1.1 Motivation to improve skin cancer diagnosis

The objective of this article is to offer a thorough examination of the diverse facets and components associated with the diagnosis of skin cancer. The objective of this review is to present a comprehensive analysis of computer-aided diagnosis (CAD) techniques employed in the detection and classification of skin cancer. To accomplish this aim, this research has conducted a systematic review to evaluate and assess existing methods based on predetermined evaluation criteria. A comprehensive examination has been carried out on selected sources to gather relevant literature review regarding the methodologies utilized, the results achieved, and the dataset available in skin cancers. This research holds considerable importance, as it underscores the heightened mortality rates associated with skin cancer, both presently and in projected future scenarios.

The primary aim of this review is to improve understanding of the problem domain and its proposed solutions through a comprehensive examination of a wide range of academic literature. The main goal is to optimize. The precision and effectiveness of skin cancer diagnosis, lead to enhanced patient outcomes and reduced mortality rates. Table 1 exhibits the comparison of the presented survey with further reviews of the skin cancer recognition system. We reviewed 20 papers from 2018, 25 papers from 2019, 37 papers from 2020, 29 papers from 2021, 13 papers from 2022, and 7 papers from this year 2023.

Skin cancer, classification, deep learning (DL), early diagnosis, dermoscopy, prevention, and many more are the preliminary search terms applied for paper selections. Subsequently, found 793 academic papers and conference proceedings fulfilling the given selection criteria. A total of 345 papers from the first cohort were chosen depending on their title relevance for the current research. The number of research papers dropped to 147 after a more thorough examination of the abstracts of the specified publications was then done to ascertain their fit. The chosen papers depending on their abstracts underwent extensive review to ascertain their general quality. After much review, the last collection of 126 academic publications was selected for close inspection and analysis. These final papers come from several sources and are distributed as follows: From IEEE publications, 15%; from Google Scholar, 25%; from the Association of Computing Machinery Digital Library, 11%; from Springer, 29%; from Science Direct. Figure 4. below shows the yearly paper selection. The selection process was meant to guarantee that the research articles selected for inclusion in this study came from a great range of reliable sources and academic venues. The main goal of this large collection is to guarantee the inclusion of important points of view from the current corpus of literature in this specific field of research, so laying a strong basis for further analysis and evaluation.

1.2 Scope of current research

The application of automated techniques for disease identification and classification would help much in skin cancer research. The concerning rise in skin cancer cases emphasizes the need for early identification to lower death rates. The aim of this work is to evaluate current approaches for skin cancer identification and classification. The processing in this study revolves mostly around image processing, computer vision, and other types of ML. The objectives, approaches, and results of the research are introduced in the first part of this article. Section 2 breaks out the method the skin cancer detection approach employs. Section 3 covers the challenges of detection, therefore offering an understanding of the issues scientists encounter. Proceeding further, Section 4 advances in the ML portion; Section 5 details the reference datasets used to confirm the accuracy of the outcomes. Section 6 also investigates the usage of mobile apps tailored for smartphones for research on skin cancer. This paper investigates how quickly and simply mobile technologies might be applied to diagnose skin cancer. Section 7 examines closely the research techniques and provides suggestions for the further study. The research report ends with a careful review and useful recommendations for the next projects. Figure 5 below shows the overall paper theme. As researchers add to the always-growing corpus of knowledge on skin cancer detection and classification, these suggestions are supposed to act as a road map and source of inspiration.

1.3 Necessity of CAD

The current diagnostic techniques employed for skin cancer are subject to various limitations that have implications for their efficacy and precision. Figure 6 depicts the need for CAD. This section elucidates the constraints and difficulties inherent in current methodologies, providing insight into areas that necessitate enhancement.¹¹

1.4 Difficulties in skin lesion analysis

Examining and analyzing skin lesions holistically calls for different challenges and nuances. This part clarifies the difficulties in microscopic analysis of skin lesions, including the resolution of natural constraints and the capture of complex characteristics.²⁵

2.1 Techniques for preprocessing

The process of preprocessing is of utmost importance in the analysis of skin lesions as it serves to eliminate extraneous information, enhance the quality of images, and optimize the performance of the system.⁴⁴ A range of preprocessing techniques are employed to mitigate artifacts and improve visual effects. Table 2 below shows the preprocessing techniques used in recent years.

