicine, science, and bioengineering related to AI, DL, and ML models guiding influential skin lesion classification tasks with NN support, this study considers a couple of research phases through which a comprehensive review process is conducted. This study used a secondary data-collection method that relied on scholarly articles and other recent publications to investigate the benefits and additional value of applying the principles of AI, DL, and ML to the problem of recognizing skin lesions in medical images.

Simultaneously, the research foundation adheres to specific criteria ensuring a comprehensive overview, leading to improved outcomes for this review task. First, the analysis and survey of several recent research publications are up to date as of 2016. This aspect enables the active inference of modern benefits and outcomes. In this extensive review, most of the surveyed articles are linked to the medical and scientific applications of AI, DL, and ML. The outcomes of these applications are examined, considering NNs in conjunction with ACO and WOA.

Validation of the Reliability, Contributions, and Efficacy of the Manuscript’s Outputs

The critical appraisal process likely involved several steps to ensure the accuracy and validity of the findings. Here are some potential methods mentioned in the section:

• Expert Judgment: The manuscript may have undergone evaluation by experts in the fields of AI, DL, ML, and dermatology. These experts likely provided feedback and insights regarding the relevance, accuracy, and significance of the study’s outputs.
• Peer Review: The manuscript might have been subjected to a peer review process, where independent experts in the field critically evaluated the study’s methodology, results, and conclusions. Peer reviewers would have assessed the study’s rigor, validity, and contribution to the existing body of knowledge.
• Professional Perspectives: Input from professionals specializing in AI, DL, ML, and dermatology could have been solicited to validate the manuscript’s outputs. Their perspectives and insights would have provided additional validation and confirmation of the study’s findings.
• Comparative Analysis: The study may have compared its findings with those of other relevant research studies to ensure consistency and accuracy. This comparative analysis would help validate the reliability and contributions of the manuscript’s outputs within the broader context of existing literature.

Overall, the validation process likely involved a thorough examination of the manuscript’s outputs through expert judgment, peer review, consultation with professionals, and comparative analysis, ensuring the reliability, contributions, and efficacy of the study’s findings.

After completing this literature review, a critical appraisal was carried out to validate and verify the reliability, contributions, and efficacy of the manuscript’s outputs with the aid of expert judgment, peer review opinions, and AI, DL, and ML professionals’ perspectives.

3. Background

In this section, rich analytical assessments and investigations of the corresponding benefits of ML and DL algorithms are carried out pertaining to their practicalities in conducting efficient SLD classification. A special concentration would be considered for addressing the roles of ACP and WOA in identifying various SLDs potentially. Also, the incorporation of NNs with those two models is taken into account in the medical context.

3.1. Brief Illustration of SLD Prevalence Universally

Skin lesion detection is a critical task in dermatology and medical imaging, aiming to accurately classify various types of skin lesions from dermoscopic images. The advent of deep learning, particularly Convolutional Neural Networks (CNNs), has revolutionized this field by enabling automated and accurate classification of skin lesions based on visual characteristics extracted from images. CNNs are adept at automatically learning hierarchical features from images, allowing them to discern patterns and features indicative of different skin lesions [2]. In the context of skin lesion detection, features could include color intensity, texture, shape, size, and other visual characteristics extracted from the images.

Despite the success of CNNs, challenges persist, including variability in lesion appearance, limited availability of labeled data, and the need for robust models capable of generalizing to diverse lesion types. Ensemble learning techniques have emerged as a powerful approach to address these challenges. Ensemble methods involve combining predictions from multiple individual models to improve overall accuracy. By aggregating diverse models, ensemble learning compensates for individual model weaknesses and enhances overall performance. This approach is particularly beneficial in scenarios with limited training data and variability in lesion characteristics [13].

Moreover, fusion-based aggregation methods have been proposed to integrate outputs from different models using various fusion strategies. These methods aim to exploit complementary information from diverse models to make more accurate predictions. By selecting the most effective fusion strategy, researchers optimize classification performance and overcome challenges associated with variability in lesion appearance and classification difficulty [14].

Early diagnosis of SLDs plays a crucial role in improving patient health outcomes and lowering death rates. Figure 2 provides a profile of age-specific levels linked to the incidence, Disability-Adjusted Life Years (DALYs), and mortality rates of malignant skin melanoma worldwide. In addition, the following are the several reports for each statistical detail and insight:

• Prevalence of Skin Lesion Disorders: In 2018, approximately 300,000 new cases of melanoma were detected globally, making it the most prevalent cancer among both men and women. Over one million new cases of Squamous Cell Carcinoma (SCC) and Basal Cell Carcinoma (BCC) were diagnosed in the same year, ranking them as the second and third most common forms of skin cancer after melanoma, respectively [7].
• Impact of Ozone Depletion and UV Exposure: The depletion of the ozone layer, which protects against harmful ultraviolet (UV) rays, has led to a considerable increase in skin cancer cases due to heightened UV exposure. Excessive exposure to harmful sunbeams, especially more than five sunburns, can double the risk of skin cancer and melanoma [7].
• Mortality Rates and Survival Rate: Skin cancer, including melanoma, is associated with significant mortality rates. Over two individuals die due to skin cancer in the USA every hour. However, the survival rate for skin cancer, especially when treated and managed during its early stages, can reach up to 90 percent [6].
• Trends in Melanoma Cases: Invasive melanoma is expected to be the fourth most recognized cancer type in both sexes in 2022, with a projected 57,180 cases in men and 42,600 occurrences in women. Melanoma cases have demonstrated an increasing trend and a higher frequency of observed cases in the USA over the past three decades, with a doubling of the rate in 2011 compared to cases recorded in 1982 [6].

According to Kassem et al. [16] and the skin care foundation (SCF) [17], the typical classifications of SLDs, especially those pertaining to skin cancer noted in different patients, are shown in Figure 3.

Additionally, reports from healthcare organizations and nonprofit skin cancer institutions indicate that invasive melanoma is expected to be the fifth most commonly recognized cancer type in both genders, with 57,180 cases in men and 42,600 occurrences in women in 2022 [18].

Moreover, certain SLDs, particularly melanoma, have demonstrated an increasing trend and a higher frequency of observed cases in the USA over the past three decades. This corresponds to a doubling rate in 2011 compared to melanoma cases recorded in 1982, with the rising trend of this skin ailment manifesting a varying number of cases and diverse skin lesion disorder situations depending on age [19].

3.2. Crucial Features and Practical Merits of ACO for the Medical Context

Simulation specialists and professional numerical analysts investigated new ML and DL algorithms such as ACO to facilitate classification and prediction tasks of different databases with considerable performance. ACO is a type of metaheuristic model characterized by its remarkable capability to analyze graphical databases that may vary and fluctuate following a dynamic behavior, overcoming limitations. Moreover, the ACO approach can prove significant potential for continuous operation and adaptation to various changes associated with real-time data analysis. These advantages have captured the interest of scholars and researchers engaged in active classification activities, graphical data investigation, and visual image detection in medical, engineering, and mathematical domains [20,21].

