Skin Cancer Detection Using Deep Learning Approaches

Abstract

Aim:

This review examined multiple deep learning (DL) methods, including artificial neural networks (ANNs), convolutional neural networks (CNNs), k-nearest neighbors (KNNs), as well as generative adversarial networks (GANs), relying on their abilities to differentially extract key features for the identification and classification of skin lesions.

Background:

Skin cancer is among the most prevalent cancer types in humans and is associated with tremendous socioeconomic and psychological burdens for patients and caregivers alike. Incidences of skin cancers have progressively increased during the last decades. Early diagnoses of skin cancers may aid in the implementation of more effective treatment and therapeutic regimens. Indeed, several recent studies have focused on early detection strategies for skin cancer. Among the lesion features that can aid the recognition and characterization of skin cancers are symmetry, color, size, and shape.

Results:

Our assessment indicates that CNNs delivered maximum accuracy in visual lesion recognition, yet GANs have surfaced as a strong tool for training augmentation through simulated image creation. However, there were significant limitations associated with existing datasets, such as provision of insufficient skin tone variability, demanding computational needs, and unequal lesion representations, which may hamper efficiency, inclusivity, and generalizability of DL models. Researchers must combine diverse high-resolution datasets within a structural framework to develop efficient computational models with unsupervised learning methods to enhance noninvasive and precise skin cancer detection.

Conclusion:

The breakthroughs in image-based computational skin cancer detection may be crucial in reducing the requirement of invasive diagnostic tests and expanding the scope of skin cancer screening toward broad demographics, thereby aiding early cancer detection in a time- and cost-efficient manner.

Introduction

The prevalence of skin cancers has increased during the previous years, and they are thought to be among the most frequently diagnosed cancer types in humans. Relative deficiencies of skin pigmentation in Caucasian populations make them particularly vulnerable to skin cancers. In addition, ozone depletion is one of the major causes of increased incidences of skin cancers. Based upon cell type affected, skin cancers are classified as nonmelanoma and malignant melanoma.¹ Nonmelanomas are basal cell or squamous cell carcinomas, while melanomas are associated with metastases of melanin-producing melanocytes.² Melanomas have reduced prevalence (approximately 1% of all cases) but are deadlier with significantly enhanced mortality rates. Skin cancer lesions are often found on sun-exposed surfaces of the hands, faces, necks, and lips. Unfortunately, if not diagnosed and/or treated properly, melanoma tumors are known to easily spread to other regions of the body, ultimately resulting in mortality.³ Regrettably, melanomas have fewer treatment choices than other cancer types. Early identification of skin cancers, particularly melanomas, is essential for evaluating and establishing more effective therapies/treatments.⁴ The most common diagnostic procedure is a biopsy of the suspicious skin lesion. However, this has time- and cost-efficiency constraints, and is painful for the subjects. Computer-based technological developments have reduced the cost, complexity, and time involved in identifying whether a person has skin cancer.⁵ Furthermore, many of these regimens involve minimally invasive or noninvasive techniques for identification of melanomas.

In recent years, several databases have been created to accommodate skin lesion images. Despite their limitations in diversity, image quality, and lesion representations, these can potentially be explored for computational analyses. The standard procedure for identifying skin cancer includes image acquisition, preprocessing, segmentation, and extraction of the desired features and outcome assessment.⁶ Figure 1 represents the basic scheme of deep learning (DL) methods, which involves augmentation of training and test data, followed by feature extraction. The extracted features train a classifier model, which allows the trained model to classify inputs as either benign or malignant, thereby detecting the presence of skin cancer. Color-coded blocks and arrows have been used in the flowchart to represent the sequential process from data preparation to final classification.

FIG. 1.

The process of skin cancer detection using deep learning (DL) algorithms.

DL algorithms outperform more conventional machine learning (ML) methods and have improved the pattern recognition of ML in multiple arenas, such as industries, bioinformatics, medical, and clinical fields.⁷ One such tool is the artificial neural network (ANN), which recapitulates the functions of the human brain network.⁸ The applications of DL algorithms have also been extended to identification and characterization of skin cancers.⁹ This review attempts to compile relevant research evidence supporting the utilities of deep neural network (DNN)-based categorization in detecting, classifying, and characterizing skin cancers. This study summarized the current findings of early detection and classification of skin cancer based on DL techniques with non-invasive diagnostic procedures. The algorithms such as ANN, convolutional neural network (CNN), k-nearest neighbors (KNN), and generative adversarial network (GAN) used for identification and classification purposes are tested in the study. This study also refers to the widely used datasets: International Skin Imaging Collaboration (ISIC) and PH2 skin cancer datasets. By summarizing recent advances, it highlights opportunities to improve the accuracy of diagnoses, develop strong models, and fill the gap in data representation and training methodologies.

Research Methodology

This research identified and summarized the most recent and relevant skin cancer diagnosis methods based on neural networks (NNs). Literature search was conducted to collect and evaluate previously published research according to predetermined standards. This research study was guided through the deployment of criteria for selection to ensure relevance and quality in the choice of studies. Only those studies published in English from recognized journals, conferences, and other reputable platforms were considered. Given the dynamic nature of the area of research, more recent publications were prioritized, as they propose/evaluate/discuss the most recent developments related to NN-based skin cancer detection methods. Most of the selected studies were topical interest-based and included those that contained experimental detail and methodological robustness directly and appropriately related to the research objectives/questions (see Research questions). Literature references within, and studies citing the selected studies were included if found necessary and suitable. Other studies, which include case reports, conference proceedings, or lack proper documentation, were excluded.

Research questions

The overarching aim was to search and critically analyze the scholarly databases for articles focused on DL-mediated diagnoses of skin cancer. The authors developed the following research questions before starting the review of the literature. 1.

What are the most common algorithms used?

Which skin cancer databases are typically used?

What are the challenges in this field?

Search strategy

To obtain data for the above questions and summarize the relevant information, the authors conducted a thorough literature search of previously published primary studies. Relevant search terms and keywords used for literature search are summarized in Table 1. The selection criteria for primary studies were the works’ language, publication year, and topical value. All potentially interesting studies, including case studies, reports, conference proceedings, and research articles, were initially selected. Later, studies from unrecognized journals/conferences and those lacking extensive reporting were removed. The citations in these primarily selected studies were also examined carefully for pertinent researchers, which might have been overlooked during the initial search.