2.1.2 Techniques for filtering and enhancing data

Image noise and artifacts have the potential to lower the overall quality of collected and transmitted images. By employing filtering algorithms to get rid of extraneous noise.⁵⁵ In addition, image enhancement methods are used to adjust an image's brightness, contrast, and sharpness. Several noise-reduction and image-enhancement techniques are put to use in the field of skin cancer identification.⁵⁶ Included are algorithms like contrast-limited adaptive histogram equalization, adaptive histogram equalization, global–local contrast stretching, color constancy with shades of gray, adaptive gamma correction, gamma correction, and more.⁵⁷ With the use of these preprocessing methods, skin lesion images can be greatly enhanced in quality and clarity, leading to increased precision and reliability in following analysis and classification steps for the detection of skin cancer.⁵⁸ Noise reduction filters, such as the median filter and Gaussian filter, are employed to mitigate salt-and-pepper noise. Edge enhancement filters like Sobel and Laplacian filters accentuate boundaries and rapid intensity changes. Frequency domain filters, like the Fast Fourier Transform, facilitate noise elimination through transformations between spatial and frequency domains.⁵⁹ Data enhancement methods include contrast enhancement techniques like histogram equalization, color space transformations such as RGB to HSV conversion, and data augmentation through rotations, scaling, and flipping for ML model robustness. Image fusion combines information from different modalities, and super-resolution techniques enhance image resolution. Figure 8, below shows data filtering, enhancement, and preprocessing techniques.⁶⁰ In the context of DL, normalization, image cropping, and data resampling are used as preprocessing steps to standardize and refine input data for optimal performance in skin cancer diagnosis and analysis.

2.2 Traditional and modern segmentation approaches

The process of segmentation plays a vital role in the examination of skin lesions as it allows for the accurate identification and separation of the specific area affected by the lesion.⁶¹ This section aims to elucidate various segmentation methodologies that have been utilized to achieve precise delineation of skin lesions from adjacent healthy skin.⁶² Thresholding-based Segmentation techniques that rely on thresholding assign a distinct threshold value to delineate the region of interest, specifically the lesion area, by considering the intensity of individual pixels.⁶² Commonly employed techniques in image processing include simple thresholding methods, such as global thresholding or adaptive thresholding. Below Figure 9 shows the traditional and modern DL segmentation methods. The aforementioned techniques involve the comparison of pixel intensity values with a predetermined threshold to generate a binary mask that effectively distinguishes the lesion from the background.⁶³

Region-based segmentation methods that are based on regions strive to categorize pixels into cohesive regions by considering their shared attributes, such as color, texture, or spatial proximity.⁶⁴ Techniques such as region growing, watershed transformation, and graph cuts employ seed points or markers to progressively expand or divide regions according to predetermined criteria.⁶⁵ The segmentation techniques that are based on edges primarily aim to identify and delineate the boundaries of skin lesions.⁶⁶ Typically, these methods utilize edge detection algorithms, such as Canny edge detection or gradient-based edge detection, to discern the edges that demarcate the region of the lesion.⁶⁷ Subsequently, the identified edges are employed to generate a boundary or contour encircling the lesion. Active contour models, commonly referred to as snakes or level sets, are flexible curves or surfaces that undergo evolution in response to image characteristics and imposed limitations.⁶⁸ The objective of these models is to identify the most optimal contour that accurately represents the boundary of the lesion. Active contour models can accurately capture the irregular shapes and boundaries of skin lesions through the iterative adjustment of the contour.⁶⁹

As far as DL segmentation, specifically convolutional neural networks (CNNs), has demonstrated significant progress in the field of skin lesion segmentation. The methodologies employed in this study involve the utilization of deep CNN architectures and training datasets to acquire knowledge of the distinctive features and patterns that differentiate the lesion region from the adjacent healthy skin.⁷⁰ Table 3 below, defines an analysis of DL-based segmentation. The trained models are capable of accurately and reliably segmenting skin lesions. Every segmentation methodology possesses distinct advantages and drawbacks. The selection of the most appropriate methodology is contingent upon various factors, including the attributes of the dataset, available computational resources, and the specific demands of the application.^{71, 72} Through the utilization of suitable segmentation methodologies, researchers and clinicians can effectively isolate the region of interest, thereby facilitating subsequent analysis and enhancing the overall precision of skin lesion diagnosis and classification.⁷³