The working principles of ACO are correlated with an ant leaving a trail with a particular chemical substance, such as a pheromone, across its path of travel as it leaves its nest to look for food. The critical advantages of this behavior are reflected in its flexible ability to identify the pheromone trail left by the ant’s predecessor. The notation “stigmergy” describes the indirect communication process, which selects the fastest path and shortest route between the nest position and the location of the food. Every path is characterized by an initial pheromone deposition degree and pheromone evaporation level. In this context, the ACO technique can be used to identify a set of peaks and vertices from a graphical representation where contours are balanced based on adaptivity and supported by pheromone strength, which is an advantageous property in ants [20]. The working principle details of ACO can be seen in the Supplementary Materials, section 2.

Ant Colony Optimization (ACO) has been increasingly applied to various classification tasks, including skin lesion classification [2,22]. The following is how ACO can be applied in this context and the specific advantages it offers:

• Feature Selection: ACO can be used to select the most relevant features from skin lesion images. By considering the importance of different features (such as texture, color, and shape) through the exploration of feature space, ACO helps in identifying the most discriminative features for classification.
• Optimization of Classifier Parameters: ACO can optimize the parameters of machine learning or deep learning classifiers used for skin lesion classification. This optimization process helps in fine-tuning the classifier to achieve better performance in terms of accuracy and robustness [23].
• Enhanced Adaptability: ACO’s ability to adapt to dynamic and varying datasets is particularly advantageous in skin lesion classification, where images may exhibit diverse characteristics due to factors like lighting conditions, camera angles, and patient demographics. ACO helps classifiers adapt to these variations, leading to more consistent and reliable classification results [24].
• Improved Interpretability: ACO-based feature selection and parameter optimization techniques often result in more interpretable classifiers. This means that the reasoning behind the classification decisions can be better understood, which is crucial for medical applications like skin lesion diagnosis [25].
• Reduced Overfitting: ACO’s optimization process can help in preventing overfitting, a common issue in machine learning models where the model performs well on training data but fails to generalize to unseen data. By optimizing model parameters in a principled manner, ACO aids in building classifiers that generalize better to new skin lesion images [25].
• Scalability: ACO algorithms are scalable and can handle large datasets efficiently. This scalability is important in skin lesion classification, where datasets may contain thousands or even millions of images [26].
• Overall, by leveraging ACO techniques in skin lesion classification, researchers and practitioners can achieve more accurate, robust, and interpretable classification models, ultimately leading to improved diagnosis and patient outcomes. However, because of its workability in different classification situations of SLDs, various researchers, including, clarified ACO’s potential in classifying the type of skin problem, relying on visual optimization, preprocessing, similarity evaluation, augmentation, and filtering of investigated images.

3.3. Serviceable Practicalities and Efficient Features of WOA for the Medical Context

WOA can be illustrated as a swarm intelligence (SI)-based optimization technique inspired by the hunting techniques of humpback whales [27]. The WOA model utilizes three operators to simulate the humpback whale’s search process for food, encirclement of prey, and bubble-net foraging technique. In recent years, the WOA has been effectively utilized in a wide variety of optimization and feature selection scenarios. For feature selection, some scholars, such as Mafarja and Mirjalili (2017) [23], created a hybrid WOA utilizing a simulated annealing approach. In their hybrid approach, simulated annealing was harnessed to better utilize the most promising locations identified the WOA relevance through exhaustive searching. Another wrapper selection method by Mafarja and Mirjalili (2018) [28] was based on WOA, improving search effectiveness by employing competition and roulette-wheel selection tactics in addition to crossover and mutation operators. Nonetheless, based on references to its practical application in the medical scenario analysis of high-dimensional medical datasets, it was carried out by Nematzadeh et al. (2019) [29], who presented a frequency-based filter selection of features that employs the WOA. They employed WOA along with a filter criterion to eliminate extraneous features. Another filtering approach, Mutual Congestion (MC), is adopted to determine the order of the remaining features.

The Whale Optimization Algorithm (WOA) [29] has garnered significant attention in the optimization community. Since its inception, researchers have explored various extensions and hybrids to enhance their performance and applicability across different domains. For instance, [30] proposed a hybrid algorithm merging the Grey Wolf Algorithm with WOA, termed the WGC algorithm. Additionally, [28] introduced a hybrid WOA combined with simulated annealing for feature selection tasks.

In diverse fields, researchers have applied WOA variants to tackle specific optimization challenges. Wang et al. [31] developed a multi-objective version of WOA for wind-speed forecasting, while El Aziz [32] utilized WOA for multilevel thresholding image segmentation. Eid [33] devised a binary variant of WOA (bWOA) for parameter estimation in photovoltaic cells and feature selection problems.

Further advancements include Reddy K. et al. [34] employing bWOA for profit-based unit commitment problems in marketing and Hussein et al. [35] presenting binary WOA variants for feature selection problems. Despite the no-free-lunch theorem’s assertion that no single optimization algorithm is universally superior, WOA’s versatility in handling various problems, including engineering design and mathematical optimization, has made it a popular choice among researchers. The next paragraph will explain the implementation of feature selection with ACO and WOA and how it contributes to the classification of skin lesions.

Feature selection is a crucial step in the classification of skin lesions using ACO and WOA algorithms. The following sub-sections give a detailed explanation of how feature selection is implemented and its contribution to the classification process.

3.3.1. Features and Merits of WOA

• Versatility and Effectiveness: WOA has gained significant attention in the optimization community due to its effectiveness in solving diverse problems [36]. Researchers have explored extensions and hybrids to improve its performance across different domains.
• Hybrid Approaches: Hybrid versions of WOA have been proposed, combining it with other algorithms like simulated annealing or the Grey Wolf Algorithm, enhancing its capabilities for specific tasks [27,30].
• Applications Across Fields: WOA variants have been applied in various fields such as wind-speed forecasting, image segmentation, parameter estimation, and feature selection [31,32,33,35].
• Popularity Among Researchers: Despite the no-free-lunch theorem’s assertion, WOA’s versatility in handling various optimization problems has made it a popular choice among researchers [37].

3.3.2. Feature Selection and Classification

• Feature Selection Process: Feature selection involves identifying relevant features from input data essential for accurate classification [27]. ACO and WOA algorithms optimize feature selection iteratively, aiming to maximize classification accuracy.
• Contribution to Classification: Selecting discriminative features improves classification efficiency and effectiveness. ACO and WOA algorithms optimize feature selection by considering feature importance, correlation, and interaction, leading to better classification outcomes [38].

3.3.3. Mathematical Evaluation of WOA

Researchers have developed algebraic correlations and formulas to describe whale hunting and food-searching methods [39,40]. Strategies involve mimicking the leader whale’s behavior, updating positions, seeking prey, encircling prey, and executing bubble-net attacks [26]. The working principle details of WOA can be seen in the Supplementary Materials, section 3.

In summary, WOA’s features, versatility, and effectiveness make it a valuable tool for optimization tasks, including feature selection for classification purposes. Researchers have explored various extensions and mathematical evaluations to enhance their performance and applicability across different domains.

Overall, ongoing research efforts aim to enhance and adapt WOA to address a wide range of optimization challenges, underscoring its significance in the field of metaheuristic algorithms.

