Hybrid Feature Fusion and Machine Learning Approaches for Melanoma Skin Cancer Detection

Skin cancer is an exquisite disease globally nowadays. Because of the poor contrast and apparent resemblance between skin and lesions, automatic identification of skin cancer is complicated. The rate of human death can be massively reduced if melanoma skin cancer can be detected quickly using dermoscopy images. In this research, an anisotropic diffusion filtering method is used on dermoscopy images to remove multiplicative speckle noise and the fast-bounding box (FBB) method is applied to segment the skin cancer region. Furthermore, the paper consists of two feature extractor parts. One of the two features extractor parts is the hybrid feature extractor (HFE) part and another is the convolutional neural network VGG19 based CNN feature extractor part. The HFE portion combines three feature extraction approaches into a single fused feature vector: Histogram-Oriented Gradient (HOG), Local Binary Pattern (LBP), and Speed Up Robust Feature (SURF). The CNN method also is used to extract additional features from test and training datasets. This two-feature vector is fused to design the classification model. This classifier performs the classification of dermoscopy images whether it is melanoma or non-melanoma skin cancer. The proposed methodology is performed on two ordinary datasets and achieved the accuracy 99.85%, sensitivity 91.65%, and specificity 95.70%, which makes it more successful than previous machine learning algorithms.


Related Work
The medical and biological datasets are increasing rapidly. To analyse such big and complex data, artificial intelligence and machine learning algorithms become most popular [10][11][12][13][14][15][16]. Therefore, it is important to implement novel techniques to uncover the medical and biological patterns. In particular Machine learning and deep learning techniques have been widely used to analyse imaging data [15,[17][18][19][20][21][22][23][24]. Skin cancer is one of the most challenging medical; many researchers have used several methodology or techniques in melanoma skin cancer detection including dermatology image of skin.
For example, the author [25][26][27] introduced a novel system to detect melanoma skin cancer. The system has preprocessed skin lesion input image to get the high-quality image. The thresholding and edge detection techniques is used for segmentation. Then the system extract features from segmented image by using geometry-based features and ABCD (Asymmetry, Border, Color, and Diameter) features. These extracted features classified the image as ordinary skin and melanoma skin cancer. The research didn't clarify the accuracy in their proposed novel system. This study [28] provided an approach to detect normal skin and abnormal skin. First of all, the system preprocessed the dermoscopy image and segmentation by using threshold value. The gray level co-occurrence matrix was utilized to extract features, and feature selection was done using principal component analysis approaches. For classification the research applied support vector machine and calculate total dermoscopy score to get good accuracy. In 2018, the research conducted by Yuexiang Li and Linlin Shen. For lesion region detection, dermoscopic feature extraction, and classification, the researcher offered two deep learning algorithms. For segmentation and classification, Fully Convolutional Residual Networks (FCRN) are utilized, and the Lesion Index Calculation Unit (LICU) approach refines the classification outcomes calculation distance heat-map. Finally, CNN is used for dermoscopic feature extraction. The highest given accuracy using their framework is 91.2% [29]. Vijayalakshmi M et al. (2019) assessed the problem using three phases like data collection and augmentation, model design and prediction. The author used CNN, SVM algorithms and augmented it with different image processing tools and got 85% accuracy [30]. A. Pramanik     3: D is for extracting each image attribute ( j, 1,570). 4: Fc ( j, 1) = R(x, 1) +Fc ( j, 1).

Dataset
By analyzing the performance of measurement matrices, the images are characterized as melanoma or non-melanoma. In this research, the paper used a standard dataset from Academic torrents and collected by HAM10000 (Human Against Machine with 10000 training images) dataset [36]. The dataset contains RGB 16,170 images in distinct types: melanoma and non-melanoma, which are publically available at

Data Preprocessing
Data preprocessing is the most essential task for analysis before applying any feature extractor and classification methods [37][38][39]. For the Medical images preprocessing part is inevitable. Each image has a group of pixels that contains noise and imperfection. Some procedures are used to remove redundant pixels and distortion pixels from images in order to produce correct outcomes. After performing the ROI technique, the proposed system converts the RGB image to Grayscale. In order to remove irrelevant text and machine annotations from training and test images, the area of interest (ROI) is extracted [40]. This approach suppresses the number of unnecessary noise and distortion. Fig. 3 demonstrates the procedure of image data preprocessing. For any supervised learning for training phase, it is necessary to collect a huge volume of labeled data. In most applications, inadequate training data might lead to an overfitting issue [21,41]. By erasing the overfitting status, the data augmentation approach is able to overcome this barrier. The best fit in our model is a machine learning based on CNN model, which overcomes the limitation of the shortage of labeled images. Some of the augmentation techniques include translation, resizing, slicing, magnification, rotation, reversing, and brightness adjustment that may be used to change the size and appearance of a lesion in a dermatological image.