2.3 Advances in ML for skin lesion analysis

In the field of dermatology, machine-learning methods have become important tools for analyzing and making sense of complicated medical data, like images of skin lesions.⁸⁹ Figure 10 below shows all the advances below. This part talks about the progress that has been made in ML methods for analyzing skin lesions, showing how they could be used to make it easier to find and classify skin cancer.⁹⁰

2.3.2 Feature extraction and representation

In ML, it is important to extract and describe features. These models are very good at picking out important details from pictures of skin tumors, which make it possible to get a full picture of the visual characteristics that help with diagnosis.⁹² With the help of CNNs and other methods for feature extraction, important patterns, and distinguishing traits can be found and extracted. A view of the CNN model is shown in Figure 11. In turn, this makes skin lesion labeling even more accurate.

2.3.5 Transfer learning

Transfer learning is a method that lets you use what you've learned in one job or set of data in another. It has shown a lot of promise in the field of skin lesion analysis.¹⁰² Skin lesion data can be used to fine-tune models that have already been made, like those that have been trained using large picture datasets like ImageNet.^{103, 104} This makes it possible to use the learned models and improve classification performance, even when there are only a few labeled samples to work with. Transfer Learning is an ML approach where a model trained on one task is adapted or fine-tuned for a related task.¹⁰⁵ In the context of skin cancer image analysis, a pre-trained neural network, such as a CNN, trained on a large dataset for a generic image recognition task, can be fine-tuned on a smaller dataset specific to skin cancer images.¹⁰⁶ Transfer learning leverages the knowledge gained from the source task to enhance the performance of the model on the target task with limited labeled data. This approach is particularly valuable when labeled data for the target task is scarce or expensive to obtain.¹⁰⁷ The pre-trained model captures general features from the source domain, and the fine-tuning process refines these features for the specific characteristics of skin cancer images.¹⁰⁸ Figure 12 below shows transfer learning general structure.

2.3.6 Interpretability and explainability

Recent improvements in ML for skin lesion research have put a lot of attention on making things easier to understand and explain. Different methods, such as attention mechanisms, saliency maps, and Gradient-weighted Class Activation Mapping (Grad-CAM), have been made to help us understand how ML models make decisions.¹⁰⁹ These methods are meant to help doctors understand and trust the predictions that these models make. Recent improvements in ML methods that can be used to analyze skin lesions have a lot of potential to change how skin cancer is diagnosed and treated.¹¹⁰ Table 5 below shows ML/DL comparative performance. This is done by doing things like automating lesion segmentation, extracting important features, making accurate classification models, using ensemble methods, taking advantage of transfer learning, and making things easier to understand.¹¹¹ Figure 13. below shows comparative performance metrics of ML and DL algorithms for above same references as in Table 4.

TABLE 5. Comparative performance of ML/DL algorithms.

Algorithm	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1-score (%)
Support vector machine (SVM)	85.6	83.4	88.2	84.5
Random Forest (RF)	87.2	85.3	89.1	86.5
CNN (VGG16)	91.5	89.2	93.4	90.7
ResNet-50	93.1	91.8	94.5	92.5
MobileNet	90.8	88.7	92.6	89.9

2.3.7 Algorithm analysis

SVM is widely used for skin lesion classification due to its robustness in small, structured datasets.¹¹² It excels in binary classification tasks but struggles with high-dimensional data and scalability. RF employs ensemble learning, which enhances performance by combining multiple decision trees. It is particularly robust to overfitting and handles imbalanced datasets effectively.¹¹³ However, it may lack interpretability. CNN models like VGG16 and ResNet-50 outperform traditional ML algorithms by automating feature extraction. ResNet-50, in particular, addresses vanishing gradient issues with residual blocks, making it suitable for diverse lesion datasets. MobileNet lightweight architecture is optimized for real-time diagnosis on mobile devices.¹¹⁴ Its efficiency makes it suitable for resource-constrained environments while maintaining competitive accuracy. Figure 14 below shows the comparison of MobileNet and ResNet architectures.¹¹⁵

