Skin cancers image classification using transformation and first order statistic features with artificial neural network classifier

Skin cancer is one of the most dangerous types of cancer. Some types of this cancer lead to death, so cancer must be discovered and indexed to avoid its spread through initial detection in the impulsive stage. This paper deals with the detection and indexing of different types of melanomas using an artificial neural network (ANN) depending on the international skin imaging collaboration (ISIC) 2018 dataset that was used. The pre-processing is the most important part because it formulates an image by insolated the cancer part from the skin image. It consists of four stages, removable, cropping, thinning, and normalization. This phase has been used to eliminate all the undesirable hair particles on the image lesion. The cropped image transforms into frequency domain coefficients using discrete cosine transform (DCT), discrete wavelet transform (DWT), and gradient transform for sub-band images to extract its feature. The statistical feature extraction is implemented to minimize the size of data for ANN training. The experimental analysis used dataset ISIC 2018 consisting of seven different types of dermoscopic images (this paper deals with four types only). For classification purposes, ANN was implemented and the accuracy obtained is about 88.98% for DWT, 85.44% for sub-band DCT, and 76.07% for sub-band gradient transform.


INTRODUCTION
Skin cancer is the most common cancer worldwide, known for its increasing incidence rates and increasing burden, it is also difficult to visually distinguish normal tumors from abnormal and general tumors. General practitioners (GPs) usually do not have sufficient experience to diagnose skin cancers [1]. That's why the search focuses on the development and implementation of skin cancer systems to improve the quality of diagnostics by using three-technique with first-order statistical features to detect border of lesions followed by feature extraction selection and classification Uninhibited development of irregular skin cells is acknowledged as a skin tumor. It happens when deoxyribonucleic acid (DNA) damage to skin cells generates changes, or genetic faults, that principal the skin cells to multiply quickly and form menacing cancers. Skin cancers have usually diagnosed by physical analysis and surgery [2]. The surgery is a simple way where part or all of the spot is removed and tried in the test center. The dermoscopy is the most effective method to generate color skin imaging, which is the most improved tool in the analysis of tumors depending on the experiment doctor's eye [3]. The standard datasets Int J Adv Appl Sci ISSN: 2252-8814  Skin cancers image classification using transformation and 1st order statistic … (Asmaa Abdulrazaq Alqaisi) 233 of skin cancer are limited and cover a few numbers of skin cancer types, with a little number of images for every type [4]. New methods for skin cancer classification are the goal of researchers works are related to perfect methods of cancer classification. Work on skin image data replicated by flipping and rotation compensates for variations between the training and the test sets and has an optimistic effect on the convolutional neural network (CNN) presentation [5]. Used neural network (NN) to categorize skin cancer into three kinds. The type of dermoscopic images has been implemented in this work, which directly works on color images of skin without pre-processing [6]. Uses an extra dense NN and tempers the features with the CNN [7]. This work attainment to explain the dataset method by counting outside data with skin lesions category. The other class lesion problem is the severe class imbalance number of images. This paper care of this property by considering die rent method input resolutions and die rent cropping stratagems, by integrating data depending on age, structural site, and sex. Classified three-dimensional (3D) image shapes were performed by converting them to images of two-level (binary 0 and 1) [8]. The new algorithm called Adaptive snake used to extract the segment features and classification by artificial neural network (ANN) has been employed.
The ANN model was built to produce abstractions accuracy in data using multiple dispensation layers. The important elements of using ANN are: i) The minimizing of cancer huge data for the training and NN models; ii) Graphic processing unit was developed to depiction for computer analysis; and iii) The linear unit, dropout, and batch normalization has been corrected in ANN methods [9]. The automatic identification of tumors from dermoscopy color images has three interesting contracts. First, the skin is very like lesions typical belongs to different lessons, such as size, texture, color, and shape. Second, is the high association between melanoma and non-melanoma lesions. Third, different environmental conditions, such as hair, veins, noise, and radiance [10]- [12].
In this work different types of feature extraction with first-order statistic calculation for preprocessing, the image has been applied to the ANN to classify the skin cancer types and reach to accurate diagnosis of classification. In this work, two cascade feature extraction methods have been used. First, is the transformer method, discrete cosine transform (DCT), discrete wavelet transform (DWT), and gradient method has been used. Second is the statistical methods, which calculate mean, standard deviation, skewness, and the fourth momentum has been implemented to minimize the data and extract features. The ANN has been used to identify and classify cancer types.

DISCRETE WAVELET TRANSFORM (DWT)
The wavelet transform converts the image pixels from the space domain to the frequency domain (DWT coefficients). It includes a cluster of low pass filters and high pass filters named filter bands. In DWT may filter can use to generate multi-level DWT. The DWT uses filter bands to take away divisions of the pixels into many frequency bands. The DWT translates the pixels into multi-scale depictions of both spatial and frequency selves to gather. These tips are for effective multi-scale exploration with lower calculation costs [13]. It produces location-sensitive data which very essential in analyzing thyroid nodules. The subbands for 1-level and 2-levels using two dimension DWT haar filters as a basis function wavelet decomposition are shown in Figure 1 [14].

DISCRETE COSINE TRANSFORM (DCT)
The DCT is a mathematical equation that takes a signal and transforms it from the spatial domain into the frequency domain, The DCT is the real part of the discrete fourier transform (DFT). The low frequency by in the left top corner, and when the frequency increased by zigzag order arrived at the right bottom corner [15]. An image of size (N*M) every pixel is denoted by p(x,y), The two-dimension DCT is used to transfer the image into P(u,v) DCT coeffic.