Modified Anisotropic Diffusion Filtering
The objective of proposed Modified Anisotropic Diffusion Filtering is preserving in detail information while speckles are being reduced. The suggested approach uses covariance and kurtosis measurements of noise to maintain the critical edge information. This speckle reduction technique is continued until the image's noise component reaches to Gaussian value. If the distortion is Gaussian, the skewness value must be 0. Equation (1) represents the noise component. The loop will remain until the kurtosis of noise part is less than the measurement. This measurement can be defined by equation (3). When the relationship with both image class and disturbance class is the smallest, the iteration will end. In below equations (1) to (7), I and I0 represent actual and noisy image, µ is used to represent the mean of noise intensity G. The kurtosis value k is determined by equation (4). Equation (6) derives the image intensity correlation, whereas Equation (7) generates the noise intensity correlation. When I and G have the lowest amount of deviation, the recommended filtering will produce reliable results.

Feature Extraction
This section demonstrates applied different feature extractor techniques and their feature extraction process. In this research, two image extractors were used: HFE and a CNN-based feature extractor. HOG, LBP, and SURF are three feature extraction techniques included in the HFE extractor. To diagnose illnesses from skin dermatological images, the feature extractor techniques such as LBP, HOG, SURF, and neural network-based features extracts hybrid features. Numerous machine learning and computer vision applications use data fusion [19,[42][43][44]. The challenge of combining many feature vectors, known as features fusion, is critical. The technique proposed is based on entropy-based feature fusion. The vector that has been fused (1×1280). The entropy is applied to the features vector for the selection of optimal attributes based on the score. Equation (8) and (9) explain how the feature selection method works mathematically. From a total of 7948 characteristics, entropy was utilized to choose 1280 score-based characteristics. In Equations (8) and (9), indicates entropy and denotes features probability. In order to classify melanoma skin cancer images, the final attributes are sent to the classifiers.

Hybrid Feature Extractor
Three extracting features strategies are included in the HFE extractor: local binary pattern (LBP), histogram-oriented gradient (HOG), and speed up robust feature (SURF), all of which result in a single fused feature vector. In the proposed system, hybrid features extractor i.e., LBP (12), SURF (11) and HOG (10) features are extracted. A fused features vector is utilized to evaluate the suggested approach. HOG properties are extracted from the images at all grid dense regions and are often utilized for object detection. HOG attributes (1×3780), LBP attributes (1×59), and SURF attributes (1×13) may all be used to define the form and appearance of skin cancer.
Furthermore, the extracted features are combined into one vector.

Histogram Oriented Gradient Features
This research used a feature extraction algorithm known as HOG. Firstly, the input image is converted into gray scale image after that image transformed into gradient image for better edge detection. The gradients or edges orientation histogram is obtained in each cell unit, divided into smaller cells, and then these histograms are combined to give a HOG description [45,46]. Fig. 6 depicts the HOG feature extraction algorithm's fundamental flow. To extract the feature in HOG, we should first create a gradient on both the x and y axis. X and y direction slopes can be determined fast since the horizontal direction's pattern is K= [-1, 0, 1], and its inversion may be used to filter an image. The following is the indication: ga = I (x+1, y)-I (x-1, y) (14) gy = I (x, y+1)-I (x, y-1) The pixel value of (x, y) is indicated by I, and the direction slope of x is represented by gx, while the orientation slope of y is indicated by gy (x, y). The slope magnitude of (x, y) is represented by g (x, y) and denoted by ∆g (x, y) = √ (gx 2 + gy 2 ) And (x,y) gradient 's direction (θ) is determined as follows: Ɵ = arctan (gy/ gx) (17)

Local Binary Pattern (LBP) Features
LBP provides texture analysis and local spatial statistics of ultrasound image [47]. A threshold value is used to level the contiguous pixels and it is represented by 0 and 1. If each pixel value is larger than the center pixel value, each adjacent pixel gray value (3 × 3) is leveled as 1, otherwise it is leveled as 0. Thus, LBP represents a set of binary digits which are used to replace center pixel value after converting into decimal. Equation (18), (19) represent LBP segmentation from test image where g (p) is gray level pixel for surrounding pixels (i, j) and g (c) is complementary constant. For neighbor (8, i), the total number of samples is 256.

