• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Deep Learning Based Automated Diagnosis of Skin Diseases Using Dermoscopy

    2022-08-24 03:29:02VatsalaAnandSheifaliGuptaDeepikaKoundalShubhamMahajanAmitKantPanditandAtefZaguia
    Computers Materials&Continua 2022年5期

    Vatsala Anand,Sheifali Gupta,Deepika Koundal,Shubham Mahajan,Amit Kant Pandit and Atef Zaguia

    1Chitkara University Institute of Engineering and Technology,Chitkara University,Punjab,India

    2Department of Systemics,School of Computer Science,University of Petroleum&Energy Studies,Dehradun,Uttrakhand,India

    3School of Electronics&Communication,Shri Mata Vaishno Devi University,Katra,182320,India

    4Department of Computer Science,College of Computers and Information Technology,Taif University,Taif,21944,Saudi Arabia

    Abstract:Biomedical image analysis has been exploited considerably by recent technology involvements,carrying about a pattern shift towards‘a(chǎn)utomation’and ‘error free diagnosis’classification methods with markedly improved accurate diagnosis productivity and cost effectiveness.This paper proposes an automated deep learning model to diagnose skin disease at an early stage by using Dermoscopy images.The proposed model has four convolutional layers,two maxpool layers,one fully connected layer and three dense layers.All the convolutional layers are using the kernel size of 3 * 3 whereas the maxpool layer is using the kernel size of 2 * 2.The dermoscopy images are taken from the HAM10000 dataset.The proposed model is compared with the three different models of ResNet that are ResNet18, ResNet50 and ResNet101.The models are simulated with 32 batch size and Adadelta optimizer.The proposed model has obtained the best accuracy value of 0.96 whereas the ResNet101 model has obtained 0.90, the ResNet50 has obtained 0.89 and the ResNet18 model has obtained value as 0.86.Therefore,features obtained from the proposed model are more capable for improving the classification performance of multiple skin disease classes.This model can be used for early diagnosis of skin disease and can also act as a second opinion tool for dermatologists.

    Keywords: Dermoscopy images; CNN; deep learning; classification;optimizer;ResNet;diagnosis;skin disease