Table 1.

Primary and Secondary Keywords Used for Retrieval of Pertinent Literature

Primary keywords	Secondary keywords
Skin cancer, lesion, melanoma, machine learning, deep learning, detection and classification, lesion detection	Lesion detection, lesion classification, medical image processing, framework, algorithm

Lesion Detection

A structured framework can be applied to assess the effectiveness of algorithms, including ANN, CNN, KNN, and GAN, in skin cancer detection. The evaluation method relies on accuracy and sensitivity alongside specificity and precision metrics to determine the algorithms’ ability to diagnose cancerous and noncancerous lesions correctly. The matrices, F1-score, receiver operating characteristic (ROC) curves, and area under the ROC curve (AUC) score, enhance evaluation of lesion categorization reliably. Evaluation of generalizability focused on diverse input data collection from databases such as ISIC, PH2, HAM10000, and DermNet NZ (the Databases section) helps reduce overfitting by providing varied information about lesion types and skin tones accompanied by demographic diversity. Many of the algorithms show robust performances during tests, affirming their ability to process noisy data and heterogeneous lesions and images of varying quality. However, these testing techniques are required to be subjected to both cross-validation and clinical dataset examination to show their utilities for advancing the development of skin cancer diagnostic tools in practical real-world settings.

DL Methods

As indicated, DNNs have increasingly been used to detect and classify skin cancer lesions. These DNNs can be divided into a group of unique nodes communicating with one another. This nerve cell network is remarkably similar to that found in the human brain. To solve a specific problem, each node adds to the effectiveness of the network as a whole.⁹ An NN can become a reliable source of knowledge about a set of tasks once trained to perform them. With regard to skin cancer detection, skin lesion images can be characterized and categorized using NNs for the presence/absence of skin cancers and their different characteristics. A variety of skin lesion images may be conveniently acquired from databases such as the ISIC dataset.¹⁰ Various algorithms, including ANN, CNN, KNN, and GAN, can be used for the prediction of skin cancer using the pictorial characterization of lesion images.¹¹ This section is dedicated to the usage of these different DNN approaches for the prediction of skin cancer.

Table 2 concisely compares the methods, highlighting their strengths and limitations for skin lesion detection in skin cancer using advanced NN-based approaches that may increase the efficiency of diagnoses. ANNs use layered structures and back-propagation for features and classification of data, while CNN is more suitable for image-based assessment through convolution and pooling layers, which are very accurate with lesion identification. KNN is also classified by proximity, making it much more effective as it can better handle nonlinear relations. GANs, by generating synthetic images, aid overcoming the issue of data scarcity, thereby improving the model’s training and overall accuracy of detection. The workings of these still developing technologies, computationally expensive and requiring a quality dataset, high image diversity, and better lesion detection with clinical applicability, are individually discussed in subsequent subsections.

Table 2.

Comparison of the Different Deep Learning-Based Skin Lesion Detection Methods

Method	Key features	Strengths	Limitations
Artificial neural network (ANN)	Uses input, hidden, and output layers; back-propagation for refinement of predictions.	Applied to nonlinear modeling and feature extraction; effective both in supervised and unsupervised training.	Computationally expensive; need large dataset to train properly.
Convolutional neural network (CNN)	Makes use of convolution layers for feature extraction and pooling; then connected layers for classification.	Well-applied with images for lesion detection; works very well with other pretrained models such as ResNet and AlexNet.	Needs high-resolution images; information-rich and not too representative. Highly dependent on the data quality and diversity used in the model.
k-Nearest neighbor (KNN)	Nonparametric, unsupervised learning; clusters data by proximity in the feature space.	Performs great on big data; best algorithm for nonlinear relation in data.	Poor performance if the number of features is very large; computation is very heavy for big data.
Generative adversarial network (GAN)	Has the generator and discriminator networks; produces synthetic data similar to real samples.	It eases the data scarcity problem; improves model accuracy.	Needs large-scale compute power; overfits easily and diverges during training.

ANN-based algorithms

An ANN generates predictions using statistical learning and nonlinear modeling. In ANN, there are three layers of neurons. Information is sent from the first “input” layer neurons to the second “intermediate” layer neurons.¹² Conventional ANNs are more complicated than they appear, with hidden intermediate layers beneath their simple surface.¹³ As illustrated in Figure 2, an NN architecture is highlighted by input layers, hidden layers, and output layers, in which the input layer has five nodes representing variables such as RPM, slope, and their interactions. The hidden layer has several interconnected neurons processing the inputs, while the output layer consists of a single neuron predicting efficiency of lesion categorization. Furthermore, the network structure depicts a classical feedforward NN typically used for regression or classification tasks. Intermediate neurons pass the information to the third “output” layer neurons.¹⁴ Back-propagation exposes the complex interconnections between the input and output layers, allowing refinement of predictions.¹⁵ Computations carried out by each network level are important to achieve reliable output.¹⁶

FIG. 2.

Basic artificial neural network (ANN) structure.

ANN is used in skin cancer detection systems to categorize the detected properties into groups (Table 3). The input photographs are split into two categories after training/classification of the training set: those with melanoma and those without. The number of photographs an ANN processes determines the thickness of its hidden layers. The input dataset links the first layer to the second layer when building an ANN.¹⁷ Both supervised and unsupervised learning techniques can be used on labeled or unlabeled data. Semisupervised learning, the workflow of which is depicted in Figure 3, includes human annotation and task-clustering classifiers. This process allows manually annotating a human annotator on unlabeled data, which then converts the unlabeled data to labeled data. Once the data are labeled, they are fed into the task-clustering classifiers to further perfect the labeling process, which can also handle new incoming unlabeled data efficiently. Such a system forms a feedback loop that increases labeling accuracy over time.¹⁸ In an NN, each link or connection has a “weight” learned either through back-propagation or feed-forward. Each system has a unique organizational structure for the underlying data source. Every communication takes place in the space between the input and output layers, a method for evaluating whether a skin lesion is benign or metastatic.

FIG. 3.

Basic scheme of semisupervised learning.

Table 3.