Speed Up Robust Feature (SURF) Features
SURF is a similarity invariant representation and comparison algorithm. Its robust feature extractor technique is used in nearest neighbor matching [48]. During augmentation it can extract features. As a scaling and rotation variant algorithm, SURF provides fast operator computation using box filtering [49]. The two functions of SURF are feature extraction and feature description. The features extraction in SURF is done with Hessian matrix-based interest point approximation. The SURF descriptor provides unique information of features generated by surrounding area of an interest point. It operates by indicating the distinctive orientation of an interesting point using Haar wavelet responses. Before calculating descriptor, interest areas of neighbor interest point are rotated to its selected orientation. The Hessian matrix H (x, y) at scaling is given by formula (20) for a given location X = (x, y). Equation (21) represents the wavelet response in x and y direction is noted by dx and dy direction. A vector V is computed for each sub region.

CNN Based Feature Extraction and Classification
CNN based feature extraction techniques are most popular process in medical image processing [50]. In proposed research a pre-trained CNN model and scratch model were applied to extract features. Comparatively CNN (VGG19) provides good performance than scratch model. Pre-trained VGG19 model was well predictive and more performable feature extractor technique for our dataset among VGGNet, VGG16, Scratch model, ResNet50 and AlexNet.
19-layers are used to develop the network model using VGGNet techniques. Performance of given dataset using VGG19 is more accurate and reliable compared with others model. Fig. 7 shows how VGG19 was constructed using 16 convolution layers and three completely linked layers. The convolution component is separated into 5 successive max-pooling layers, with a nonlinear ReLU function acting as an activation function to ensure that each convolution layer's output is more accurate. Depth of the 5 consecutive layers is 64,128,256,512 and 512 respectively. Each of the layer formatted with sub regions where pooling layer decrease the learnable parameter. Ending layer played a vital rule to get feature vector of proposed VGG19 model. Every fully attached layer besides the dropout layers was regularized with L2 to reduce overfitting problems during implementation of the fine tune model. Applying ReLU function on VGG19 based CNN model produce 4096 tuned features for further process. The features of VGG19 based CNN model and Hybrid feature extractor features are fused on a fused vector. The SoftMax feature aids in determining if the disease is Melanoma or Non-melanoma. The graphical illustration is shown in fig. 7.

Finding a Cancer Region
If the image is identified as a skin infection, the output section in this research is advised to determine the possibly damaged region. Depending upon the nature of the lesion, skin lesions appear in a range of forms and sizes. A circle is drawn around any spot that has been identified as a lesion.  Sometimes variant segmentation techniques are used to tumor detection which are not convenient and efficient due to deformation growth of lesion. Finding accurate region of skin lesion and symmetry of axis is time consuming and challenging. Researchers nowadays uses fast bounding box to detect lesion faster and robust. A fast bound box technique is fast and robust process of segmentation which overcome the above problem by locating an axis parallel box or bounding box around the skin lesion. This process is scored based on gray scale intensity analysis. Score function provides a linear search method for bounding boxes. FBB is an unsupervised and real time basis process where images fixation is not mandatory. Using this boundary box symmetry method tumor region can be detected alone finally which provides an outlier around the tumor.

Experimental Setup
The suggested system was evaluated using four performance criteria: accuracy, sensitivity, specificity, and precision. The parameters used to compute each measure are True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN). In Table 1 In

