CNN-based classification models must be trained with each source image having the same size. Mechanisms like cropping, resizing, and rescaling can be used to combine the dataset configurations in order to accomplish this. When new and distinct instances are added to the dataset for training, data augmentation makes machine learning models more effective. To ensure that the architecture runs effectively, there needs to be an adequate repository that can deliver more data and has several desirable features. Using data augmentation approaches to incorporate variations that the model typically encounters in practice could strengthen machine learning models and increase the likelihood of issues such as over-fitting. Additionally, the data are unbalanced because the number of instances of various skin disorders is higher than that of melanoma. Data augmentation can be used, for example, to balance a dataset. Several studies have employed widely used methods for augmenting data, including scaling [86–90], flipping, rotation, and cropping [85–90]. To reduce the noise in the dataset, researchers employed a range of methods, including guided filters, Gaussian low-pass filters, blur filters, motion blur, noise addition, and histogram equalization [85,88,91,92]. Adjustments in color and brightness have also been made in [70,91]. After training generative adversarial networks (GANs) on digital images of cutaneous lesions, the researchers presented a DL approach for controlling variable background illumination [93]. These days, researchers are using DL layouts a lot to identify skin lesions from medical imaging. The lighting in images of skin lesions may be adjusted with the use of these deep learning techniques. In order to correct for variable background intensity in dermatological images, the study introduces a novel transform known as IECET [94].

Segmentation is an approach that partitions the picture into numerous segments or pixels for the purpose of enhancing image analysis [95]. It facilitates identifying tumors from an analyzed picture easier. However, it is essential to recall that image segmentation is a complicated procedure that may necessitate further processing. The 2 leading approaches to segmentation are K-means clustering and Otsu’s approach [75,96,97]. Sagar et al. [65] implied a color channel-based technique that merged an edge detection mechanism with a modified 2D Otsu approach.

The researchers report a 93.7% accuracy rate using their suggested strategy for lesion identification from a source image. A number of researchers used the K-means clustering approach with K = 2 for segmentation [82,83]. Another study suggested combining Otsu thresholding with K-means, which produced accuracy and AUC values of 0.91% and 89.07%, respectively [98]. An improved K-means variation with a 94% accuracy rate has been suggested in a different study [83]. In addition to these methods, image segmentation in [72] was done using Chan Vese’s active contour approach [99]. In [64,100], a novel segmentation technique combining Chan level set border detection with rapid independent component analysis (FastICA) has been demonstrated. A novel method that combines specific morphological analysis with Chan Vese active contour computing was put forth by Roberta et al. [79]. The TDR value outperformed the Otsu technique’s results while using the same set of data [101]. Furthermore, Sabouri et al. segmented photographs using CNN boundary-identification technology, producing an 86.67% Jaccard Index score [102].

Using GAN and a U-Net generator, Udrea et al. identified and suggested pigmented lesions with 91.40% accuracy [103]. Another research effort [104] used a U-Net with four subsampling layers to segment medical images. The method was first proposed in [105], where 92%, 98%, and 95% percent accuracy, specificity, and sensitivity were obtained, respectively. Another research study [106] has demonstrated a fuzzy C-means clustering algorithm with sensitivity, specificity, and accuracy of 90.02%, 99.15%, and 95.69%, respectively. A strategy called interactive object recognition was used in another investigation for the segmentation [107]. A further study has presented a unique segmentation approach that uses edge detection, thresholding, and evaluation of a correlated element of the mask to extract geometric features based on ABCD [108]. For determining the threshold level from medical macro photos, a CNN method called DTP-Net was introduced [109].

The processes of feature extraction and selection are fundamental components of machine learning. The latest studies on the diagnosis of skin cancer include a variety of attributes, particularly distinct aspects derived from the dermatologist’s ABCD rule. The feature selection approach is usually utilized to limit the feature set’s size. However, a number of research employed DNN and charged CNNs with feature extraction. PSLs are evaluated using the ABCD rule; further assessment by a specialist is required for these lesions [110]. By computing asymmetry, the authors in [101] were able to estimate the lesion’s major and minor axes [111]. The region of the lesion was split into 2 parts, employing an axis derived from the longest diagonal vector corresponding to the Euclidean distance of the affected area [112].