Artificial Neural Network-Based Skin Cancer Detection Paradigms

Reference	Skin cancer diagnoses	Classifier	Results (accuracy)
⁵	Melanoma	Artificial neural network (ANN) with back-propagation method	97
²²	Melanoma/nonmelanoma	ANN with back-propagation algorithm	87
²¹	Cancerous/noncancerous	ANN with back-propagation algorithm	83.6
²⁵	Malignant/benign	A feed-forward artificial neural network (NN) trained using the back-propagation algorithm	80
²⁶	Malignant/nonmalignant	As an NN classifier, back-propagation NNs are used	71
²⁴	Cancerous/noncancerous	ANN with back-propagation algorithm	86
²⁷	Cancerous/noncancerous	ANN utilizing the back-propagation technique	88
²⁸	Basal cell carcinoma (BCC)/non-BCC	ANN	93
²⁹	Melanoma/nonmelanoma	Implementing the ANN learning procedures of LM, RBP, and SCG	94
³⁰	Malignant/benign	A fuzzy NN and a BPN are components of a meta-ensemble model for ANNs	95

A self-generating NN can be used to initially filter out photographs with lesions.¹⁹ Tumor margins, texture, and color are characterized in the second stage. Principal component analysis, which reduces the number of feature dimensions, can make it possible to choose the best features. In skin cancer detection, the effectiveness of the three ANN learning algorithms, Levenberg–Marquardt, robust back-propagation, and scaled conjugate gradient, was also investigated recently.²⁰ Levenberg–Marquardt’s learning algorithm successfully identified benign lesions even as the epoch number increased. With a sensitivity of 92.6%, the scaled conjugate gradient learning algorithm generated better outcomes in the interim. Furthermore, developing a mole categorization system increased the feasibility of detecting melanoma skin cancer early.²¹ The asymmetrical border, color, and diameter (ABCD) rule can be applied to determine how the lesions appeared. Indeed, asymmetry and edges of a lesion can be identified using the Mumford–Shah algorithm and the Harris–Stephen method.²² Such analyses can be performed on moles to determine their benignity. Thus, it may be benign if a mole is black, cinnamon, or brown. As melanoma moles are often bigger, the diameter threshold for melanoma detection may be set at 6 mm. Back-propagation feed-forward ANN can divide moles into three classes as follows: common, unusual, and melanoma moles. Such automated skin cancer diagnosis systems have been reported.²³ A “maximum entropy thresholding” technique has also been used to classify lesions into carcinogenic and noncancerous using a gray-level co-occurrence matrix for extraction of visual traits.²⁴ Similarly, the ANN-based classifier resulted in an accuracy rate of about 88%.²⁵

CNN methods

CNN is frequently used to identify, gather, and organize images as it can generate more complex features such as corners using information pertaining to angles from curves and edges.²⁶ Before the completely integrated layers in a CNN, there may be numerous convolutional, pooling, and fully connected hidden layers. Figure 4 depicts the working scheme of the CNN architecture for image classification, in which the process begins with a raw image input, preprocessing, augmentation, and normalization. Thereafter, convolutional layers of the CNN framework extract image features, while the pooling layers are used for reducing the dimensionality, and the fully connected layers are involved in the final classification of images.²⁷ Therefore, CNN brings about effective feature learning and accurate classification in computer vision tasks.²⁸ Its architecture for image classification includes the input first, followed by several convolutional layers with ReLU activation, pool layers for dimensionality reduction, full convolution layers, followed by fully connected layers with softmax activation and classification labels (Fig. 5). This structure efficiently extracts hierarchical features from images and can be used for accurate classification in DL-based skin cancer diagnoses. In recent years, several studies have used CNNs to identify skin cancer (Table 4). The data can be grouped using techniques such as support vector machine (SVM) and softmax classifiers. Without segmentation, melanoma can be correctly grouped with an accuracy of 82.8% and 85.5% with segmentation. A multiscale CNN from an inception v3 DNN was trained on ImageNet to further aid in the detection of melanoma. Classification of skin cancer was significantly improved using coarse-scale (lesion environment and form as a whole) and finer scale (lesion textual descriptions) parameters for high-resolution images of skin lesions.²⁹ Diagnosis of skin lesions involves extracting properties from deep CNNs that have already received training. Some of the trained versions used as deep feature generators include Alex Net, ResNet-18, and VGG16.³⁰ For culminating image sorting, combining all the classifiers’ outputs is required.³¹ Such protocols are effective in classifying seborrheic keratosis (SK) and melanoma images with high accuracy.³⁰ A pretrained version of ResNet-152 was used as the foundation for a deep CNN architecture for grouping 12 different types of skin lesions. Similarly, after extracting features from a trained deep CNN termed AlexNet, an error-correcting output coding SVM has been used for skin lesion classification, resulting in scores for sensitivity and specificity.³² Identification of squamous cell carcinoma, actinic keratosis, and basal cell carcinoma can have high accuracy rates (of > 85%) using CNN designs.³

FIG. 4.

Basic scheme of convolutional neural network (CNN) architecture.

FIG. 5.

Basic convolutional neural network (CNN) architecture.

Table 4.

A Comparative Analysis of Skin Cancer Detection Using Convolutional Neural Network-Based Approaches

Reference	Diagnoses	Classifier	Results (accuracy)
³⁵	Benign/malignant	Light Net is a deep learning-based method for classifying objects.	Accuracy (81.6)
³⁶	Melanoma/benign	Convolutional neural network (CNN)	Accuracy (81),
³⁴	Basal cell carcinoma (BCC)/ squamous cell carcinoma (SCC)/melanoma/actinic keratosis (AK)	Support vector machine (SVM), CNN	Accuracy (SCC: 95.1, AK: 98.9, BCC: 94.17)
³⁷	Melanoma	Deep CNN	Accuracy (72.1)
³²	Malignant melanoma and BCC	CNN using the architecture of ResNet 152	AUC (melanoma: 96, BCC: 91)
³⁸	Melanoma/nonmelanoma	SVM-trained CNN was employed	Accuracy (93.75)
³⁴	Seborrheic keratosis (SK)	CNN	Average AUC:9
³⁸	BCC/non-BCC	CNN	Accuracy (95.93)
³⁸	Cancerous/noncancerous	CNN	Accuracy (86.67)
³	Benign/melanoma	Visual geometry group 16 (VGG-16) and CNN	Accuracy (78)
³⁹	Malignant/benign	CNN	Accuracy (89.5)

CNN, convolutional neural network; SVM, support vector machine; AUC, area under the ROC curve.