IMAGE GRADIENT
In image processing, the process of changing the values of color pixels is called color gradation. in measured, gradient of a two-dimension function at each point in the image p (x, y) produce two dimension vector with the values given by the derivatives in the vertical gv (i, j) and horizontal directions gv (x, y) at points in locations x row and y column [16]. The gradient vector values depend on pixels value and it distribution in the sub-block of size (3*3) [17], [18]. The gradient equation cconvertsimage pixels p (x, y) into two vectors gradient magnitude M (x, y) and gradient direction A (x,y).

THE FIRST ORDER STATISTICAL METHOD
In skin cancer images many statistical methods have been used for the extraction features of data [17]. Among these, are color means variance skewness, and fourth momentum. It is very simple and is invariant to transformation coefficients information regarding an image, the statistic equation used in this work are [18]- [20]. a) Mean, the mean of coefficients X (i, j) of size M, N.

ANN NETWORKS
The ANN network comprises of three layers call input, output, and one or more hidden layer [21]. Training methods to build the networks accuracy depend on the values of the features like gradient or statistical methods (that used in this work) [22]. The network is built by calculated layers weight values, these values update by repetition of calculation to minimize the error between calculated output values and expected values. This problem is called the disappearing gradient [23], [24].

PROPOSED SKIN CANCER CLASSIFICATION METHOD
In this work, different skin image classification methods have been implemented. The proposed systems block diagram is shown in Figure 3.

THE PROPOSED SYSTEMS CONTAIN THREE CASCADE STAGES 8.1. Color image pre-proses
The preprocessing steps to extract cancer part from another skin part, namely i) Noise removal: Here unwanted pixels like hair are eliminated from the red, green, blue (RGB) images using median filters; ii) Cropping: Our area of interest is the cancer region. The white spaces surrounding cancer have been cropped; iii) Thinning: Cancer strokes are represented with minimum cross-sectional width by eliminating a few foreground pixels; and iv) Normalization: The RGB images are resized to three 128x128 pixels for cropping cancer regions only. The preprocessing step is shown in Figure 4.  Average Filter (5x5) List3 Full Space Average Filter (17x17) List4 Resize image (128*128) End

Feature extraction 8.2.1. First step
The normalized cropped cancer region has been transformed into frequency domain. Three different transformations (DCT, DWT, and gradient explain above) have been used for (RGB) to extract the difference between the image's pixel and transform it. The proposed system divided the cropped image into sub-band (8*8), and the DCT for every sub-band in RGB images has been implemented to extract features from the neighborhood, this is shown in Figure 5. DCT for (RGB) Image is shown in Figure 5(a), The Sub-Block DCT for (RGB) image is shown in Figure 5(b). In Figure 6 shows the DWT coefficient for 1-level and 2-level. The 1-level DWT coefficients of R, G, and B image; and respectively shown in Figures 6(a), 6(b), 6(c); and the 2level DWT of R, G, and B image, respectively shown in Figures 6(d), 6(e), 6(f). The magnitude of gradient transform for sub-band (8*8) in RGB images has been implemented as shown in Figure 7. For whole R image is shown in Figure 7(a) and for sub-block R image is shown in Figure 7

Second step
The statistical method has been calculated (mean, std. dev., skewness, and fourth momentum) for RGB coefficient results in DCT, DWT, and gradient. When the contrast was calculated for cropped cancer image. This step helps to collect the result and minimized the size of data for learning. Figure 8 shows 1 st order statistic feature for the DCT transform for the first 20 images, Figure 9 shows 1 st order statistic feature for the DWT (Haar) 1'st level transform for the first 20 images, Figure 10

ANN network
The ANN (backpropagation method) has been used to build an artificial network for the classification of Skin cancer. The ANN consists of two hidden layers every layer built from 20 nodes. The 200 iterations have been used to minimize the error. Figure 12 shows the diagram of DCT training, Figure 13 shows the diagram of gradient training, and Figure 14 shows the diagram of DWT (Haar) transform training.

RESULT OF PROPOSED SYSTEMS
In this work, the dataset of four different types of cancer images has been used, see Table 1. The three proposed method has been implemented to build three different classification ANN networks. The proposed method for feature extraction is implemented for sub-band DCT, DWT (1-level and 2-level), and sub-band gradient transform. In the second step, five statistic calculation has been applied to extract the minimum size of features (every color Image have 3*5=15 feature values). The ANN used 70% of data for training and 25% of data for testing. The ANN train has a mean square error (MSE) as shown in Table 2. The confusing matrix used to implement three different skin cancer classification methods the testing results are shown in Table 3. The unbalance of skin cancer images in all types of the percentage of confusing matrix is shown in Table 4.

CONCLUSION
The proposed methods have been established that a group of feature extraction methods with ANN built network can realize low-priced classification performance for skin cancer color image classification to skin cancer lesion. The ANN appears as highly preset and accurate. The inequality dataset with a big difference in total images for each class makes it complex to simplify the image features classification of the lesions. The result shown in Table 3 obtain that DWT methods give high perfection in cancer classification in a low number of cancer images. When sub-band DCT gives less ratio in error detect of on other types this falling the risk and expand the robustness of the method.