3.4. Engagement of ACO and WOA in SLD Identification

There were multiple efficient algorithms created and modified to conduct accurate categorization of the current clinical problem relying on the data from visual databases, like SLD, such as naïve Bayes (NB), support vector machine (SVM), decision tree (DT), and k-nearest neighbor (KNN) [41].

Nevertheless, there are other practical numerical models on which HCPs rely to identify different SLDs. The Internet of Things (IoTs), ML, DL, and other AI paradigms developed over the last few years contributed to an enhanced potential of classification tasks and remarkably accurate recognition of visual medical databases. Doctors have benefited from these innovative models in classifying benign or malignant SLDs and skin dermoscopy problems, streamlining decision-making for appropriate therapy and optimal treatment. Early identification of the existing SLD in patients can contribute to promoted healthcare outcomes through earlier therapy and rapid treatment before the SLD or cancerous skin problem might affect other skin areas [42,43].

Huang et al. [44] have made a noteworthy contribution by creating an enhanced whale optimization technique specifically designed for denoising gray images. Their work focuses on rectifying the limitations of conventional denoising algorithms, particularly in situations characterized by intense noise, thereby improving both the visual quality and the peak signal-to-noise ratio of images.

Xu et al. [45] provide a thorough study that expands the application of bio-inspired algorithms in image processing. The review covers a range of swarm intelligence optimization techniques, such as Ant Colony Optimization (ACO) and WOA. Their review thoroughly analyzes the practical uses of algorithms in image segmentation, matching, classification, feature extraction, and edge detection. It highlights the extensive possibilities of these methods in improving image processing techniques. The review comprehensively explores theoretical research and improvement methodologies while also emphasizing practical applications. This approach offers a comprehensive perspective on the present status and future prospects of swarm intelligence in image processing. These works jointly emphasize the efficacy of bio-inspired optimization algorithms in tackling intricate issues in the field of image processing. The incorporation of algorithms such as ACO and WOA into image processing workflows brings about a new era of innovation, enabling improvements in medical diagnosis accuracy and image quality enhancement through advanced denoising techniques. By harnessing the distinctive capabilities of these algorithms, researchers have created opportunities for the advancement of image-processing techniques that are more effective, precise, and resilient. This represents notable progress in the realm of artificial intelligence and its practical implementation.

To achieve its role in identifying the accurate type of SLD from patients’ visual databases, the ACO can enhance the edge contours of SLD images. The image quality of the SLD can be improved by a three-step preprocessing procedure comprising color-space conversion, contrast augmentation, and filtering. Three common edge-detection techniques, Canny, Sobel, and Prewitt, are also utilized to produce the edge map. Following this step, the ACO can be then applied to the image to help enhance the sharpness of the edge contours [46,47]. As a result, optimum identification of the SLD can be accomplished.

Various scholars validated the ACO rationale in conducting precise SLD identification. For instance, Sengupta et al. (2020) [48] conducted an analysis that investigated the critical relevance and vital rationale of the ACO model for the active process of skin lesion identification. Scholars have explained that automatic skin lesion picture recognition is crucial to creating a completely automated computer-aided skin analysis system. This will aid physicians in treating skin lesions as effectively and efficiently as possible. Their article proposed two image-processing methods for the reliable identification of skin lesions. The first method utilizes a subset of AI called nature-inspired algorithms to optimize edge detection. ACO significantly improves the accuracy of edge identification in skin lesions. The second method combines a global thresholding segmentation process with an edge-smoothing operation to perform a split-and-merge operation in the color space. The entropy performance evaluation parameter is used to compare the results of the two methods. Their numerical analysis indicates that the output images obtained using the ACO-optimized Canny edge detection method outperform those obtained using the ACO-Sobel, ACO-Prewitt, and Edge Smoothing-Color Space approaches. When compared to the ACO-Sobel, ACO-Prewitt, and Edge Smoothing-Color Space methods, ACO-Canny Edge detection was clearly the most effective method for identifying skin lesions.

3.5. New Cutting-Edge Concepts and Novel Approaches Involved in Skin Lesion Detection Process

Research and development (R&D) on numerical computer simulations continued in creating helpful innovations that could support existing AI and ML models in performing their classification tasks more efficiently and robustly. The following paragraphs provide more details regarding the critical gains and contributions of those new concepts to upgrade existing AI and ML models’ categorization accuracy. Several breakthrough technologies and novel concepts related to the categorization of medical deficiencies, specifically in skin lesion detection, are mentioned. Let us break down some of these examples and how they contribute to overcoming challenges:

• Deep Convolutional Neural Networks (D-CNNs): D-CNNs are a type of deep learning model specifically designed for image classification tasks. They excel in detecting intricate patterns and features within medical images, particularly in dermoscopy images used for skin lesion classification. By leveraging D-CNNs, researchers can achieve higher accuracy in classifying skin lesions into different categories, such as melanoma, nevus, and seborrheic keratosis. These networks contribute to overcoming challenges by providing accurate and reliable automated screening systems for early detection of malignant skin lesions, thereby assisting clinicians in timely diagnosis and treatment.
• Ensemble Learning: Ensemble learning involves combining multiple models to improve predictive performance over any single model. In the context of skin lesion detection, researchers combine the outputs of multiple deep neural network architectures to enhance classification accuracy. By aggregating the classification results of different models, ensemble learning compensates for individual model weaknesses and enhances overall performance. This approach helps address challenges related to the limited availability of training data and variability in lesion characteristics by leveraging the collective knowledge of diverse models.
• Fusion-Based Aggregation Methods: Fusion-based aggregation methods involve combining the outputs of individual neural networks using various fusion strategies. These methods aim to integrate complementary information from different models to make more accurate predictions. By selecting the most effective fusion strategy, researchers optimize classification performance and overcome challenges associated with variability in lesion appearance and classification difficulty. Fusion-based aggregation methods enhance the robustness and reliability of skin lesion classification systems, improving diagnostic accuracy and patient outcomes.

Overall, these technologies and concepts contribute to overcoming challenges in skin lesion detection by leveraging advanced deep learning techniques, ensemble learning principles, and fusion-based aggregation methods to enhance classification accuracy, reliability, and efficiency. They enable the development of more effective automated screening systems for early detection and diagnosis of skin cancer, ultimately improving patient care and outcomes. The following paragraphs provide more details regarding the critical gains and relevance of these avant-garde strategies.

3.5.1. Neural Networks (NNs)

Neural Networks (NNs) are computational models inspired by the structure and function of the human brain. In the context of skin lesion detection, deep neural networks (DNNs), particularly Deep Convolutional Neural Networks (D-CNNs), have shown remarkable success. D-CNNs excel at automatically extracting hierarchical features from images, enabling them to achieve high accuracy in classifying skin lesions [2]. However, one of the challenges in training DNNs for skin lesion detection is the requirement for large, labeled datasets. Obtaining such datasets can be challenging in medical imaging tasks. To address this challenge, researchers have explored ensemble methods that combine multiple DNN architectures to improve classification accuracy. By leveraging the collective knowledge of diverse models, ensemble learning has been shown to enhance performance, even with limited training data [49]. The details descriptions and studies of NNs can be found in the Supplementary Materials, section 4.