KNN models

Similar to CNNs, KNNs do not rely on external aid because they learn via unsupervised learning, requiring only a basic understanding of the properties of the input data.³³ The two-dimensional plane of KNN typically has two levels, an input layer and an output layer, connected via bridges.³⁴ Figure 6 illustrates a decision-making process based on KNN. A message is encoded into several embeddings, which are stored in different storages. Each store retrieves a set of nearest neighbors, and the scoring function processes them to decide the final outcome. This structure is typically used in recommendation systems, anomaly detection, and natural language processing applications. Without knowing the relationships between the individuals in the input data set, a KNN can be used to cluster data.³⁵ Each node in the competitive layer must act as both an input and an output since a KNN lacks a specific output layer. A KNN preserves the topological structure of the initial data space as it transitions from high to low dimensions. The “preservation” distance calculates how evenly distributed the information is in a given area.³⁶ By mapping the data points farther apart, this method considers the distance between data points in the input data space. Hence, KNN is the best method for handling enormous amounts of data. The generalizability of KNN is another important feature. Unknown incoming data can be recognized and categorized by the network as KKN can map complex, nonlinear connections between data points effectively.³⁷

FIG. 6.

Basic k-nearest neighbor (KNN) architecture.

Due to these advantages, KNNs have been used in many skin cancer detection systems (Table 5). A KNN-based detection was used for synchronized fluorescence spectra from melanoma, nevus, and healthy skin samples to train an NN. The spectra’s high dimensionality was significantly reduced via principal component analysis (PCA).³⁶ Similarly, KNN and ANN have been used to identify melanomas, with the error rate being 3%–4% on the test dataset.³⁸ Basal cell carcinoma, melanoma, and squamous cell carcinoma may be identified using a self-organizing NN and a radial basis function NN.³⁹ Skin lesion images have been categorized using color, GLCM, and shape attributes, using KNN, ANN, and Naive–Bayes classifiers.⁴⁰ An automated method based on KNN to lessen ambient noise in the filtered images was used to recognize skin cancer with an exceptionally high accuracy.⁴¹ A similar KNN-based method has been compared with SVM, back propagation neural network (BPN), and 3-layer NN algorithms for its utilities in lesion classification performance and found to be highly accurate.⁴²

Table 5.

A Comparative Analysis of Skin Cancer Detection Using k-Nearest Neighbor-Based Approaches

Ref	Diagnoses	Classifier	Description	Results (%)
⁴⁹	Melanoma	Combining feed-forward self organizing map (SOM) and neural network (NN)	Principal component analysis (PCA) can be utilized to reduce the dimensionality of a spectrum	Accuracy (96–98)
⁴⁵	Basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and melanoma	SOM and radial basis function (RBF)	15 characteristics were generated via the extraction of morphological and color features from the GCM	Accuracy (93.15)
⁴⁷	Cancerous	Modified k-nearest neighbor (KNN)	The extraction of feature automatic threshing technique was used to separate the grains	Accuracy (98.3)

GAN-based techniques

Deep CNNs such as GAN are powerful DNNs built on the idea of negligible games. Two NNs, discriminator and generator, compete to learn and record available data.⁴³ The generator module creates highly biased data, and the discriminator component differentiates between fake and real image.⁴⁴ Both networks will carry out these tasks repeatedly during training.⁴⁵ Figure 7 shows a GAN comprising two networks as follows: the generator network, which takes real images and creates fake images, and the discriminator network, which tries to determine whether an image is real or fake. Each evaluates images from the other, for concerted improvement of the capabilities of image generation and distinguishment of both networks. Overall, the generator becomes better at fooling the discriminator, and the discriminator becomes better at telling the real from the fake. The main benefit of GAN is its capacity to produce fakes, including photorealistic images, with the same data distribution as real samples.⁶ It can aid DL in overcoming the deficiency of good training examples. Figure 8 highlights working of a GAN consisting of two NNs, the generator that generates fake samples from noise, and the discriminator that can differentiate real from fake samples. The discriminator is trained using real data as well as generated samples, which is fed back into the system via back-propagation. Both the networks are controlled by loss functions, so that the generator keeps improving the capability of generating real samples with time.

FIG. 7.

Basic generative adversarial network (GAN) architecture.

FIG. 8.

Working scheme of the generative adversarial network (GAN)-based image differentiation.

Studies have evaluated the effectiveness of GAN models in skin lesion detection (Table 6). For example, a GAN-based technique was used for training to generate fake images resembling real skin lesions. The CNN-based discriminator module was then used as a classifier for categorizing skin lesions into seven different categories.⁴⁶ GAN models have also been compared with benchmarks such as ResNet-50 and DenseNet for their effectiveness in the detection of skin cancer and found to result in increased accuracy of classifying dermoscopy images as melanoma, SK, or nevus.⁴⁷ Ahmad et al. have suggested a new approach for characterizing skin lesions using self-attention progressive GAN with significantly increased accuracy of detection.⁴⁸

Table 6.

A Comparative Analysis of Skin Cancer Detection Using Generative Adversarial Network-Based Approaches

Ref	Skin cancer diagnoses	Classifier and training algorithm	Description	Results (%)
⁶	BCC	Generative adversarial network (GAN)	The suggested system used convolutional neural network (CNN) for the generator and a deep & cross network (DCN) for the discriminator	Accuracy (86)
⁸	Melanoma	Deep convolutional generative adversarial network (DC-GAN)	Heterogeneous adaptive networks for feature extraction using deep convolutional layers	ROC AUC (92), accuracy (86.4)
⁵⁵	BCC	Self-attention-based probabilistic generative adversarial network (PGAN)	A stability strategy was used to improve a generative model	Accuracy (70.1)

Databases

Reliable image collection is indispensable for evaluating skin cancer diagnostic effectiveness of ML/DL techniques.⁴⁹ However, insufficiency of data types and numbers make this difficult. This section discusses the datasets of lesion images, which can be used for skin cancer detection. Research incorporating new datasets should fill important gaps regarding dataset variability and patient representation to enhance technological reliability, clinical applicability, and predictive accuracy for DL models that detect skin cancer.