    1 Introduction

    Skin is an outer covering that separates the environment and body of human beings.It is the largest organ of the body.It acts as a barrier guarding the body against foreign materials, harmful chemicals, UV radiations etc.The population affected from this skin disease was 15.1 crore in the year 2013 and then increased to 18.8 crore in the year 2015.The two major categories in which the lesion is spotted are malignant and benign classifications.Benign is not dangerous whereas malignant skin lesions can spread to other parts.Research prompts that one-fifth of the population is affected by skin disease,therefore the classification becomes more complex.For classification of skin disease,there are various signs that may be used like color of the lesion,body size distribution of lesion etc.It is very difficult to analyze each lesion individually.The difficulty of classification tasks is increased as the manual feature extraction method is not suitable for classification of lesions.Diagnosis of a skin disease is a difficult task as there exists many types of skin disease[1–4].Various types of disease may have similar characteristics like some skin diseases are the same in color.There also exists inter and intra class similarities.Therefore,appropriate selection of features is important that will help in proper identification of skin disease.The success of a system depends upon how well it can perform and show results[5,6].There are many technologies that are used for diagnosis of disease in medical images.Machine Learning (ML) and Deep Learning (DL) is an area of Artificial Intelligence (AI)which deals with the techniques that allow a machine to learn and solve a given problem.ML basically teaches computers to perform tasks that humans can naturally perform.Whereas,DL learns different features directly from the given data.The“deep”in deep learning comes from the use of multiple layers in the network.DL techniques are capable of handling data with high dimensionality and give better performance.It is efficient in focusing on the right features of the image on its own.However,machine learning is less accurate to work with large amounts of dataset for prediction purposes.Therefore,deep learning proves to be more efficient as compared to ML techniques as deep learning techniques are able to work with large databases and with more accuracy[7].From the last few years,improvements in deep learning Convolutional Neural Networks(CNN)have shown favorable results and also became a challenging research domain for classification in medical image processing[8].Meng et al.[9]had performed feature classification and automatic recognition in a non-invasive way.Authors had used polarized-light dermoscopy image technology.The data was collected from the hospital and a total of 642 images were used for evaluation of the model.The average classification accuracy achieved was 93.65%.Garnavi et al.[10]had presented a novel CAD system for diagnosis of melanoma.The proposed diagnostic system is applied on a set of 289 dermoscopy images.The system achieves an accuracy of 91.26% and area under curve value of 0.937.Mahmoud et al.[11] had used a total of 448 images and had achieved an accuracy of 51.1%and 75.6%with wavelet and curvelet transforms respectively.They had used wavelet and curvelet transforms.Barata et al.[12] had used a total of 176 dermoscopy images for classification of disease.They had used kNN and SVM classifiers and had achieved sensitivity of 96%, specificity of 80% for global methods and for local methods it had achieved specificity of 75%.Fassihi et al.[13]had used morphologic operators and wavelet transforms and had achieved an accuracy of 90%.They had used a total of 91 images taken from hospitals and websites.Xie et al.[14] had designed a new classifier and used two datasets.First dataset was xanthous race dataset which consists of 240 dermoscopy images and second dataset was caucasian race dataset which includes 360 dermoscopy images.They had achieved sensitivity of 95%,specificity of 93.75%,accuracy of 94.17%on xanthous race dataset and sensitivity of 83.33%,specificity of 95%,accuracy of 91.11% on caucasian race dataset.Breneman et al.[15] had used the Otsu method for segmentation and morphological operators.After that, they had ported the algorithm to a mobile platform for segmenting the lesion from the background in the live images.Smaoui et al.[16] had done pre-processing followed by a segmentation process in which region growing was used.After that feature extraction was done followed by ABCD rule.A set of 40 dermoscopic images was used.It had attained values, accuracy as 92%, sensitivity as 88.88% and specificity as 92.3%.Garnavi et al.[17] had presented a clustering-based histogram thresholding algorithm for segmentation.By using the algorithm, morphological operators are utilized to obtain the segmented lesion.The algorithm was tested on 30 high resolution dermoscopy images.It had achieved an accuracy of 97%.Lu et al.[18]had proposed a novel method,for the purpose of segmentation of the melanocytes.The dataset consists of 28 different histopathological images of skin tissue.It showed a sensitivity rate of 80%and positive prediction rate of 70%.Amin et al.[19] performed pre-processing to resize the images and used Otsu algorithm to segment the skin lesion.The publicly available datasets were combined to form a large dataset for the validation of the proposed method.The obtained results show sensitivity as 99.52%,specificity as 98.41%,positive predictive value as 98.59%,false negative rate as 0.0158 and accuracy as 99.00%.Codella et al.[20] combined latest ML techniques with deep residual networks and CNN for segmentation of dermoscopy images in order to resolve the skin disease classification issue.They worked on the dataset consisting of 900 dermoscopic images.A system was presented that includes the latest developments in DL combined with ML techniques.Also, examining the detected area of disease for detection of melanoma.It had achieved accuracy of 76%, specificity of 62% and sensitivity of 82%.Mahbod et al.[21] investigated image down-sampling and cropping of skin lesions and a three-level fusion approach.From the ISIC archive and HAM10000 dataset 12,927 images were extracted.It had achieved an accuracy of 86.2%.Koohbanani et al.[22] proposed a method by including transfer learning algorithms.The proposed method was encouraged by the UNet architecture.It consists of a validation set with 100 images and a test set of 1000 images extracted from ISIC 2018.Hekler et al.[23]examined the effects of label noise by training and evaluating CNN with 804 images.The dataset consists of 804 images taken from HAM10000 and ISIC archive.For dermatology the achieved accuracy was 75.03%and for biopsy accuracy was 73.80%.Dorj et al.[24]mainly focused on classification of skin cancer using SVM and CNN.They had used a pre-trained AlexNet model for extracting features.It consists of a set of 3753 images.The values of achieved average accuracy for actinic keratosis was 92.3%, basal cell carcinoma was 91.8%, squamous cell carcinoma was 95.1%,melanoma was 94.2%.Zafar et al.[25–28]proposed a method by combining two architectures called Res-Unet.The dataset used was taken from the PH2 dataset with 200 dermoscopic images and ISIC-17 test data consisting of a total of 600 images.On the ISIC-17 dataset the value of Jaccard index was 77.2% and Dice coefficient was 0.858.Whereas on the PH2 dataset, the value of Jaccard index was 85.4% and Dice coefficient was 0.924.Also, authors had used different datasets and different techniques for detection of skin disease.In 2020, Masni et al.[29] had used a total of 2750 images from ISIC 2016,2017 and 2018 dataset.They had worked only on three classes such as NV,MEL and AKIEC.They had worked on four pre-trained models such as Inception-v3,ResNet-50,Inception-ResNet-v2 and DenseNet201 and had achieved an accuracy of 81.29%,81.57%,81.34%and 73.44%respectively.Huang et al.[30]had developed a convolution neural network and had achieved an accuracy of 85.8%and More et al.[31]had achieved an accuracy of 75.03%on HAM10000 dataset.In this paper,a model is proposed and pre-trained networks are used for detection of skin disease.The last fully connected layer has been dropped out and is replaced with a Softmax layer with weights for the classification purpose.The major contributions of the study are as follows:

    1.A new Convolutional Neural Network (CNN) based model has been proposed for the classification of skin disease using dermoscopy images.

    2.Proposed model has been compared with three ResNet models i.e.,ResNet18,ResNet50 and ResNet101.

    3.The data augmentation technique has been applied to increase the dataset and to overcome the imbalance in dataset.

    4.The models are simulated using various performance parameters such as precision,sensitivity,specificity,F(xiàn)1 score and accuracy.

    5.The proposed model has shown promising results which will act as a second opinion tool for dermatologists.

    The remaining paper is structured as follows.Section 2 gives description of materials and methods.Experimental setup and comparison of results are shown in Section 3.Section 4 gives the discussion and conclusion of the study.

    2 Materials and Methods

    In this study, a complete explanation of the proposed methodology is given for skin disease detection and the dataset used for the validation of the proposed model.

    2.1 Dataset

    For the validation of the proposed method, the dermoscopy images are obtained from the HAM10000 dataset [32].It contains 10015 skin disease images and its detailed levels are tabulated in Tab.1.For HAM10000 dataset, Nikon Coolscan 5000 ED was used to digitized the diapositives and files are stored in jpg format.After that scanned images are cropped with lesion in center and histogram corrections are applied manually for contrast enhancement.In this dataset, metadata is provided which includes a set of information about a particular image.It provides information about a particular image belonging to a specific type of skin disease.The images are RGB and all the images are in jpg format.The complete dataset is split into training and testing.For the training purpose,approximately 80%of the data is used and for testing 20%of the data is used.

    Table 1:Dataset description

    The dataset consists of seven classes:Actinic Keratoses(AKIEC),Benign Keratosis-like Lesions(BKL),Vascular Lesions(VASC),Basal Cell Carcinoma(BCC),Melanocytic Nevi(NV),Dermatofibroma (DF), and Melanoma (MEL).The proposed model classifies skin disease image into seven categories that includes mainly two steps:pre-processing (normalization and augmentation) and classification using pre-trained CNN architectures.The detailed explanation of each stage is given in the subsequent sections.

    2.2 Normalization

    Normalization of image is used to keep numerical stability to the CNN architectures[33].Initially,skin disease images are in the RGB format and have been normalized in between the range of 0–1 by multiplying each pixel value with 1/255.While using normalization,a model is expected to learn faster.

    2.3 Data Augmentation

    As the available training dermoscopy images in the dataset are 10015 which is very less for training the deep learning model.To resolve this issue,offline data augmentation techniques have been implemented to increase the number of sample images[34].A major aim is to develop a technique that can function in the presence of noise and variation on a large number of images.The data augmentation on images is done using different transformation techniques like flipping, rotation, zooming etc.Flipping operation is achieved in both horizontally and vertically direction.The number of training images before and after augmentation are shown in Tab.2.With data augmentation techniques,the number of images increased to approximately 5000 in all the cases as shown in Tab.2.The AKIEC disease class images increased to 5055,which was earlier having only 297 images.

    Table 2:Training images before and after augmentation

    Tab.3 shows seven classes of skin disease images taken for testing purposes.The testing images for NV were 883,whereas the testing images for VASC were 13.