The edges of the lesion, deformity, four features based on the radius [113], asymmetric edge indices [114], compactness index [115], Heywood circularity index, mean curvature [116], best-fit ellipse indices [117], bulkiness index [118], bending energy, area, perimeter of the convex hull, convexity index, indentation and protrusion index, and fractal dimensions were among the features extracted for border irregularity. In the three color channels, the average gradient magnitude and variance in the lesion expanded rim were retrieved in [119]. The gray-level co-occurrence matrix (GLCM) [120,121] has been employed by numerous authors for acquiring textural information [82,98,101,122] and is, therefore, the most frequently employed method. Furthermore, the authors in a distinct study employed the gray-level run length matrix (GLRLM) [123]. For acquiring the textural attributes of cutaneous lesions, the authors employed the local binary pattern (LBP) [124]. In another study, the authors employed a mechanism called Color Image Analysis Learning Vector Quantization (CIA-LVQ) [125].

In another study, the local features were acquired employing the interest points that were identified with the help of the difference of Gaussian (DoG) [126]. Pacheco et al. [86,91] suggested an approach to merging CNN-extracted attributes with medical information and this proposed methodology was called MetaBlock. According to the study’s results, MetaBlock can be a more effective feature combination technique than conventional concatenation approaches. The authors in [127] adopted the same technique on the model they utilized in a recent investigation. A pre-trained ResNet50 had been employed by the authors in their model for withdrawing deep features [128].

The acquired attributes are then used to assess and provide the PSLs’ detailed classification data. A training set of data is used to create a classification method, which is subsequently used by several classifiers. As a result, the features and classifier drive the model’s performance. Furthermore, comparing methods of classification with the same dataset and features is often beneficial [46]. Compared to fully linked networks, CNNs are frequently simpler to train and require fewer hyper-parameter modifications. Custom CNN models have been trained to recognize skin lesion photos in certain investigations [81,88,102]. The researchers in [83] used two convolutional layers in each of two similar CNN frameworks to assess the overall structure and distinct texture of skin pictures. The researchers combined them to create a fully connected layer. CNNs had been employed in a different study to extract features from photographs [92]. CNNs that were initially trained to respond to classification issues had undergone modifications as a result of numerous investigations [129]. A few researchers used ResNet50 [130] and GoogleNet [85,91,131]. ResNet-152 was tuned to function as a classifier in a different study [132]. The same researchers trained a CNN to categorize images into 134 groups in a separate test [133]. The researchers in Reference [123] used a customized artificial neural network (ANN) to achieve the categorization goal. The researchers of a different study [61] proposed a linear soft margin SVM as an alternative to the linear SVM utilized in [67]. In [119], an SVM architecture with a kernel for the histogram’s intersection was employed as the classification strategy. Additionally, the researchers in [129] classified many categories using a weighted SVM and adjusted them according to each category’s degree of complexity. Two machine learning models—KNN and a hybrid classifier created by fusing KNN and Decision Tree (DT)—were employed in a different investigation [60]. The KNN was used for binary classification in a different study [108].

Non-invasive detecting techniques have emerged from using incident light, oil immersion, and magnification. The epiluminescence microscope (ELM) stimulates these mechanisms. Nonetheless, experience as a medical professional is still required for accuracy. Recently, a variety of methods have been used in experiments with automatic skin cancer diagnosis. Computer-Aided Diagnosis (CAD) approaches for dermoscopic pictures have evolved to readily stimulate the diagnosis of worrisome lesions, hence aiding dermatologists in their diagnostic evaluation. Furthermore, inexperienced practitioners may utilize it as a supplementary tool to carry out an initial inquiry [46]. Considering the key elements which are excluded from the dermoscopic pictures, these therapies might be grouped into broadly 2 categories. By automatically eradicating the identical attributes, a subset seeks to resemble therapeutic recognition strategies. Another discipline has its foundation in machine learning and computational pattern recognition utilized for common visual elements. In [134], a depiction of an assessment network affected by therapy aspects is shown.

Gola et al. [135] demonstrated a method for diagnosing melanoma. The globular and reticular patterns are identified by pattern analysis, which is the basis for this automated approach. An additional study that presents automatic melanoma identification is [136]. The dermoscopic rules calculated using the ABCD rule are the foundation of the framework.