3.5.2. Multi-Layered Perceptron (MLP)

Multi-layered perceptron (MLP) is another type of neural network model commonly used in skin lesion detection. MLPs are organized into a hierarchical structure, with input nodes gathering input patterns from the dataset and output nodes mapping these patterns onto classifications or signals. Hidden layers in MLPs play a crucial role in extracting essential features from input data to forecast outputs [42]. In the context of skin lesion classification, researchers have explored the use of fuzzy MLP models to address the ambiguity of inputs. These models consider uncertain input data during training, improving performance by reducing the effects of equivocal inputs on the learning process. Additionally, novel optimization functions and activation functions have been proposed to enhance the performance range of fuzzy neural networks in skin lesion detection tasks [42].

3.5.3. Data Dimensionality Reduction Techniques

Data dimensionality reduction techniques aim to reduce the computational burden associated with analyzing large datasets by eliminating irrelevant or redundant information. In the context of skin lesion detection, these techniques play a crucial role in improving efficiency and scalability. By minimizing the size of the dataset without compromising the quality of classification results, data dimensionality reduction techniques enhance the performance of machine learning and deep learning models. Researchers have explored various methods for dimensionality reduction in skin lesion analysis, leading to more efficient and scalable classification systems [50]. A taxonomy of data dimensionality reduction algorithms connected to various biomedical areas of data analysis is explained in Figure 4 [51].

Apart from the positive effects of data dimensionality reduction paradigms and technologies, scholars have observed that employing this innovative procedure can yield additional advantages and enhanced quality collected with the numerical findings obtained from intelligent MKL and DL algorithms. Figure 5 illustrates some of these effects.

3.5.4. Data Augmentation

Data augmentation techniques involve generating new training samples by applying transformations to existing data. In skin lesion classification, data augmentation is used to increase the diversity of the training dataset and improve the generalization ability of machine learning models. Horizontal Flip Augmentation (HFA) is a commonly used technique in skin lesion classification, where images are horizontally flipped to create additional training samples. By augmenting the dataset with transformed images, researchers enhance the robustness of skin lesion classification models and improve overall accuracy [52]. Figure 6 shows a flip augmentation process applied to the visual dataset of SLD deficiencies to improve the effectiveness of the dermatological classification task based on the horizontal direction.

3.6. Commonly Examined Datasets Related to SLD

Various publicly available datasets containing dermoscopic images of skin lesions are used for training and evaluating skin lesion classification models. These datasets provide researchers with a diverse range of skin lesion images for analysis, enabling the benchmarking of different algorithms and techniques which include: the PH2 Dataset, MED-NODE Dataset, HAM10000 Dataset, Derm7pt Dataset, BCN20000 Dataset, and the ISIC Dataset are some examples of publicly available datasets used in skin lesion detection research. These datasets play a crucial role in advancing the development of machine learning and deep learning models for skin lesion classification [2]. The details’ data description can be found in the Supplementary Materials, section 5.

The following sub-sections are the references for each statistical detail, with insight provided.

3.6.1. Prevalence of Skin Lesion Disorders

In 2018, approximately 300,000 new cases of melanoma were detected globally, making it the most prevalent cancer among both men and women [53]. Over one million new cases of Squamous Cell Carcinoma (SCC) and Basal Cell Carcinoma (BCC) were diagnosed in the same year, ranking them as the second and third most common forms of skin cancer after melanoma, respectively.

3.6.2. Impact of Ozone Depletion and UV Exposure

The depletion of the ozone layer, which protects against harmful ultraviolet (UV) rays, has led to a considerable increase in skin cancer cases due to heightened UV exposure [54]. Excessive exposure to harmful sunbeams, especially more than five sunburns, can double the risk of skin cancer and melanoma.

3.6.3. Mortality Rates and Survival Rate

Skin cancer, including melanoma, is associated with significant mortality rates. Over two individuals die due to skin cancer in the USA every hour. However, the survival rate for skin cancer, especially when treated and managed during its early stages, can reach up to 90 percent.

3.6.4. Trends in Melanoma Cases

Invasive melanoma is expected to be the fourth most recognized cancer type in both sexes in 2022, with a projected 57,180 cases in men and 42,600 occurrences in women. Melanoma cases have demonstrated an increasing trend and a higher frequency of observed cases in the USA over the past three decades, with a doubling of the rate in 2011 compared to cases recorded in 1982.

In summary, advancements in deep learning techniques, such as ensemble learning and fusion-based aggregation methods, combined with large-scale skin lesion datasets, contribute to improved accuracy and reliability in skin lesion detection and classification. These technologies hold great promise for enhancing diagnostic accuracy and patient outcomes in the early detection and diagnosis of skin cancer, ultimately improving patient care and treatment outcomes.

Table 1. A summary of famous datasets linked to SLD explored broadly worldwide.

No.	Name of the Dataset	Overall Number of Photos	Photo Category	Number of Skin Defects	Format of Images	Year Issued	Crucial Relevance
1	MED-NODE	170	Macroscopic	Two	.jpg	2015	To identify skin cancer problems
2	HAM-10,000	10,015	Dermatological	Eight	.jpg	2018	To consider the improper variety in datasets and their small size
3	PH2	200	Dermatological	Three	.bmp	2013	To provide active segmentation and classification process of melanoma
4	Derm7pt	2000	Dermatological Structured Database	Fifteen	.jpg	2018	To actively detect skin lesion problems for a seven-point malignant analysis
5	BCN-20,000	19,424	Dermatological	Nine	.jpg	2019	To help therapists identify challenging skin defects in difficult areas, like mucous membranes and nails
6	ISIC Archive	More than 13,000	Dermatological	Nine	DICOM and .jpg	From 2016 to 2020	To boost the adoption of numerical skin lesion detection and enhance treatment efficiency and outcomes

summarizes the datasets, highlighting their distinctive characteristics. This tabulated database provides information on prominent datasets utilized in skin cancer classification. Simultaneously, scholars have been invited to discover additional categories of skin lesion datasets from various sources [55,56]. Dermoscopic images dominate the skin disease datasets, with clinical and histopathological images being less common. Nonetheless, most skin disease datasets contain a small number of images compared to datasets of natural images, posing classification challenges for skin cancer.

4. Contributary Statistical Outcomes and Medical Gains of ACO, WOA, and NNs

Comparative analyses of state-of-the-art AI, DL, and ML approaches such as ACO, WOA, and NNs have proven the superiority of intelligent algorithms and numerical strategies for the efficient classification and optimal detection of skin lesion disorders. To present a comprehensive database and supporting evidence for the valuable applicability of models (ACO, WOA, NNs, and other models) in classifying diverse issues within visual datasets related to skin lesion deficiencies, we conducted a comparative analysis. This involved a tabulated representation of the database, facilitating a streamlined review process and offering a flexible perspective on the essential contributions and relevance of ACO, WOA, NNs, and other models in categorizing various skin lesion defects with notable accuracy, robustness, and performance, as documented in the available literature [26,37,43].