    Table 3:Testing images

    2.4 Proposed Method

    A model is designed for classification of skin disease in HAM10000 dermoscopy images.Fig.1 shows the structure of the proposed model.First layer of convolution is applied with 16 filters of size(3*3)on(28*28)size of images which generates 448 parameters.After that second convolution layer is applied with 32 filters of size (3 * 3) on output shape (26 * 26) which generates 4640 parameters.Then it is forwarded to (2 * 2) sized max pooling layer, that shows output image of size (13 * 13).Again,convolution layer is applied with 32 filters and output shape of(13*13)that obtained 9248 parameters.Next convolution layer is applied with 64 filters and obtained output shape of(11*11)that obtained 18496 parameters.Max pooling layer is applied with a pool size of(2*2)and has obtained output shape of 6*6.Flattening layer flattens all the features and then three dense layers are applied for classification of skin disease into seven different categories.

    Figure 1:Structure of proposed model

    The main part of CNN is the convolutional layer.A convolution is sliding a filter over the input.CNN works by extracting features directly from images.Convolution between an input signal x[n]with a system having impulse response h[n]is shown in Eq.(1)

    In the pooling layer,the image stack is shrinked into a smaller size.The most common methods for reduction are max pooling and average pooling.Max pooling operates by finding the highest value within the window region and discarding the rest of the values.Average pooling on the other hand uses the mean of the values within the region instead.All the flatten features are input into ANN for further processing.The flattened output is fed to a feed-forward neural network and backpropagation applied to every iteration of training.Over a series of epochs,the model is able to classify using the Softmax Classification technique.Tab.4 describes the size of filter,number of filters,input image size,output image size and total number of parameters used at each layer for the proposed model.

    Table 4:Parameters of proposed model

    3 Experimental Setup and Performance Analysis

    The models are evaluated using dermoscopy images taken from the HAM10000 dataset.The proposed model is compared with all the ResNet models [35,36] i.e., Resnet18, ResNet50 and ResNet101 in terms of precision,sensitivity,accuracy and F1 Score.Adadelta[37]optimizer is used for evaluation of models using 32 batch size.The features obtained from the proposed model are capable of improving the performance of classification for multiple skin disease classes.Batch Size specifies the images that are managed in only one iteration.In this paper,batch size with 32 values is used for evaluation and validation.Epochs is the total number of times;the neural network receives the entire dataset.For seven types of skin disease,the output layer was used with the softmax activation function.Respective values of these simulation parameters selected for the tuning deep neural networks are tabulated in Tab.5.

    Table 5:Simulation parameters

    3.1 Proposed Model

    The proposed model is evaluated using parameters such as sensitivity, precision, and F1-score.Fig.2a shows the graph of precision,sensitivity F1 Score and specificity values taken on all the seven classes of the HAM10000 dataset.From the graph it can be seen that the value of sensitivity and F1 Score is approximately 97% in case of NV disease.The value of precision is approximately 90% in case of MEL.The values of sensitivity are nearly 90%in case of disease BCC and DF.The value of precision is higher for MEL disease as compared to other diseases.Fig.2b shows the confusion matrix of the ResNet18 model.The matrix provides a clear view of correct and incorrect predictions.Each and every column and its corresponding row is labelled by class name.In this study,classes 0,1,2,3,4,5 and 6 represent AKIEC,BCC,BKL,DF,NV,VASC and MEL respectively.Diagonal values of matrices represent the accurate number of images classified by the particular model.