The majority of the CAD observed with the ML method has been reported [137–139]. A methodology that uses the K-nearest neighbors (KNN) classifier and removes distinguishing variables including color, texture, and border was given by Ganster et al. [19]. The methodology achieved 92% specificity (SP) and 87% sensitivity (SE). Another study involved Artificial Neural Networks (ANN) utilized as classification entities [138] and the features withdrawn were border and color. In another study, a skin lesion classification formulated on SVM is presented [111]. The features withdrawn include shape, texture, and color; attaining a SE of 93% and a SP of 92%.

Tumpa et al. [140] presented melanoma recognition and classification based on ANN utilizing composite texture features. The procedure employs Otsu’s thresholding segmentation and the features withdrawn are ABCD, LBP, and GLCM, attaining an accuracy of 97.7%.

While each investigating organization chooses a distinct data source, results evaluation becomes complicated regardless of whether the CAD practices offer desirable outcomes. It is therefore prevalent to have to create a dataset of dermoscopic pictures that researchers can employ as ground truth.

Skin cancer assessment and categorization are largely processed through DL algorithms especially CNNs. ML and hybrid methods are utilized as well from time to time.

A procedure of removal of artifacts and noise from the raw data file is referred to as pre-processing. The most important phase in the procedure of segmentation is to discover and describe the area of lesion in the pictures.

Feature extraction is a crucial step in obtaining relevant data from the partitioned area. Domain-specific features are often used in machine learning settings. Thanks to DL, the field of ML has completely changed in the last few decades. It is believed that ANN algorithms represent the most sophisticated area of ML. These frameworks got inspiration from the formation and functioning of the human brain. In many different situations, DL methodologies have demonstrated impressive results contrary to distinct conventional ML techniques. Many DL approaches have been used in computer-based skin cancer screening throughout the last few decades.

Summary: In the latter half of the 20th century, investigators started to investigate the possibility of computers augmenting and analyzing medical images. The primary focus of CAD software was basic pattern recognition in radiographic images. The journey of CAD started as a quest to augment human vision with computing capacity. In recent times, artificial intelligence (AI) and machine learning evolved to an extent where CAD has emerged as an exceptionally developed platform for discerning minute inconsistencies that are largely overlooked by human vision. The advancement was thwarted by certain challenges. Owing to the lack of precision and inability to overtake human experience, early CAD methods were viewed with skepticism But persistent exploration and development in technology made it suitable for CAD to enhance our potential and emerge as an effective testing technique.

CAD is a technique that effectively merges modern innovations in technology and medicinal attributes to offer a rather prudent solution. Optimizing the precision as well as the efficacy of disease observation by means of the utilization of several feasible techniques is a primary goal of the comprehensive architecture. A lot of advantages stem from CAD synthesis in the therapeutic domain. Most specifically it allows for identifying cancers at an early stage and enhances the accuracy of diagnosis while minimizing the duration of time necessary for interpreting. Diagnoses might be conducted more efficiently and precisely attributable to the practical equipment termed computer-aided diagnostic (CAD). However, there are actually plenty of problems with CAD operation. Among the primary issues is the fact that algorithm training relies on reliable information that may prove a limitation for specific circumstances. Furthermore, the integration of CAD into conventional healthcare services implies significant infrastructure and mechanical modifications.

It is utilized for implementing automatic image thresholding in image processing. This method divides a pixel into foreground and background by returning a single intensity threshold. Fig. 4 illustrates skin cancer images after applying Otsu thresholding in the proposed methodology. The method examines the threshold which reduces the intra-class variance, characterized as a weighted aggregate of variances of the 2 categories. The clip limit (CL), as well as block size (BS), are two significant CLAHE parameters. These two factors primarily regulate enhanced image quality. Due to the input image having poor intensity and the large clip limit obtaining histogram flatter, the image becomes brighter when the clip limit is increased. The dynamic range expands along with the block size, which also increases image contrast. When using the entropy of the image, the two factors define at point having the highest entropy curvature and generate inputs with a subjectively high quality [48]. Each contextual region is subjected to histogram equalization using the CLAHE method. The clipped pixels from the original histogram are then distributed among each grey level. The redistribution histogram differs from the regular histogram since each pixel intensity is restricted to the selected maximum. However, enhanced and the input image carry minimal and maximal grey values. The steps of the CLAHE algorithm to enhance an input image are as follows:

Stage 1: Separate an input intensity image into disjoint contextual areas. The overall number of image tiles is equivalent to M × N, and 8 × 8 is a good score to maintain the chromatic information of an image.