Results and Discussion
Dermatology skin images with various speckles, noises, and resolutions were employed as test data for the proposed approach. Information reserve and noise reduction are required when working with relevant features. In the image preprocessing stage, the desired system employed modified anisotropic diffusion filtering. Three evaluation metrics are used to quantify performance in modified anisotropic diffusion filtering: Edge Preservation Factors (EPF), Minimum Square Error (MSE), and Signal to Noise Ratio (SNR). Higher SNR and EPF values indicate superior noise removal and edge detail preservation, respectively, when using the proposed filtering strategy. The minimum MSE value, and on the other hand, shows that there is less inconsistency between the input and filtering images. Table 2 represents the performance measurement using evaluation metrics. Although all known filters perform well in MSE, the proposed modified anisotropic filtering approach has a higher SNR and EPF. CNN models like VGG-16, AlexNet, VGG-19, and ResNet50 are increasingly often used to feature extracted from training and testing data sets. VGG-19 outperforms other pre-trained algorithms in the proposed approach. All of these models provide the same result when given a similar training dataset. Table 3 represents performance of different CNN models for experimental data. However, proposed system also compares with some scratch models where performance is not satisfactory like fine-tuned models. CNN classifier is used to detect accuracy as classification scheme. Before classifying, CNN extracted features to train. Test characteristics were retrieved from a test image using multiple preprocessed models to evaluate the performance of CNN models.  Table 3 shows that fine-tuned VGG19 pre-trained model provides better result in feature extraction using CNN. ResNet50 and VGG19 pre-trained models perform significantly better than VGG16 and AlexNet models. In term of accuracy VGG19 is selected for our test dataset though ResNet50 shows better sensitivity. Table 4   Without preprocessing the performance result proposed fused features and CNN is not satisfactory. Table 5 shows accuracy, sensitivity and specificity result of noisy and speckle test dataset. Aside from CNN, several machine learning methods such as Decision Tree (DT), Random Forest (RM), Artificial Neural Network (ANN), and K closest Neighbor (KNN) have been used to produce the final classification. Proposed fused vector features are fed to classification methods to find better classifier. Finally, as demonstrated in Table 6, CNN shows better performance as a classifier. The proposed system used a starburst pattern and poor dermatological images during the testing phase. A modified apostrophic diffusion filtering approach is used to eliminate multiplicative noise present in the test image. It is capable of efficiently overcoming the challenges of a noisy image. As a result, the preprocessing approach for the input text image is more successful in reliably extracting features using the suggested fused vector. The bulk of study based on dermatological images has poor generalization accuracy. For generalization, the proposed system employed a standard dataset from an academic torrent. In generalization, we have used 16,170 images with two classes of 11170 images are melanoma and 5,000 images are non-melanoma. Table I displays the generalization findings' confusion metrics.
The cancer affected mole can come in many variant colors like brown, black and tan. The variety of mole in the same mole could be cancerous [20,51]. Finding the limitations of dermatological image was the first issue in this study. The area of interest (ROI) is retrieved from training images to eliminate extraneous text and machine classifications. Prior to that, the image was transformed from RGB to grayscale to eliminate the problem of multiple colors.
Detecting skin cancer using the same dataset or a related dataset is a tough issue. Preprocessing techniques, feature extraction approaches, and classification methods were all utilized by the researchers. It's now challenging to recommend a prospective strategy or combination of procedures for removing speckles from dermatological images that is more beneficial. In the phase of feature extraction and noise removal, the suggested approach presents a feasible technique. Table 7 shows the comparative analysis of relevant research. Table 5 shows that the suggested approach of feature fusion utilizing HFF and CNN (VGG19) performs greater accuracy than the CNN classifier. For dermatological datasets, CNN performs well as a binary classifier. To improve performance, modified anisotropic filtering of the input test image is essential. This preprocessing technique suppresses the limitation of test noisy image.
To obtain a more detailed experimental outcome, a k-cross validation procedure is used. The CNN system is built using a 5-fold classification technique after extracting features. This stage divides the feature vector into five sub folds at random. Four sub folds are chosen from the training dataset, whereas only one sub fold is chosen from the testing dataset. Table 8 shows the various results obtained utilizing the individual and combined methods.  The proposed deep learning model has a great potential to be used on healthcare imaging data analysis including CXR images, brain imaging, etc. Using a deep neural network to detect and classify skin cancer is a tough challenge.

Conclusion
Today, it is important to use innovative techniques to identify medical biomarkers and classify disease including cancer [50][51][52][53][54]. The proposed system in this study provides an exploratory analysis through hybrid feature extractor and convolutional neural network to gain more features information and thus achieve a promising accurate result. A modified anisotropic filtering technique is used in dermatology test image to diverge speckle from noisy images. The proposed fused vector with CNN and HFF has proven 99.49% accuracy. The best performing CNN classifier used to detect whether it's melanoma or non-melanoma skin cancer and also established 99.85% accuracy. In future this work will be extended using more promising machine learning algorithm and design a system for remotely checkup from home.