This section presents significant findings from a comparative analysis of state-of-the-art AI, DL, and ML approaches, including ACO, WOA, and NNs, for classifying skin lesion disorders. Some key findings and their implications for the field of skin lesion classification and diagnosis may include:

• Superiority of Intelligent Algorithms: The study demonstrates the superiority of intelligent algorithms such as ACO (Ant Colony Optimization) and WOA (Whale Optimization Algorithm) over traditional methods inefficiently classifying skin lesion disorders. This suggests that leveraging advanced computational techniques can lead to more accurate and reliable diagnoses [57].
• Optimal Detection of Skin Lesions: The comparative analysis reveals that AI, DL, and ML approaches offer optimal detection of skin lesions compared to conventional methods. This finding is crucial, as early and accurate detection is paramount for the timely intervention and treatment of skin disorders [2].
• Performance Metrics: This study assesses various performance metrics such as accuracy, robustness, and sensitivity across different models. Highlighting the performance of each algorithm provides insights into their effectiveness in handling diverse skin lesion datasets and informs researchers and practitioners about the most suitable approach for specific diagnostic tasks. The details performance metrics are provided in the Supplementary Materials, section 6.
• Flexible Perspective: By providing a tabulated representation of the database, the analysis offers a flexible perspective on the contributions and relevance of different models. This aids in understanding the strengths and limitations of each approach, enabling researchers to make informed decisions when selecting algorithms for skin lesion classification tasks.
• Implications for Clinical Practice: Ultimately, the findings from this comparative analysis can have direct implications for clinical practice. The adoption of advanced AI techniques in dermatology can enhance diagnostic accuracy, assist dermatologists in decision-making processes, and improve patient outcomes by facilitating early detection and treatment of skin lesions [58].

4.1. ACO with or without NNs to Perform Precise SLD Classification

Comparative analyses of state-of-the-art, DL, and ML approaches specifically the ACO model (supported or not by) NNs are reviewed from different peer-reviewed articles to prove the superiority of in carrying out efficient classifications of SLDs flexibly.

Table 2. The critical outcomes and numerical outturns related majorly to ACO and NNs roles in classifying skin lesion disorders.

No.	Paper (Researcher(s) and Year)	Kind of ML/DL Model Analyzed and Its Use	Dataset Classification	Contributory Findings
1	Sengupta et al. (2019) [25]	ACO Model	Skin Lesion Visual RGB Dataset	The introduction of ACO has enhanced efficiency levels.
2	Anjum et al. (2020) [58]	ACO, ResNet-18, tinyYOLOv2, Optimized Support Vector Machine (O-SVM) and Optimized Naive Bayes (O-NB)	Skin Lesion MICCAI ISIC 2017, 2018, and 2019 Datasets	DL principles, optimization algorithms, active classifiers, and ACO-driven feature selection enhanced accuracy in localizing, segmenting, and classifying skin lesion defects in visual data.
3	Singh et al. (2021) [59]	A Novel Hybrid mode (SLICACO) from (A) the Simple Linear Iterative Clustering (SLIC) and (B) ACO Models	Visual Skin Lesion Benchmark Dermatoscopic PH2 Archive Dataset	Simulation results from the novel hybrid model incorporating ACO indicate a classification accuracy rate of approximately 95.9% for multiple skin lesion images.
4	Ahmed et al. (2019) [60]	Transductive Support Vector Machine (TSVM) Identification with the Suppiort of ACO and GA as Clustering Optimizers	Dermatological Visual Dataset Comprising Skin Lesion Disorders, Skin Cancer, and Other Skin and Dermatological Deficiencies	The TSVM model facilitated an active and efficient classification of various skin and dermatological problems, categorizing 24 different patterns with an elevated accuracy of approximately 95%.
5	Sengupta et al. (2020) [48]	Nature Inspired Algorithm (NIA) and ACO, ACO-Prewitt, Edge Smoothing Color Space, and ACO-Sobel	Skin Lesion Visual Dataset Consisting of RGB Dermoscopy Images of Skin Lesion Issues	Their proposed method demonstrates more robust performance and higher accuracy percentages compared to other optimization algorithms such as Edge Smoothing-Color Space, ACO-Sobel, and ACO-Prewitt models.
6	Sarada et al. (2021) [61]	A Hybrid Multi-Layer Algorithms of (I) K-Means and (II) WOA	Dermatological Dataset of Visual Information Containing Different Skin Lesion, Skin Cancer, and Other Skin Defects	Numerical results demonstrate that integrating AI and DL models with firefly optimization procedures can decrease the early stage detection error rate.
7	Zhao et al. (2023) [62]	An improved ACO (LACOR)	Visual BSDS500 Dataset Comprising Skin Defects, particularly Skin Cancer and Melanoma	The LACOR model, alongside other ML and DL models, facilitates active segmentation and high-performance classification processes for skin lesions and melanoma disorders, offering significantly higher quality rates compared to other algorithms.
8	Dalila et al. (2017) [63]	ACO, KNN, and Artificial Neural Networks (ANNs)	Visual Skin Lesion Dataset Containing Melanoma Problems and Dermoscopic Images	ANNs classified skin lesion defects with an accuracy of 93.60%, surpassing the 86.60% achieved by traditional techniques.
9	Yang et al. (2022) [64]	Enhanced ACO for Continuous Ranges (EACOR)	Visual Melanoma Pathology Dataset	Their proposed model could significantly contribute to the high-quality analysis and classification of visual datasets related to melanoma pathology detection and investigation.
10	Mirunalini et al. (2017) [65]	NN Model		The proposed NN model efficiently learned, aiding in the categorization of visual datasets related to melanoma and skin lesion issues during training, leading to improved levels of accuracy.
11	Shetty et al. (2022) [41]	CNN Model	HAM-10000 Visual Dataset consisting of 10,015 images of Various Dermoscopic and Skin Lesion Issues	Numerical simulations revealed that utilizing the CNN model enhanced the accuracy of skin lesion classification.
12	Attique et al. (2022) [66]	Moth–Flame Optimization Algorithm, MobileNetV2, Multiclass Extreme Learning Machine Classifier (MELMC)	Three imbalanced Visual Skin Datasets, (I) HAM10000, (II) ISBI2018, (III) ISIC2019	A comparison with current ML and DL techniques demonstrated the enhanced efficiency of the suggested paradigm.
13	Maqsood and Damaševičius (2023) [67]	Four pre-trained CNN algorithms, namely (A) Xception, (B) ResNet-50, (C) ResNet-101, and (D) VGG16, besides Multi-Class SVM (MC-SVM)	(1) HAM10000, (2) ISIC2018, (3) ISIC2019, (4) PH2 datasets	The accuracy for the four investigated and analyzed datasets, respectively, reached values of 98.57%, 98.62%, 93.47%, and 98.98% in skin lesion classification.
14	Shan et al. (2022) [68]	DenseSFNet-45, Dense Convolutional Network (DenseNet), Squeeze-and-Excitation (SE) Block	Three public Visual Skin Lesion Datasets: (i) ISBI 2017 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset (ISBI-skin-2017), (ii) ISBI 2018 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset (ISBI-skin-2018), and (iii) PH2 dataset	The proposed method surpassed conventional machine learning algorithms, current classical model classification schemes, and state-of-the-art techniques.
15	Tan et al. (2020) [69]	Hybrid Learning Particle Swarm Optimization (HLPSO), D-CNN, and Firefly Algorithm (FA), KM Clustering Algorithm	(I) Dermofit Image Library, (II) PH2, (III) ISIC 2017 Dataset	The simulation investigation revealed that the proposed HLPSO model outperformed both advanced and classical search models in managing skin lesion classification, based on benchmark activities related to mathematical landscapes and recognition tasks associated with the complex CEC 2014 testing suite.
16	Bi et al. (2017) [70]	Deep Residual Networks (ResNets)	Visual Skin Lesion ISIC 2017 Dataset	The ResNets model significantly enhanced performance and accuracy in classifying skin lesions, achieving an average AUC of 91.50%, surpassing that of other conventional ML models.

summarizes the significant findings and corresponding datasets and the AI, DL, or ML approaches implemented, focusing on ACO, NNs, and other types of paradigms. The following is a summary of the significant limitations of the ACO model with or without NNs.