PH2

Dermoscopic images (∼200 typical and atypical nevi, and melanoma) of PH2 dataset⁴⁹ are based in Pedro Hispano Hospital Dermatology Center, Portugal. The pictures are taken by Tuebinger-Mole-Analyzer at 20 × magnification under similar illumination. The images in PH2 are 768 × 560 pixels with 8 bit of each RGB color. Images can be evaluated for features such as stripes, colors, melanin networks, regression zones, and blue-white veil globules.⁵⁰

International Skin Imaging Collaboration

The ISIC repository is another database initiated in 2016 at the International Symposium on Biomedical Imaging.⁴² Training and testing images of benign nevi, SH, and malignant melanomas are organized separately.⁵¹ ISIC regularly updates its image library, and includes additional information such as patient demographic data.

Dermatology Information System

Dermatology information system (DermIS)⁵² was created as a collaborative effort of the Universities of Erlangen and Heidelberg. It contains >6500 lesion images and has been used to create two new online image atlases for dermatology and pediatric dermatology.

DermQuest

DermQuest dataset⁵³ harbors images in excess of 22,000. It has lesion tags for images; however, it was transferred to Derm101 in 2018, and since December 31, 2019, it is no longer in use.

Ham10000

The HAM10000 dataset⁴² includes >10,000 dermatoscopic images of numerous lesion types, enabling researchers to create durable predictive models to properly identify disorders in underrepresented groups, including people with dark skin.

DermNet NZ

The DermNet NZ database⁵⁴ expands DL model diagnostic abilities through its extensive coverage of dermatological conditions, including complex and rare lesions, even though skin cancer remains its secondary focus.

MED-NODE

The MED-NODE dataset⁵⁵ functions as an essential resource for algorithm testing under controlled conditions because of its precise melanoma focus and its emphasis on high-quality data curation.

PAD-UFES-20

PAD-UFES-20 dataset⁵⁶ that includes clinical images taken from smartphones, alongside dermoscopic photographs of multiple skin conditions, may boost model generalization by presenting algorithms to different lesion characteristics among varied patient groups.

XiangyaDerm

XiangyaDerm database⁵⁷ stores >150,000 skin images, mainly from individuals of Asian ethnicity. In addition to adding racial diversity for skin cancer diagnoses, the repository in XiangyaDerm encompasses >570 skin diseases and has been shown to efficiently aid in the identification of multiple dermal conditions, including psoriasis.⁵⁸

Discussion

Recent progresses in the applications of DL algorithms for skin lesion detection/categorization are commendable. The DL methodologies achieved exceptional diagnostic effectiveness in detecting skin lesions through automated scalable and noninvasive solutions. CNNs have demonstrated exceptional performance in handling high-resolution pictures while finding intricate features that particularly suit image-based classification needs. GANs resolve the elimination of medical data by creating artificial images to develop stronger training abilities as well as exemplary model results. Overall, DL models eliminate human dependency in diagnostic work, which may enhance error-free standardized outcomes that match data from different sources. Early skin cancer detection benefits substantially from these techniques because they excel at recognizing minor skin lesion patterns to enhance patient treatment results.

However, there are significant obstacles that must be addressed before widespread clinical implementation of DL methods for skin cancer diagnoses. A notable hurdle is the requirement for time- and cost-intensive extensive training, which in turn relies on powerful software, extensive hardware, and computing facilities.⁵⁹ These are all challenging for clinicians and hospital staff, particularly considering their cost-ineffectiveness, hindering their widespread use under clinical settings. Computational needs are among the major drawbacks of DL models because effective deployment and technical know-how require advanced hardware and software systems, which limit their accessibility, especially in resource-limited areas. Another critical aspect is that these models suffer from the absence of diverse and high-resolution images that can severely hamper the generalizability of the models. Lesions can be quite heterogeneous and come have different geometries and sizes, and many images lack diagnostic clarity and resolution. This is a significant challenge for detecting small lesions (1–2 mm).⁶⁰ In fact, lesion heterogeneity with regard to their geometries, sizes, and background is a significant hurdle that compromises the accuracy of detection/prediction results. Identification of lesion shapes is equally critical.⁶¹ Similarly, skin color is a critical issue as most databases are overrepresented with fair-skinned subjects. Indeed, databases lean toward a more favorable complexion population and neglect the darker complexion population, affecting the diagnostic precision widespread use across different populations. Hence, accurate training of detection systems and identification of lesions in individuals of color need recognition of background skin color.⁶² This, in turn, requires larger samples of both dark- and light-skinned subjects to be included in the databases. In addition, available dermoscopy images of these rare types of skin cancer and age group information are not completely available. Most of these limitations must be addressed while taking DL-based diagnostic systems ahead. Likewise, the insufficiency of dermoscopy images of rare lesion types in databases is another limitation.⁵¹ Another point of concern is the age of subjects; several skin cancer types, such as Merkel, basal, and squamous cell carcinomas, are more prevalent in people older than 65 years.⁶³ An adequate number of lesion images from this age group is required for effective NN-based training and detection. Similarly, skin cancer types are strongly influenced by a plethora of environmental and genetic factors,^64,65 which must be represented and taken into account while using DL/ML models. Finally, patient data privacy and unconsented sharing of lesion images are another issue that must be addressed for implementation of DL methods for skin cancer diagnoses.

Many NNs exist as uninterpretable “black boxes” that prevent clinicians from understanding the basis of their diagnostic decisions. Model robustness faces challenges because data reliability problems lead to degraded performance, while encounters with uncommon lesion presentations present algorithm difficulties. The evaluation of DL techniques shows their ability to reinvent skin cancer diagnosis. Yet, additional work on dataset biases alongside computational hurdles and interpretability barriers is required to achieve broad clinical acceptance and equitable health care results.