Stage 2: Calculate a histogram of each contextual area by grey levels present in the array image.

Stage 3: Compute a contrast limit histogram of the contextual regions by the clip limit value as: (1)σw2=ω1(t)σ2(t)+ω2(t)σ22where ω1 and ω2 represent weights for the probabilities of 2 categories segmented by t, i.e., threshold and σ2, σ22 represent variances.

The class probability ω((1, 2)) (t) is calculated from N bins of histogram as: (2)ω1(t)=∑m=0t−1P(m)where NCLIP stands for normalized clip limit in the range [0, 1], where NCL is the actual clip limit. The pixels will be clipped if the number of pixels exceeds NCL. If N∑clip is the total amount of clipped pixels, then the mean of the remaining pixels for distribution to each grey level as: (3)ω2(t)=∑m=tN−1P(m)

According to this approach for MT, the gray levels of the image are divided into groups or divisions. Let R be the whole area of space that an image occupies. The goal of image segmentation is to split the region R up into smaller sections R1, R2, ..., RN.

For this procedure a thresholding level (t) is specified, the rule set for bi-level thresholding can be stated as follows: (4)R1←g if 0≤g≤t,R2←g if t≤g≤N−1

Here R1 and R2 are regions, N − 1 represents the maximum pixel term, g represents any gray term from the specified range {0, 1, 2, …, N − 1}.

In case the pixel value lies below threshold t, it corresponds to Region R1, else it corresponds to Region R2. The rule set for n MT can be stated as:

Else if (Hregion(i)+Navggray)>NCL then (5)R1←g if 0≤g≤t1 (6)R2←g if t1≤g≤t2 (7)Ri←g if ti≤g≤ti+1 (8)Rn←g if tn≤g≤tn+1

Here t1, t2, ti, ti+1, tn and tn+1 represent different thresholds. The overall grayscale is precisely categorized into numerous exclusive regions through preference or by evaluating the threshold value (t). In this paper generation of threshold for 2 level, 3 level, and 7 level using Otsu’s approach is considered for evaluating the optimized t values. Fig. 5 illustrates skin cancer images after applying Multilevel Otsu thresholding in the proposed methodology.

This is a local thresholding approach where instead of manually defining the threshold value, it is adapted and determined automatically in accordance with image pixels and arrangement for transforming the pixels of the image to grayscale or a binary image. Primarily, this approach advocates selecting threshold values automatically for segmenting the prime entity from its background in a state where there are distinct illumination, colors, or contrast in the image. The pixel intensity upon which the pixels related to the background and foreground are dissociated is termed as threshold value for an image. In other words, the higher intensity image pixels than the threshold will be dissociated from the lower intensity image pixels. For executing adaptive thresholding, as opposed to providing a single threshold value by primitive attempt, it is possible to partition the image foreground and background more effectively by employing a much more precise threshold value.

The segmentation of lesions can be attained by correlating the color of every pixel by a threshold t. In this paper for adaptive segmentation, small blocks are used to adapt the images. Here the function provides a rate Y running on X to evaluate the adapted threshold. Further standard deviation (SD) of X is evaluated. If the rate of SD of the block of pixels is less than 1, it is labeled as a background.

The Niblack and Sauvola thresholding approach has particularly evolved to enhance the image quality. These are local thresholding techniques that change the threshold in accordance with the local mean and SD for every pixel in a sliding window. The final local pixel value of these thresholding approaches is felicitated by distinct positive parameters also. It is performed to ensure the division between the entity and the background. (9)SV=X¯∗(1+c∗(σ/p−1)) (10)NV=X¯+c∗σ−awhere SV represents Sauvola value, NV represents Niblack value, X¯ represents mean, σ represents standard deviation (SD), c represents positive parameter, p represents dynamic range of SD.