    Figure 2:Results of proposed model(a)Performance parameters(b)Confusion Matrix

    3.2 Results on ResNet18

    It is an Artificial Neural Network(ANN)that enables the design of a deep neural network by using shortcuts or skip connections.Fig.3 shows the structure of ResNet18[35].Every ResNet architecture performs convolution and max-pooling using 7*7 and 3*3 filter sizes respectively.Curved arrows are showing the shortcut connections.Here the solid curves maintain the input feature dimension with the same number of channels as in the previous layer.Moreover, it does not use any stride for convolution.Whereas dotted curve lines double the number of channels and halve the feature map dimension by using stride of 2.It consists of 4 layers having four convolutional blocks each.Layer 1 of the model consists of 4 convolution blocks.The number of channels used to perform the operation is 64 and size of filter is 3*3.Layer 2 of the model consists of 4 convolution blocks.The number of channels used to perform the operation is 128 and size of filter is 3*3.Layer 3 of the model consists of 4 convolution blocks.The number of channels used to perform the operation is 256 and size of filter is 3*3.Layer 4 of the model consists of 4 convolution blocks.The number of channels used to perform the operation is 512 and size of filter is 3*3.Layers 2,3 and 4 are using a convolution stride of 2,due to which it reduces the size of the feature map dimension to its half.The ResNet18 model is evaluated using parameters such as sensitivity,precision,and F1-score.Fig.4a shows the graph of precision,sensitivity,F(xiàn)1 Score and specificity values taken on all the seven classes of the HAM10000 dataset.From the graph it can be concluded that optimizer Adadelta has shown good results in case of NV and BCC disease with a value of sensitivity approximately 90%.The value of precision is higher for NV disease in comparison to other diseases.Fig.4b shows the confusion matrix of the ResNet18 model.The matrix provides a clear view of correct and incorrect predictions.Each and every column and its corresponding row is labelled by class name.In this study,classes 0,1,2,3,4,5 and 6 represent AKIEC,BCC,BKL,DF,NV,VASC and MEL respectively.Diagonal values of matrices represent the accurate number of images classified by the particular model.

    Figure 3:Structure of ResNet18

    Figure 4:Results of ResNet18(a)Performance parameters(b)Confusion Matrix

    3.3 Results on ResNet50

    Fig.5 shows the structure of ResNet50[38].Every ResNet architecture performs convolution and max-pooling using 7*7 and 3*3 filter sizes respectively.This model consists of 4 stages with each stage shown with blue and orange boxes alternatively.Curved arrows are showing the shortcut connections.Here the solid curves maintain the input feature dimension with the same number of channels as in the previous layer.Moreover, it does not use any stride for convolution.Whereas dotted curve lines double the number of channels and halve the feature map dimension by using stride of 2.Stage 1 of the model consists of 1 residual block having 2 stacked convolution layers with filter size of 3*3 and 64 channels.The stage 1 block is repeated three times.Stage 2 of the model consists of 2 residual blocks that contain 2 stacked convolution layers each.The second residual block of the second stage is repeated three times.Similarly,the second residual block of the third stage and fourth stage is repeated five times and two times respectively.

    Figure 5:Structure of ResNet50

    The ResNet50 model is evaluated using parameters such as sensitivity, precision, and F1-score.Fig.6a shows the graph of precision, sensitivity, F1-score and specificity values taken on all the seven classes of the HAM10000 dataset.From the graph it can be seen that optimizer Adadelta has outperformed on DF disease with value above 98%.Fig.6b shows the confusion matrix of the ResNet50 model.This model has performed best on Adadelta optimizer with batch size 32.The matrix provides a clear view of correct and incorrect predictions.Each and every column and its corresponding row is labelled by class name.In this study,classes 0,1,2,3,4,5 and 6 represent AKIEC,BCC,BKL,DF,NV,VASC and MEL respectively.Diagonal values of matrices represent the accurate number of images classified by the particular model.

    Figure 6:Results of ResNet50(a)Performance parameters(b)Confusion Matrix

    3.4 Results on ResNet101

    Fig.7 shows the structure of ResNet101 [38].This model consists of 4 stages with each stage shown with blue and orange boxes alternatively.Curved arrows are showing the shortcut connections.Here the solid curves maintain the input feature dimension with the same number of channels as in the previous layer.Moreover, it does not use any stride for convolution.Whereas dotted curve lines double the number of channels and halve the feature map dimension by using stride of 2.Stage 1 of the model consists of 1 residual block having 2 stacked convolution layers with filter size of 3*3 and 64 channels.The stage 1 block is repeated three times.Stage 2 of the model consists of 2 residual blocks that contain 2 stacked convolution layers each.The second residual block of the second stage is repeated three times.Similarly,the second residual block of the third stage and fourth stage is repeated twenty-two times and two times respectively.

    Figure 7:Structure of ResNet101

    The ResNet101 model is evaluated using parameters like sensitivity, precision, and F1-score.Fig.8a shows the graph of precision,sensitivity and F1-score values taken on all the seven classes of the HAM10000 dataset.VASC disease has shown values as 89–90%approximately for all the parameters.In case of MEL disease,the values of precision are very less.In case of sensitivity parameter,optimizer Adadelta has outperformed on NV disease with value above 95%.Fig.8b shows the confusion matrix of the ResNet101 model.The matrix provides a clear view of correct and incorrect predictions.Each and every column and its corresponding row is labelled by class name.In this study,classes 0,1,2,3,4,5 and 6 represent AKIEC,BCC,BKL,DF,NV,VASC and MEL respectively.Diagonal values of matrices represent the accurate number of images classified by the particular model.