2.4 Classification model performance analysis

A comprehensive evaluation of the efficacy of the classification algorithms utilized in the selected scholarly articles. Various performance metrics, such as accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic (ROC) curve were used to evaluate the performance of these algorithms.^{116, 117} The primary objective of the study was to comprehensively evaluate the accuracy of various classification algorithms in accurately categorizing cases of skin cancer, with a focus on identifying their respective strengths and limitations.¹¹⁸ The performance metrics provided insightful information regarding the algorithms' overall accuracy, their precision in accurately detecting positive cases, their recall in capturing all positive cases, and the F1 score, which accounts for the balance between precision and recall.^{119, 120} In addition, the area under the ROC curve was analyzed to evaluate the algorithms' ability to distinguish between positive and negative cases. The metric provided significant insights into the sensitivity and specificity performance of the algorithms.¹²¹ The primary objective of this review paper was to provide a thorough understanding of the advantages and disadvantages of various classification algorithms for the detection of skin cancer by performing a thorough analysis of performance metrics.^{122, 123} Figure 15 below shows the performance analysis of different classification algorithms across key metrics.

The results of this study have the potential to aid researchers and clinicians in selecting the algorithms best suited for accurate and reliable skin cancer classification. Table 6 above shows the performance analysis of different classification algorithms and their accuracy rate. Using the ISIC dataset,¹³⁴ researchers were able to accomplish our goal with 0.957% precision. When applied to the ISIC dataset, VGG16, VGG19, and Inception V3 all achieved accuracy rates of 77%, 76%, and 74%, respectively. Over 24 000 cutaneous cancer samples were classified using ResNet, VGG19, and InceptionV3 models in a separate investigation. Among the three models, InceptionV3 had the best results.¹³⁵ The International Skin Imaging Collaboration (ISIC) 2020 melanoma classification dataset was used alongside EfficientNets and ensemble models to categorize cutaneous lesions. All EfficientNet B6 models plus one EfficientNet B5 model in an ensemble produced the best results (ROC curve of 0.944%).¹³⁶ The researchers achieved a 91.51% classification accuracy when they used a CNN model to categorize pigmented lesions in the HAM10000 dataset.¹³⁷

The VGG16 and AlexNet models were gradually combined and improved using principal component analysis (PCA). An astonishing success rate of 99.9% was reached by this method.¹³⁸ In this study, the HAM-10000 dataset was used using five individual CNN models and four different ensemble models. The purpose of this study was to identify subtypes of cutaneous cancer. Accuracy of 93.20% and 92.83% were achieved using pre-trained ResNetXt101 and ensemble InceptionResNetV2 + ResNetXt101 models, respectively.¹³⁹ Using the ISIC 2020 dataset, the EfficientNetV2-M model correctly categorized cutaneous lesions 99.23% of the time.¹⁴⁰ Hyper-connected convolutional neural network (HcCNN) was created to identify cutaneous lesions. Multimodality testing with 7-point criteria had an accuracy of 74.9%.¹⁴¹ On the ISIC dataset,¹⁴² we devised a novel technique using a CNN with a unique regularizer for the binary classification of malignant and benign lesions. There was a 97.49% success rate with this method. For melanoma detection, we used the CNN models InceptionV4, SENet154, InceptionResNetV2, and PNASNeT-5-Large. On the ISIC 2018 dataset, PNASNeT-5-Large achieved a validation result of 0.76. The ARL-CNN model was created to deal with the problem of limited training data by taking intra-class variance and inter-class similarity into account. The accuracy of this method was 85.0% when tested on the ISIC 2017 dataset.¹⁴³ To solve the problem of cutaneous lesion classification in the HAM-10000 dataset, the MobileNet CNN architecture was created. The first version of the model had an accuracy of 83.93% without any data augmentation procedures.¹⁴⁴ Seven classes from the ISIC 2018 dataset were classified using GoogLeNet, VGG, and their ensemble models in a separate study.¹⁴⁵ Accuracy of 79.7%, 80%, and 81.5% were attained. Figure 16 defines classification models and their accuracy on skin cancer datasets.