In this paper, Niblack, Sauvola, and binarized Sauvola techniques are used for classifying skin cancer lesions. Fig. 6 illustrates skin cancer images after applying Niblack and Sauvola thresholding in the proposed methodology.

This segmentation technique utilizes energy forces and limitations to split up the pixels of interest out of the remainder of the image for additional refining and evaluation. For segmentation, an active contour interprets a particular edge or curvature for the areas of the target entity. The contour relies on certain constraints, and on this basis, it is categorized into distinct types. The approach is used in several image processing applications, especially in therapeutic imaging, by segmenting the regions from distinct therapeutic images [153,154]. The contours are edges formulated for the targeted region required in an image. It is a cluster of points that experience interpolation procedures that might be linear, splines, or polynomials, representing the curve in the image [155]. Distinct models of active contours are implemented for the segmentation approach in image processing, and the prime objective of is to specify a smooth shape and create a closed contour for the region. These models outline the edges of entities or distinct attributes of the image to produce a contour. The curve of the models is ascertained by numerous contour methods by applying external and internal forces. The energy function is related to the curvature specified in the image. The synthesis of forces because of the image that is precisely employed for administering the orientation of the contour on the picture is referred to as external energy, while internal energy regulates the deformable variations. The appropriate contour is attained by determining the minimum of the energy functional. The cluster of points that observe a contour is considered to be contour-deforming. By reducing the energy function this contour fits the preferred image contour defined by minimizing the energy function.

In this paper, the Chan Vese model for active contour is utilized for segmenting skin cancer lesions. It is an effective and flexible approach that can segment distinct types of images, which are relatively difficult to segment through traditional segmentation approaches [155]. The approach is demonstrated by the Mumford-Shah functional that is utilized in the therapeutic imaging domain.

The segmentation is performed using the image segmenter app by evolving 100 iterations. The components connected to the edges of the image are suppressed. The masked illustration is then created through the input, which is the final segmented image. Fig. 7 illustrates skin cancer images after applying the Chan Vese segmentation approach to the proposed methodology.

The watershed algorithm is formulated by withdrawing certain background and foreground and later utilizing markers to create a watershed run and determine the accurate edges. The approach typically helps in determining connected and overlapping entities in an image.

In this paper, the utilization of the watershed algorithm is initiated by modifying the RGB images into grayscale. Consequently, to provide a smooth image, a morphological top-hat filter is utilized. In the next step, the images are transformed into binary format employing grayscale and are later complemented. In the next step, the distance is computed and the watershed transformation is calculated. The segmented regions are displayed by employing distinct colors. Fig. 8 illustrates skin cancer images after applying the watershed segmentation approach.

It is an arithmetical method that is used to interpret boundary value problems of the Eikonal equation: (11)|∇T(y)|=1v(y)

The problem generally illustrates the progression of a closed surface as a function of time T and speed v in the usual direction at point y on the propagating area. The speed function is defined, and the time when the contour traverses the point y is attained by solving the equation. The FMM uses this optimal control evaluation of the complication so as to come up with a solution emerging from the boundary values. The FMM works on the fact that data only moves outward from the seeding region.

In this paper, FMM is utilized to segment the ROI of the images on the basis of variations in gray intensity in contrast to the seed locations. A mask is then created by defining seed locations. In the subsequent step, a weighted array is computed on the basis of grayscale intensity variations, and the images are segmented using these weights. Further, by thresholding the geodesic distance matrix by distinct thresholds, different segmentation results are computed. Fig. 9 illustrates skin cancer images after applying FCM.

By synthesizing pixels in the image plane based on their proximity and color homogeneity, this method produces superpixels. It is the most often used method for superpixel segmentation; however, it requires too much processing power. Images of skin cancer after using the simple linear iterative clustering technique are shown in Fig. 10.

It is utilized for implementing automatic image thresholding in image processing. This method divides a pixel into foreground and background by returning a single intensity threshold. Fig. 11 illustrates skin cancer images after applying K-means clustering in the proposed methodology. The method examines the threshold which reduces the intra-class variance, characterized as a weighted aggregate of variances of the 2 categories. It is an unsupervised clustering algorithm that simply indicates that there is no availability of labeled data. It is employed to ascertain distinct groups or categories in the specified data derived from the similarity of the data. Data points in the same category are indistinguishable from distinct data points in a similar group compared to those in different groups.