• Sengupta et al. (2019) [25]: The primary limitation of the ACO model proposed by Sengupta et al. lies in its computational complexity and the potential for premature convergence to local optima. This can lead to suboptimal classification of skin lesions under varying conditions and complexities presented in visual RGB datasets.
• Anjum et al. (2020) [58]: Anjum et al.’s research, while pioneering in integrating ACO with deep learning models for skin lesion classification, faces limitations in generalizability across different datasets beyond MICCAI ISIC 2017, 2018, and 2019. The model’s performance might vary significantly, with datasets having different characteristics, potentially affecting its applicability in real-world scenarios.
• Singh et al. (2021) [59]: The novel hybrid model SLICA-CO, combining SLIC and ACO, although highly accurate, may suffer from scalability issues when dealing with very large dermatoscopic image archives. Additionally, the complexity of parameter tuning for the hybrid model can pose challenges in achieving optimal performance across diverse image sets.
• Ahmed et al. (2019) [60]: While the use of TSVM with ACO and GA for classifying dermatological issues shows high accuracy, the approach’s critical limitation is its dependency on high-quality labeled data for training. In dermatology, where subtle distinctions exist between different conditions, the lack of extensive, well-annotated datasets can hinder the model’s learning and generalization capabilities.
• Sengupta et al. (2020) [48]: The ACO-optimized Canny edge detection method, despite its robust performance, may encounter difficulties in edge detection for low-contrast skin lesion images. The limitation stems from the inherent challenge of accurately detecting edges in images where the lesion boundaries blend with the surrounding skin.
• Sarada et al. (2021) [61]: The hybrid multi-layer algorithms of K-Means and WOA, while effective in reducing detection error rates, face limitations in handling highly imbalanced datasets, where the prevalence of one class significantly outweighs others. This imbalance can skew the classification performance, affecting sensitivity and specificity.
• Zhao et al. (2023) [62]: The improved ACO model, LACOR, although advantageous for skin lesion, cancer, and melanoma classification, has limitations in adaptability and computational efficiency when applied to extremely large and complex visual datasets. The model may require significant computational resources for training and inference, impacting its practical deployment.
• Dalila et al. (2017) [63]: While the integration of ACO, KNN, and ANN demonstrates promising results in melanoma detection, a critical limitation is the potential for overfitting, especially in scenarios where the diversity of the training dataset does not adequately represent the variance found in real-world conditions.
• Yang et al. (2022) [64]: The Enhanced ACO for Continuous Ranges (EACOR) model faces limitations in dynamic adaptability to new or evolving types of melanoma pathology, potentially requiring frequent retraining or adjustment of its optimization parameters to maintain high levels of effectiveness and performance.
• Mirunalini et al. (2017) [65]: The proposed NN model, despite its learning efficiency, is limited by its relatively lower accuracy rate of approximately 65.8%. This suggests a need for further model optimization or incorporation of additional features to improve its classification performance for skin lesion and melanoma problems.
• Shetty et al. (2022) [41]: The CNN model’s critical limitation lies in its intensive computational demands for training and inference, particularly when processing large datasets of dermoscopic images. This can pose challenges in resource-constrained environments or in real-time applications.
• Attique et al. (2022) [66]: The reliance on Moth–Flame Optimization Algorithm and absence of data augmentation phases may limit the model’s ability to generalize across different skin conditions and variations within the datasets, potentially affecting its robustness in diverse clinical settings.
• Maqsood and Damaševičius (2023) [67]: While the four pre-trained CNN algorithms show high accuracy, their limitation is the potential for model overfitting due to the high capacity of these networks. Balancing model complexity and generalization remains a challenge, especially in multi-class skin lesion classification tasks.
• Shan et al. (2022) [68]: The DenseSFNet-45 model, despite outperforming conventional algorithms, may face limitations in interpretability and transparency, making it challenging for practitioners to understand the basis of its classification decisions, which is crucial in medical diagnostics.
• Tan et al. (2020) [69]: The HLPSO model’s limitation is the potential for increased computational complexity due to the hybrid nature of the optimization process, which combines several algorithms. This complexity can lead to longer training times and may require more computational resources.
• Bi et al. (2017) [70]: The Deep Residual Networks (ResNets) model, while achieving high performance, is limited by its sensitivity to hyperparameter settings and the need for extensive computational resources, which may limit its accessibility for some research or clinical environments.

From Table 2, it is evident that the diagnosis and categorization of skin cancer, along with discerning variations in skin textures and injuries, pose significant challenges. The manual detection of skin lesions in dermoscopic images is a complex and cumbersome task. Recent advancements in the IoT and AI for clinical applications have resulted in a significant increase in precision and processing time, especially with the support of ACO and WOA simulations. Consequently, substantial attention has been directed towards DL models, which yield effective outcomes in identifying cancer cells. As a result, the level of diagnosis and accuracy can be significantly elevated by categorizing benign and malignant dermoscopy images.

4.2. WOA with or without NNs to Conduct Accurate SLD Recognition

In comparison to the previous findings attained for the ACO model, this section similarly identifies many constructive rationale aspects and some limitations of the WOA when incorporated with (or not) NNs. These outcomes are shown in

Table 3. Summary of previous works on utilizing WOA (with/without) NNs.