Conclusion and Future Perspectives

The review discusses the ability of various DL-based NN algorithms, mainly ANN, CNN, KNN, and GAN, to identify/categorize skin cancer lesions. Such characterizations are particularly interesting as they are noninvasive and do not require surgery or biopsy. The benefits are accuracy in diagnosis, auto-categorization of lesions, and minimization of reliance on interventional procedures and human subjection. Using preprocessing, image segmentation, feature extraction, and classification techniques of ML/DL, it is possible to identify and characterize skin cancers with high accuracy. While each algorithm has benefits and drawbacks, it is crucial to choose the best classification strategy for optimum performance, and develop it further. Because CNN is more closely related to computer vision than other NN types, it is more effective at classifying images and predicting whether a lesion is cancerous or not. While CNNs are highly effective for image-based lesion classification, GANs can be beneficial with a dearth of data by generating synthetic images. Hence, DL models using CNNs alongside GANs may demonstrate a stronger potential for delivering precise noninvasive skin cancer diagnoses. DL method advancement must include work on diverse datasets that use high-resolution inputs while implementing unsupervised learning methods to maximize feature detection ability and overall model consistency. Unsupervised ML techniques such as auto-organization may be used to find features, connections, or patterns in image datasets to enhance the accuracy of their feature representations.⁶⁶ Such image processing systems could be improved through research so that even minute features may be identified. In addition, it is essential to improve computational efficiency and minimize dataset biases to achieve broader clinical deployment. Developing interpretable automated models may enhance acceptance of the prediction results by medical practitioners and clinicians, thereby extending the use of computerized systems for actual clinical implementation. In conclusion, despite the significant progresses, further steps to address the challenges associated with diversity in the datasets, computational complexity, and underrepresentation of some populations, are warranted to diversify image acquisition and improve the protocols.

Footnotes

Acknowledgments

The authors gratefully acknowledge the funding of the Deanship of Graduate Studies and Scientific Research, Jazan University, Saudi Arabia, through Project Number: RG24-L06. The authors, S.H., D.M.M. and A.B. are grateful to Jazan University, Saudi Arabia, for providing access of the Saudi Digital Library for this work.

Authors’ Contributions

S.H.: Conceptualization, methodology, formal analysis, investigation, resources, writing—original draft, writing—review and editing, visualization, supervision, and project administration. F.A.: Conceptualization, methodology, formal analysis, investigation, writing—original draft, writing—review and editing, and visualization. V.S.: Conceptualization, methodology, formal analysis, investigation, writing—original draft, writing—review and editing, and visualization. D.M.M.: Conceptualization, methodology, formal analysis, investigation, writing—original draft, writing—review and editing, and visualization. A.B.: Conceptualization, methodology, formal analysis, investigation, writing—original draft, writing—review and editing, and visualization.

Disclosure Statement

The authors declare no conflicts of interests.

Funding Information

Deanship of Graduate Studies and Scientific Research, Jazan University, Saudi Arabia – Project No. RG24-L06.

References

Ashraf

, Afzal

, Rehman

, et al. Region-of-Interest based transfer learning assisted framework for skin cancer detection. IEEE Access, 2020; 8:147858–147871; doi: 10.1109/ACCESS.2020.3014701

Cives

, Mannavola

, Lospalluti

, et al. Non-Melanoma skin cancers: Biological and clinical features. Int J Mol Sci, 2020; 21(15):143–155; doi: 10.3390/ijms21155394

Dildar

, Akram

, Irfan

, et al. Skin cancer detection: A review using deep learning techniques. Int J Environ Res Public Health, 2021; 18(10):5479; doi: 10.3390/ijerph18105479

Khan

, Hussain

, Rehman

, et al. Classification of melanoma and nevus in digital images for diagnosis of skin cancer. IEEE Access, 2019; 7:90132–90144; doi: 10.1109/ACCESS.2019.2926837

Almansour

, Jaffar

. Classification of Dermoscopic skin cancer images using color and hybrid texture features. Int J Comput Sci Netw Secur, 2016; 16(4):135–139.

Rashid

, Tanveer

, Khan

. Skin Lesion Classification Using GAN Based Data Augmentation. In: 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) IEEE; 2019; pp. 916–919; doi: 10.1109/EMBC.2019.8857905

Codella

NCF

, Nguyen

Q-B

, Pankanti

, et al. Deep learning ensembles for melanoma recognition in dermoscopy images. IBM J Res & Dev, 2017; 61(4/5):5:1–5:15; doi: 10.1147/JRD.2017.2708299

, Sarda

, Gupta

, et al. Use of artificial intelligence in dermatology. Indian J Dermatol, 2020; 65(5):352–357; doi: 10.4103/ijd.IJD_418_20

Nawaz

, Mehmood

, Nazir

, et al. Skin cancer detection from dermoscopic images using deep learning and fuzzy k ‐means clustering. Microsc Res Tech, 2022; 85(1):339–351; doi: 10.1002/jemt.23908

10.

Hosny

, Kassem

. Refined Residual Deep Convolutional Network for Skin Lesion Classification. J Digit Imaging, 2022; 35(2):258–280; doi: 10.1007/s10278-021-00552-0

11.

Allugunti

. A machine learning model for skin disease classification using convolution neural network. Int J Comput Programming Database Manage, 2022; 3(1):141–147.

12.

Dongare

, Kharde

, Kachare

. Introduction to artificial neural network. Int J Eng Innov Technol, 2012; 2(1):189–194; doi: 10.1016/j.compeleceng.2022.108259

13.

Tito Anand

, Anandakumar

, Pare

, et al. Modeling of process parameters to predict the efficiency of shallots stem cutting machine using multiple regression and artificial neural network. J Food Process Engineering, 2022; 45(6):e13944; doi: 10.1111/jfpe.13944

14.

Al-Azzawi

, Chenchamma

, Hamad

, et al. Use of Radiation Circuits for Diagnosis of Melanoma Skin Cancer in Images of Skin Lesions Using Convolutional Neural Networks. J Nanomater, 2022; 2022(1):1–7; doi: 10.1155/2022/8135715

15.

Shetty

, Shah

, Reddy

, et al. Skin Cancer Detection Using Image Processing: A Review. In: Proceedings of the 2nd International Conference on Recent Trends in Machine Learning, IoT, Smart Cities and Applications. Lecture Notes in Networks and Systems, Vol 237 2022; pp. 103–121; doi: 10.1007/978-981-16-6407-6_11

16.