It is one of the most extensively employed clustering approaches, where K denotes the number of clusters.

The algorithm works as follows:

1. Select the number of clusters that is K.

2. Arbitrarily allocate the data points to the K clusters.

3. Evaluate the center of the clusters.

4. Evaluate the length of the data points from the centers of every cluster.

5. Derived from the length of data points from the cluster, redistribute them to the adjacent clusters.

6. Then compute the new cluster center.

7. Reiterate Steps 4, 5 and 6 until data points turn the clusters, or till the specified number of iterations is attained.

It is a particular case of K-means when the probability function applied is merely 1 in case the data point is nearest to a centroid and 0 contrarily.

The algorithm works as follows:

• Presume a certain quantity of clusters, i.e., K.

• Arbitrarily compute the K-means μ_K related to the clusters and evaluate the probability that every data point x is a component of a specified cluster K.

• Re-compute the centroid of the cluster as the weighted centroid provided the likelihood of components of each data point x: (12)μK(n+1)=∑xϵKx∗P({μK|x}b)∑xϵKxP({μK|x}b

Reiterate till convergence or till the number of iterations specified by the user has been obtained.

In this paper, fuzzy c-means clustering of images is executed which divides the image into n clusters automatically by random initialization. The counts of clusters and iterations might be controlled and determined by the user. The outcomes of the function are upgraded cluster centers and segmented images illustrated in Fig. 12.

An image processing approach that detects the edges of entities within an image is known as the edge detection technique. It is utilized for segmenting images by identifying discontinuities in brightness. Fig. 13 illustrates skin cancer images after applying (a) Sobel (b) Prewitt (c) Robert (d) Log (e) Zero Crossing (f) Canny. The distinct edge detection approaches used in this paper are discussed in Table 2.

The PH2 database was developed by a mutual research association among the Universidade do Porto, T’ecnico Lisboa, and the Dermatology service of Hospital Pedro Hispano in Matosinhos, Portugal [156]. This dataset consists of 200 dermoscopic pictures (8 bits RGB pictures) having a resolution of 768 * 560. The pictures are categorized into 80 common nevi, 80 atypical nevi and 40 melanomas.

This step involves transforming the RGB images to grayscale, and noise filtering by using the dullrazor software. It is utilized since some of the acquired images are noisy because of the presence of hair and bubbles around the lesion. So Dullrazor is administered to reduce the consequence of hair being obscured on the skin in the final picture utilized for categorization, making the segmentation more accurate. Fig. 15 illustrates the output image from the pre-processing phase.

The subsequent step is identifying and segmenting the region of interest (ROI) that depicts the suspected area. This procedure comprises steps as follows: thresholding, computing the histogram, and conversion to a black-and-white image in the case of global thresholding using OTSU segmentation. For multi-level thresholding using the Otsu approach, for a 2-level approach, a single threshold value is computed and the image is segmented into 2 levels. For the 3-level approach, 2 threshold values are computed, and the image is segmented into 3 levels, which are then converted into a colored image. For the 7-level approach, thresholds for 7 levels from the entire RGB image are generated. The threshold for each plane of the RGB picture is then generated. Process the picture with an array of threshold terms enumerated from the picture. Process each RGB plane separately using the threshold vector computed from the given plane. Quantize every RGB plane through the threshold vector prompted for that plane.

Feature extraction depicts approaches aiming at withdrawing additional value details off the images. The withdrawn elements called features might be statistical, boundary, morphological, and textural characteristics. These features are applied as input data for distinct image processing procedures such as segmentation and classification.

The proposed algorithm extracts textural features of the images in the test phase. Let X represent the feature database, which has 200 images. Let FVq represent query feature vector (FV) having length 34 (representing features extracted). The 200 images are divided into 3 categories: atypical Nevi, common Nevi, and Melanoma. These categories are subdivided into benign (containing atypical and common nevi mages) and malignant (containing melanoma images) forms. The final FV is a combination of both benign and malignant classifications forming a vector of 200 * 34. The feature extraction mechanisms utilized in the paper are discussed below:

Image classification, or deriving information from a picture through CV and ML techniques, is the final phase. The 12 distinct metrics have been utilized to quantify the textural feature extraction.