Paper (Researcher(s) and Year)	Kind of WOA (with/without NNs)	Dataset Classification	Critical Limitations	Contributory Findings
Mirjalili and Lewis [27]	Standard WOA (without NNs)	Not specific	Potential for premature convergence and sensitivity to parameter settings.	Introduced the Whale Optimization Algorithm, demonstrating its effectiveness in solving optimization problems.
Mafarja and Mirjalili [23]	Hybrid WOA with simulated annealing	Not specific	May require careful tuning of hybrid algorithm parameters to ensure a balance between exploration and exploitation.	Proposed a hybrid approach combining WOA and simulated annealing for improved feature selection in machine learning.
Mafarja and Mirjalili [71]	WOA for feature selection (without NNs)	Eighteen UCI datasets were used in the experiments	Risk of not adequately exploring the search space in high-dimensional datasets.	Explored various whale optimization approaches for wrapper feature selection, enhancing the selection process in machine learning models.
Amiriebrahimabadi and Mansouri [72]	WOA for feature selection (without NNs)	Genomics	Limited by the complexity and high dimensionality of genomic data.	Utilized WOA for frequency-based feature selection in genomics, showcasing the algorithm’s applicability in bioinformatics.
Sharawi and Zawbaa [73]	WOA for feature selection (without NNs)	The data set is provided by a power enterprise, including the electricity consumption data of 11,860 high-voltage users who have had arrears	Dependency on the quality of the dataset and the risk of overfitting.	Demonstrated the effectiveness of WOA in feature selection, improving the performance of computational intelligence systems.
Sarada and Murthy [61]	WOA for data classification (without NNs)	Generated Data Analytics	Challenges in adapting WOA to diverse and dynamic data analytics environments.	Applied WOA to improve data classification for analytics, highlighting the algorithm’s versatility in handling various data types.

to allow medical specialists, HCPs, and scholars of the same interest to recognize the major key strengths and weaknesses of WOA in conducting SLD recognition with more accuracy.

5. Leveraging ACO, WOA, and NNs for Accurate Skin Lesion Detection

Healthcare providers face significant challenges in accurately detecting and classifying skin lesions due to various factors such as subjectivity in visual inspection, time constraints, limited access to expertise, scarcity of training data, and the complexity of lesions. Skin lesions encompass a wide range of abnormalities, including benign moles, malignant tumors, infections, and inflammatory conditions. Accurate diagnosis and classification are critical for appropriate treatment and management of skin disorders, particularly in cases where early detection can significantly impact patient outcomes. Traditional methods of skin lesion classification rely heavily on visual inspection by trained professionals, but they are subjective and prone to errors, leading to misdiagnosis or delayed diagnosis [74].

5.1. Addressing Challenges with ACO, WOA, and NNs

5.1.1. Enhanced Accuracy

Ant Colony Optimization (ACO) and Whale Optimization Algorithm (WOA) are metaheuristic optimization techniques that can effectively optimize the parameters of neural network models used for skin lesion classification. By fine-tuning these parameters, ACO and WOA algorithms improve the accuracy of classification, minimizing misclassification errors and enhancing diagnostic outcomes. The optimization process helps neural networks better capture the underlying patterns and features present in skin lesion images, leading to more precise classifications [25].

5.1.2. Automation and Efficiency

Integration of ACO and WOA with neural networks automates the diagnostic process, reducing the time and effort required by healthcare providers. Automated systems can analyze skin lesion images rapidly and provide accurate diagnoses, allowing for timely interventions and treatments. This automation streamlines the diagnostic workflow, enabling healthcare providers to focus their expertise on cases requiring further examination or treatment planning [25].

5.1.3. Robustness to Variability

Neural networks, particularly deep learning models like Convolutional Neural Networks (CNNs), excel at learning complex patterns and features from diverse datasets. CNNs can adapt to variations in skin lesion appearance and texture, making them robust to variability in lesion characteristics. This adaptability enables neural networks to generalize their learning to new, unseen skin lesion images, ensuring consistent and reliable classification across different patient populations and healthcare settings [22].

5.1.4. Optimization of Resources

ACO and WOA algorithms optimize the parameters of neural networks efficiently, maximizing the utilization of available resources, such as computational power and training data. This optimization ensures that healthcare providers can achieve accurate diagnoses even with limited resources, making skin lesion detection more accessible in regions with healthcare resource constraints. By optimizing resource allocation, ACO and WOA contribute to the scalability and sustainability of automated skin lesion detection systems in diverse healthcare settings [75].

5.1.5. Generalization

Trained neural networks can generalize their learning to new, unseen skin lesion images, allowing for consistent and reliable classification across different patient populations and healthcare settings. This generalization capability is crucial for ensuring the scalability and applicability of automated skin lesion detection systems in real-world clinical practice.

By providing consistent and reliable classification results, ACO, WOA, and NN models enhance the trust and adoption of automated diagnostic tools among healthcare providers and patients [2].

5.2. Integrating Ant Colony and Whale Optimization with NNs, AI, ML, and DL

In this study, a combination of Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) models is utilized for skin lesion detection. Specifically, Ant Colony Optimization (ACO) and Whale Optimization Algorithm (WOA) are integrated with Neural Networks (NNs) to enhance the performance of skin lesion detection systems. Here is a detailed explanation of the models and their integration.

5.2.1. Neural Networks (NNs)

Neural networks are a class of machine learning algorithms inspired by the structure and function of the human brain. In the context of skin lesion detection, NNs are employed to learn complex patterns and features from input images. They consist of interconnected layers of artificial neurons, including input, hidden, and output layers, and are trained using labeled datasets to classify images into different categories, such as malignant or benign lesions [76].

5.2.2. Artificial Intelligence (AI)

AI encompasses a broad range of techniques and algorithms that enable machines to perform tasks that typically require human intelligence. In the context of skin lesion detection, AI algorithms are used to optimize the parameters of NNs and improve their performance in classifying skin lesions accurately. ACO and WOA are examples of AI-based optimization algorithms that can be integrated with NNs to enhance their efficiency and effectiveness [77].

5.2.3. Machine Learning (ML)

ML algorithms enable computers to learn from data and make predictions or decisions without being explicitly programmed. In the context of skin lesion detection, ML techniques are employed to train NNs using large datasets of labeled skin lesion images. ACO and WOA can be used as optimization algorithms within the ML framework to fine-tune the parameters of NNs and improve their ability to classify skin lesions accurately [78].

5.2.4. Deep Learning (DL)

DL is a subset of ML that utilizes neural networks with many layers (deep architectures) to learn intricate patterns and representations from data. In skin lesion detection, DL models, such as Convolutional Neural Networks (CNNs), are commonly used due to their ability to automatically extract relevant features from images. ACO and WOA can be integrated with DL models to optimize the architecture and parameters of CNNs, leading to improved performance in skin lesion classification tasks [78].

ACO and WOA are employed as optimization algorithms to optimize the parameters (e.g., weights and biases) of NNs. During the training process, ACO and WOA iteratively adjust the parameters of NNs based on the evaluation of fitness functions, which measure the performance of the NNs in classifying skin lesions.

The optimized NNs are then used to classify skin lesion images, where they process input images and predict the presence or absence of lesions based on learned patterns and features. By integrating ACO and WOA with NNs, researchers aim to improve the efficiency and accuracy of skin lesion detection systems, ultimately contributing to better diagnostic outcomes in dermatology.

5.3. Numerical Results

The performance of ACO and WOA algorithms in skin lesion detection is quantified through various numerical results and metrics reported in the literature. These findings demonstrate the effectiveness and performance of ACO and its variants in addressing the challenges of skin lesion detection.

5.3.1. ACO Model [58]

The ACO model promotes accuracy in localizing, segmenting, and classifying defects in skin lesion images. By integrating deep learning principles with optimization algorithms and active classifiers, the model achieves robust performance in skin lesion detection tasks. The reported metrics showcase the model’s ability to achieve high accuracy in identifying and classifying various types of skin lesions, including both common and rare abnormalities.