Malibari

, Alzahrani

, Eltahir

, et al. Optimal deep neural network-driven computer aided diagnosis model for skin cancer. Comput Electr Eng, 2022; 103:108318; doi: 10.1016/j.compeleceng.2022.108318

17.

Arif

, Philip

, Ajesh

, et al. Automated detection of nonmelanoma skin cancer based on deep convolutional neural network. J Healthc Eng, 2022; 2022:1–15; doi: 10.1155/2022/6952304

18.

Wenger

, Tirdad

, Dela Cruz

, et al. A semi-supervised learning approach for bladder cancer grading. Mach Learn with Appl, 2022; 9:100347; doi: 10.1016/j.mlwa.2022.100347

19.

Tabrizchi

, Parvizpour

, Razmara

. An improved VGG model for skin cancer detection. Neural Process Lett, 2023; 55(4):3715–3732; doi: 10.1007/s11063-022-10927-1

20.

Shirke

, Gadekar

. A reviewed study of deep learning techniques for the early detection of skin cancer. Tianjin Daxue Xuebao (Ziran Kexue yu Gongcheng Jishu Ban)/. J Tianjin Univ Sci Technol, 2022; 55:2022; doi: 10.1760/5/OSF.IO/7WZ9C

21.

Kaur

, Kaur

. Analysis of melanoma cancer detection techniques. SN Comput Sci, 2022; 3(1):93; doi: 10.1007/s42979-021-00955-4

22.

Kavitha

, Prasad

, Sujeeth

, et al. A perspective study on scalable computation model for skin cancer detection: Advancements and challenges. Scpe, 2025; 26(1):48–63; doi: 10.1269/4/scpe.v26i1.3501

23.

Verma

, Gautam

. Advanced Method for Melanoma and Nonmelanoma Skin Cancer Detection Using Convolutional Neural Network:. A Survey., n.d.

24.

Khanra

, Kuila

, Patra

, et al. Survey on Computational Techniques for Pigmented Skin Lesion Segmentation. Opt Mem Neural Networks, 2022; 31(4):333–366; doi: 10.3103/S1060992X2204004X

25.

Kar

, Siddiqui

, Akhtar

. An Improved CNN-Based Feature Extraction Algorithm for Identifying Pneumonia from Chest X-Ray Images. In: Intelligent Computing & Optimization. ICO 2022. Lecture Notes in Networks and Systems, Vol 569 2023; pp. 507–517; doi: 10.1007/978-3-031-19958-5_48

26.

Shrimali

. Development of a Mobile Application for the Early Detection of Skin Cancer Using Image Processing Algorithms, Transfer Learning, and AutoKeras. In: 2022 5th International Conference of Computer and Informatics Engineering (IC2IE). IEEE; 2022; pp. 100–105; doi: 10.1109/IC2IE56416.2022.9970048

27.

Afridi

, Alam

, Khan

, et al. A Multimodal Memes Classification: A Survey and Open Research Issues. In: Innovations in Smart Cities Applications. Springer; Volume 4 2021; pp. 1451–1466; doi: 10.1007/978-3-030-66840-2_109

28.

Zhang

, Cai

Y-X

, Wang

Y-Y

, et al. Skin cancer diagnosis based on optimized convolutional neural network. Artif Intell Med, 2020; 102:101756; doi: 10.1016/j.artmed.2019.101756

29.

Kumar

, Sinha

. Skin Cancer Classification for Dermoscopy Images Using Model Based on Deep Learning and Transfer Learning. In: Computational Intelligence and Data Analytics. Lecture Notes on Data Engineering and Communications Technologies, Vol 142 2023; pp. 257–271; doi: 10.1007/978-981-19-3391-2_19

30.

Mendes

, da Silva

. Skin Lesions Classification Using Convolutional Neural Networks in Clinical Images. arXiv, 20181812.02316;; doi: 10.4855/0/arXiv.1812.02316

31.

Hasan

, Barman

, Islam

, et al. Skin Cancer Detection Using Convolutional Neural Network. In: Proceedings of the 2019 5th International Conference on Computing and Artificial Intelligence ACM: New York, NY, USA; 2019; pp. 254–258; doi: 10.1145/3330482.3330525

32.

Dorj

U-O

, Lee

K-K

, Choi

J-Y

, et al. The skin cancer classification using deep convolutional neural network. Multimed Tools Appl, 2018; 77(8):9909–9924; doi: 10.1007/s11042-018-5714-1

33.

Shalu

. A Color-Based Approach for Melanoma Skin Cancer Detection. In: 2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC) IEEE; 2018; pp. 508–513; doi: 10.1109/ICSCCC.2018.8703309

34.

Edupuganti

. Intelligent Signatures Using Embeddings and K-Nearest-Neighbors. 2020.

35.

Murugan

, Nair

SAH

, Kumar

KPS

. Detection of Skin Cancer Using SVM, Random Forest and kNN Classifiers. J Med Syst, 2019; 43(8):269; doi: 10.1007/s10916-019-1400-8

36.

Vidya

, Karki

. Skin Cancer Detection Using Machine Learning Techniques. In: 2020 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) IEEE; 2020; pp. 1–5; doi: 10.1109/CONECCT50063.2020.9198489

37.

Manasa

, Murthy

DGV

. Skin cancer detection using VGG-16. Eur J Mol Clin Med, 2021; 8(1):1419–1426.

38.

Nida

, Irtaza

, Javed

, et al. Melanoma lesion detection and segmentation using deep region based convolutional neural network and fuzzy C-means clustering. Int J Med Inform, 2019; 124:37–48; doi: 10.1016/j.ijmedinf.2019.01.005

39.

Mengistu

, Alemayehu

. Computer Vision for skin cancer diagnosis and recognition using RBF and SOM. Int J Image Process, 2015; 9(6):311–319.

40.

Adegun

, Viriri

. Deep learning techniques for skin lesion analysis and melanoma cancer detection: A survey of state-of-the-art. Artif Intell Rev, 2021; 54(2):811–841; doi: 10.1007/s10462-020-09865-y

41.

Sajid

, Rajesh

. Performance evaluation of classifiers for automatic early detection of skin cancer. J Adv Res Dyn Control Syst, 2018; 10:454–461.