3.1 Key datasets and preprocessing techniques

3.2 Accuracy of different algorithms

The accuracy of different algorithms is influenced by several factors. First, dataset size plays a crucial role, as larger datasets, such as ISIC and HAM10000, provide sufficient diversity and representation, enabling models like ResNet-50 and VGG16 to achieve higher accuracy. In contrast, smaller datasets often result in overfitting, negatively impacting generalizability.¹⁶⁶ Second, preprocessing techniques significantly affect accuracy, with methods like normalization, data augmentation, and oversampling (e.g., SMOTE) improving model performance by reducing noise and addressing class imbalance.¹⁶⁷ For instance, data augmentation enhances the accuracy of CNNs by increasing data diversity. Finally, architectural differences among models also influence their accuracy, as different architectures possess varying strengths and weaknesses. ResNet-50, for example, excels due to its deep residual connections, which effectively mitigate vanishing gradient issues. On the other hand, lightweight models like MobileNet prioritizes computational efficiency, resulting in a slight trade-off in accuracy.¹⁶⁸ Below Table 8 shows the comparative analysis of algorithms.

4.1 Miiskin

This application uses high-resolution photography to capture detailed images of vast body regions. Individuals can compare the lesions on their skin over time, allowing them to identify any changes that may be indicative of skin cancer.¹⁷³

4.2 MoleMapper

An application developed through a collaboration between Oregon Health & Science University, Apple, and Sage Bionetworks, is presently available at no cost to consumers. This technology allows physicians to remotely monitor potentially concerning lesions, thereby reducing the need for frequent in-person consultations.¹⁷⁴

4.3 MoleScope

To use the MoleScope application, individuals must purchase a smartphone-compatible accessory. Through the use of the accompanying device, individuals can capture images and submit them to dermatologists for online opinions and evaluations.¹⁷⁵

4.4 SkinVision

A council of dermatologists created SkinVision, an application that uses a deep-learning algorithm to analyze photographs of lesions. This application is offered for a fee. In 1 min, the system assesses the level of risk associated with the mole and provides users with valuable information regarding potential skin cancer dangers.¹⁷⁶

4.5 UMSkinCheck

The UMSkinCheck application, developed by the University of Michigan Medical School, allows users to perform comprehensive skin cancer examinations and monitor any changes that occur over time, at no cost. The application provides instructions for performing skin examinations, stores initial photographs for comparative analysis, and sends periodic notifications to encourage consistent self-evaluations. Table 9 shows the launch year, cost, and horizon of each mobile app. These applications provide a variety of functionalities and methodologies to aid individuals in monitoring their skin health, providing significant assistance and guidance in the early identification of potential skin cancer risks.

4.6 Application value of ML/DL techniques in clinical settings

The ISIC dataset has been extensively analyzed to demonstrate the practical application of ML and DL techniques for skin cancer detection. Through the use of DL models like ResNet-50 and MobileNet, the analysis highlighted significant improvements in diagnostic accuracy, sensitivity, and specificity. Preprocessing techniques, such as data augmentation and normalization, played a crucial role in enhancing the performance of these models, particularly in clinical-like settings. Similarly, the HAM10000 dataset was leveraged to train and evaluate ML/DL models for classifying diverse skin lesions. This study emphasized the need to address dataset imbalances and successfully applied oversampling techniques, such as SMOTE, to improve the detection of rare lesion types, including melanoma. Furthermore, MobileNet-based models were deployed on mobile devices to showcase the potential of edge AI for real-time diagnostics. This implementation demonstrated the feasibility of using such models in remote areas, providing greater accessibility for early skin cancer detection and diagnosis. Table 10 below shows case studies and clinical application details.

4.7 Challenges in clinical applications of ML/DL

To address the challenges in applying ML and DL techniques in healthcare, several solutions have been proposed. Federated learning enables collaborative training of models across institutions without compromising sensitive patient data, thereby addressing privacy concerns and enhancing the diversity of training datasets. Model fine-tuning, which involves adapting pre-trained models to domain-specific datasets, significantly improves performance on clinical data while reducing the need for large volumes of labeled samples.¹⁷⁸ Additionally, the establishment of standardized clinical protocols and benchmarks for validation ensures the robustness and reliability of ML/DL models in clinical applications, promoting their effective and trustworthy deployment in real-world scenarios.

6.1 Analysis of previous studies' conclusions

The conclusions derived from prior research investigations were examined to discover the principal innovations and difficulties in the Skin cancer detection process. This survey covered which methodologies are being used, and what type of dataset is being used and this analysis facilitates the identification of the strengths and limitations inherent in these approaches, thus contributing a valuable insight into the progress achieved in the domain of skin lesion analysis.