5.3.2. Hybrid SLICACO Model [59]

The hybrid SLICACO model achieves a classification accuracy of approximately 95.9% for multiple skin lesion images. By integrating the Simple Linear Iterative Clustering (SLIC) algorithm with ACO, the model enhances performance in skin lesion segmentation and classification. The reported accuracy metrics demonstrate the effectiveness of the hybrid model in accurately identifying and classifying diverse skin lesions, underscoring its potential for clinical applications.

5.3.3. ACO Support [60]

The ACO support model demonstrates active and efficient classification of various skin issues and dermatological diseases with approximately 95% accuracy. By classifying 24 different patterns of skin problems, the model showcases its versatility and robustness in addressing a wide range of clinical scenarios. The reported accuracy metrics validate the model’s effectiveness in accurately identifying and classifying different types of skin lesions, including both common and rare abnormalities.

5.3.4. ACO-Optimized Edge Detection [48]

ACO-optimized edge detection methods exhibit significantly robust performance and enhanced accuracy compared to other optimization algorithms. By leveraging ACO for edge detection in skin lesion images, the model achieves superior performance in lesion localization and segmentation tasks. The reported metrics highlight the effectiveness of ACO-optimized edge detection techniques in accurately delineating the boundaries of skin lesions, facilitating precise classification and diagnosis.

5.4. Advantages of Integration

The integration of ACO, WOA, and NN models offers several advantages that contribute to mitigating substantial costs, time, and effort associated with traditional classification approaches in various ways:

5.4.1. Optimization of Classification Parameters

ACO and WOA algorithms optimize neural network parameters efficiently, reducing time and effort for classification. By fine-tuning these parameters, the integration ensures that neural networks can adapt to the complexities and variations present in skin lesion images, leading to improved accuracy and efficiency [79].

5.4.2. Improved Accuracy and Efficiency

Enhanced adaptability to variations in skin lesion images leads to higher accuracy and efficiency in classification. The integration of ACO and WOA with neural networks ensures that classification models can capture subtle nuances and features indicative of different types of skin lesions, resulting in more precise and reliable diagnoses [80].

5.4.3. Automation of Diagnostic Process

Automated classification streamlines the diagnostic process, saving time and effort for healthcare providers. Once trained and optimized, integrated models can quickly and accurately classify skin lesion images without the need for extensive manual intervention, enabling healthcare providers to focus on other critical tasks [81].

5.4.4. Consistent and Reliable Classification

Standardized and reproducible outcomes across settings ensure reliability in skin lesion classification. By minimizing subjectivity and variability inherent in traditional approaches, the integration of ACO, WOA, and NN models ensures consistent and reliable classification results, enhancing overall diagnostic reliability and confidence [82].

5.4.5. Reduction of Human Error

Minimization of human error through automation leads to more reliable diagnoses and improved patient outcomes. By reducing reliance on manual intervention and subjective assessment, integrated models help mitigate the risk of human error in skin lesion classification, resulting in more accurate and timely diagnoses for patients [2].

In conclusion, the integration of ACO, WOA, and NN models offers a promising approach to overcoming the challenges faced by healthcare providers in skin lesion detection. By leveraging these advanced technologies, healthcare systems can improve diagnostic accuracy, efficiency, and accessibility, ultimately leading to better patient outcomes in dermatology. Moving forward, continued research and development in this area holds the potential to further enhance the capabilities of automated skin lesion detection systems, driving innovation and advancement in clinical practice.

6. Future Work and Recommendations

Relying on the imperative extensive overview outcomes, this study proposes a collection of important future work prospects to focus on as a future work, suggesting scholars, medical researchers, and other scientists in the same field to inspect to improve the robustness of the findings attained from the comprehensive review. These noteworthy suggestions are reflected in the following ideas:

• To review other breakthrough concepts and approaches that contributed to considerable upgrading to the numerical ML and DL models’ performance.
• To investigate other technical tactics that can be adopted during the classification procedure to alleviate the adverse effects of massive data dimensionality and computational effort, time, and cost involved in the classification task of visual medical datasets, like SLD.
• To propose hybrid forms of ACP or WOA that can overcome conventional ACO challenges and WOA limitations in classifying different SLDs.
• To conduct a comparative inspection pertaining to the accuracy, robustness, and effectiveness of self-supervised learning (SSL) models compared with ACO and WOA, knowing that SSL models do not need costly and time-consuming data annotation of visual SLD patient datasets.
• To classify other key strengths and other common weaknesses (to handle) in ACO and WOA to upgrade their SLD categorization capability.

In this paper, we present two recommendations:

• The first recommendation is to use both algorithms together when working with convolutional neural networks and other networks to compensate for any shortcomings of one algorithm with the other. For example, algorithm (ACO) can be used in conjunction with algorithm (WOA), denoted as algorithm (A) and algorithm (B), respectively. If algorithm (A) produces weak prediction accuracy, algorithm (B) can be used to verify the results and adopt the more accurate one.
• The second recommendation is to merge both algorithms to work with deep neural network models and determine if they produce better results.

7. Conclusions

This study was initiated to shed more light and provide a special focus on practical ACO and WOA applications necessary for medical specialists and even for highly experienced HCPs to consider in their medical work to aid them in classifying different types of SLDs, including benign or malignant skin problems. Accordingly, optimum therapeutical processes can be adopted to make sure that patients are taking the correct medicine.

It relied on a comprehensive overview through which ACO and WOA contributory benefits have been addressed and explained. However, to offer more practicality and robustness of this work, modern state-of-the-art approaches have been illustrated, which can be involved in those two ML models to carry out necessary SLD classification with more upgraded levels of accuracy, reliability, and capability.

One of those innovatory approaches was the NNs. A brief review was provided pertaining to the SLD and its prevalence universally. Then, contributory workabilities of ACO and WOA, when supported (or not) with NNs, have been surveyed and investigated. On the other hand, to ameliorate the critical findings of this extensive overview, certain criteria were taken into consideration. Relying on the major extensive review outcomes, the core results of this paper can be summarized in the following points:

• Both ACO and WOA models, which can be supported (or not) with NNs, have proved their contributory impacts and beneficial practicalities in handling variant difficult-to-identify SLDs that might be also challenging for high-experienced HCPs to achieve.
• Significant computational costs, time, and effort corresponding to the necessary recognition processes of SLDs could be reduced compared with traditional numerical or manual/classification approaches, utilized commonly by most HCPs.
• The rates of accuracy, reliability, potential, and effectiveness of SLD classification procedures could be remarkably upgraded when ACO or WOA are utilized with or without the incorporation of NNs.
• Earlier prediction of the SLD can be flexibly accomplished due to escalated proportions of accuracy and efficiency in identifying the current problem in patients. Responsively, corrective medical actions can be implemented earlier before the SLD may expand, helping alleviate mortality rates.

8. Research Limitations

Despite the successful findings attained from the thorough overview carried out in this paper, it is essential to report a few restrictions and concerns that limit the broad feasibility of this extensive overview, including the following:

• There were some peer-reviewed articles with robust findings and noteworthy rationale. Unfortunately, they had no open axis.
• The corresponding number of papers addressing ACO or WOA when combined with other novel approaches is not abundantly available in the literature.