42.

Codella

NCF

, Gutman

, Celebi

, et al. Skin Lesion Analysis toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC). In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE; 2018; pp. 168–172; doi: 10.1109/ISBI.2018.8363547

43.

Qin

, Liu

, Zhu

, et al. A GAN-based image synthesis method for skin lesion classification. Comput Methods Programs Biomed, 2020; 195:105568; doi: 10.1016/j.cmpb.2020.105568

44.

Hirte

, Platscher

, Joyce

, et al. Realistic generation of diffusion-weighted magnetic resonance brain images with deep generative models. Magn Reson Imaging, 2021; 81:60–66; doi: 10.1016/j.mri.2021.06.001

45.

Little

, Elliot

, Allmendinger

, et al. Generative Adversarial Networks for Synthetic Data Generation: A Comparative Study. arXiv, 2021; doi: 10.4855/0/arXiv.2112.01925;2112.01925;

46.

Ren

, Guo

, Yu

, et al. Improve Image-Based Skin Cancer Diagnosis with Generative Self-Supervised Learning. In: 2021 IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE). IEEE; 2021; pp. 23–34; doi: 10.1109/CHASE52844.2021.00011

47.

Medi

, Nemani

, Pitta

, et al. SkinAid: A GAN-Based Automatic Skin Lesion Monitoring Method for IoMT Frameworks. In: 2021 19th OITS International Conference on Information Technology (OCIT) IEEE; 2021; pp. 200–205; doi: 10.1109/OCIT53463.2021.00048

48.

Ahmad

, Jun

, Palade

, et al. Improving skin cancer classification using heavy-tailed student t-distribution in generative adversarial networks (TED-GAN). Diagnostics (Basel), 2021; 11(11):2147; doi: 10.3390/diagnostics11112147

49.

Mendonça

, Celebi

, Mendonca

, et al. Ph2: A public database for the analysis of dermoscopic images. Dermoscopy Image Anal, 2015; 2:1–24; doi: 10.1016/S2589-7500(21)00252-1

50.

Senan

, Jadhav

, Kadam

. Classification of PH2 Images for Early Detection of Skin Diseases. In: 2021 6th International Conference for Convergence in Technology (I2CT) IEEE; 2021; pp. 1–7; doi: 10.1109/I2CT51068.2021.9417893

51.

Milton

MAA

. Automated skin lesion classification using ensemble of deep neural networks in ISIC 2018: Skin lesion analysis towards melanoma detection challenge. 2019; doi: 10.4855/0/arXiv.1901.10802

52.

Brooks

, Ashton

, Pethybridge

. DERMIS: A computer system for assisting primary-care physicians with dermatological diagnosis. Br J Dermatol, 1992; 127(6):614–619; doi: 10.1111/j.1365-2133.1992.tb14875.x

53.

Liao

, Li

, Luo

. Skin Disease Classification versus Skin Lesion Characterization: Achieving Robust Diagnosis Using Multi-Label Deep Neural Networks. In: 2016 23rd International Conference on Pattern Recognition (ICPR) IEEE; 2016; pp. 355–360; doi: 10.1109/ICPR.2016.7899659

54.

Jalaboi

, Faye

, Orbes-Arteaga

, et al. DermX: An end-to-end framework for explainable automated dermatological diagnosis. Med Image Anal, 2023; 83:102647; doi: 10.1016/j.media.2022.102647

55.

Giotis

, Molders

, Land

, et al. MED-NODE: A computer-assisted melanoma diagnosis system using non-dermoscopic images. Expert Syst Appl, 2015; 42(19):6578–6585; doi: 10.1016/j.eswa.2015.04.034

56.

Pacheco

AGC

, Lima

, Salomão

, et al. PAD-UFES-20: A skin lesion dataset composed of patient data and clinical images collected from smartphones. Data Brief, 2020; 32:106221; doi: 10.1016/j.dib.2020.106221

57.

Xie

, He

, Zhao

, et al. XiangyaDerm: A Clinical Image Dataset of Asian Race for Skin Disease Aided Diagnosis. In: Large-Scale Annotation of Biomedical Data and Expert Label Synthesis and Hardware Aware Learning for Medical Imaging and Computer Assisted Intervention 2019; pp. 22–31; doi: 10.1007/978-3-030-33642-4_3

58.

Huang

, Jiang

, Li

, et al. The classification of six common skin diseases based on xiangya-derm: Development of a chinese database for artificial intelligence. J Med Internet Res, 2021; 23(9):e26025; doi: 10.2196/26025

59.

Argenziano

, Soyer

, De Giorgi

, et al. Interactive Atlas of Dermoscopy. 2000.

60.

Stanley

, Stoecker

, Moss

, et al. Color and Structural Features for Automatic Skin Lesion Discrimination in Dermoscopy Images. In: Proceedings of the 6th World Congress on Melanoma, Vancouver, BC, Canada (Pp. 6-10). 2005; pp. 6–10.

61.

, Chen

, Dou

, et al. Automated melanoma recognition in dermoscopy images via very deep residual networks. IEEE Trans Med Imaging, 2017; 36(4):994–1004; doi: 10.1109/TMI.2016.2642839

62.

Brinker

, Hekler

, Utikal

, et al. Skin Cancer Classification Using Convolutional Neural Networks: Systematic Review. J Med Internet Res, 2018; 20(10):e11936; doi: 10.2196/11936

63.

Garcovich

, Colloca

, Sollena

, et al. Skin cancer epidemics in the elderly as an emerging issue in geriatric oncology. Aging Dis, 2017; 8(5):643–661; doi: 10.1433/6/AD.2017.0503

64.

Parker

. The influence of climate change on skin cancer incidence – A review of the evidence. Int J Womens Dermatol, 2021; 7(1):17–27; doi: 10.1016/j.ijwd.2020.07.003

65.

Wong

QYA

, Chew

. Defining skin aging and its risk factors: A systematic review and meta-analysis. Sci Rep, 2021; 11(1):22075; doi: 10.1038/s41598-021-01573-z

66.

Taye

. Understanding of machine learning with deep learning: Architectures, workflow, applications and future directions. Computers, 2023; 12(5):91; doi: 10.3390/computers12050091