    Figure 8:Results of ResNet101(a)Performance parameters(b)Confusion Matrix

    3.5 Comparison of Proposed Model with Three ResNet Models

    The proposed model is evaluated using Adadelta optimizer and 32 batch size.The parameters used for evaluation of the proposed model are precision,sensitivity,F(xiàn)1 score and accuracy.The proposed model is compared with the ResNet model.Fig.9 shows the comparison of precision values of the proposed model with ResNet18, ResNet50 and ResNet101.Precision refers to the closeness of two or more measurements to each other.Precision is the positive predictive value means the correctly identified cases from all the predicted cases.From Fig.9 it can be seen that the proposed model is showing best results for each and every skin disease.It is showing 89%precision value in case of MEL whereas 84% in case of BCC.So, the overall proposed model is showing outperforming results for precision.

    Figure 9:Precision of proposed model and all ResNet models

    Fig.10 shows the comparison of sensitivity values of the proposed model with the ResNet18,ResNet50 and ResNet101 model.Sensitivity is the true positive rate and it is the proportion of actual positives that are identified correctly.A good sensitivity value is approximately 90%.It is analyzed from the figure that the best value of sensitivity is 96%for NV disease whereas in case of DF disease value of sensitivity is 90%.

    Figure 10:Sensitivity of proposed model and all ResNet models

    F1 Score is used for a better measure when there is an uneven class distribution.It is the harmonic mean of precision and recall.Fig.11 shows the best value of F1 score as 96%for disease NV.

    Figure 11:F1 score of proposed model and all ResNet models

    Fig.12 shows the comparison of specificity values of the proposed model with the ResNet18,ResNet50 and ResNet101 model.Specificity is the ability of a test to correctly identify people without the disease.It is analyzed from the figure that the best value of specificity is 99%for disease AKIEC.The proposed model has shown best results on all skin disease classes.

    Figure 12:Specificity of proposed model and all ResNet models

    Fig.13 shows the comparison of ResNet18, ResNet50, ResNet101 and the proposed model in terms of Adadelta optimizer with 32 batch size.The proposed model has obtained accuracy of 0.96 whereas ResNet101 model has obtained 0.90, ResNet50 has obtained value as 0.89 and ResNet18 model has obtained value as 0.86.

    Figure 13:Accuracy of proposed model and all ResNet models

    4 Discussion and Conclusion

    In this study,a deep learning based model is proposed for the classification of seven skin disease classes using dermoscopy images.Various pre-trained ResNet architectures are studied and their results are analyzed using publicly available skin disease dataset.Although,the proposed model has only 10 layers as compared to ResNet 18, ResNet 50 and Resnet 101 that are having 18, 50 and 101 layers respectively.Even then the proposed model outperforms these ResNet models in terms of accuracy,sensitivity,precision and F1 Score.The dermoscopy images are taken from the HAM10000 dataset.The models are simulated with 32 batch size and Adadelta optimizer.The reason for choosing the Adadelta optimizer is that it is an extended version of Adagrad that adapts learning rates based on a moving window of gradient updates,instead of accumulating all past gradients.All the models are compared in terms of accuracy and the proposed model has a best accuracy value of 0.96 whereas the ResNet101 model has obtained 0.90,ResNet50 has obtained 0.89 and ResNet18 model has obtained value as 0.86.Therefore, features obtained from the proposed model are capable of improving the performance of classification for multiple skin disease classes.Dorj et al.[24]had used a pre-trained AlexNet model and used a total of 3753 images for extracting features.The values of achieved average accuracy for actinic keratoses was 92.3%,basal cell carcinoma was 91.8%,squamous cell carcinoma was 95.1%, melanoma was 94.2%.In 2020, Huang et al.[30] had developed a convolution neural network and had achieved an accuracy of 85.8% and More et al.[31] had achieved an accuracy of 75.03% on HAM10000 dataset.In future, the accuracy of the proposed model can be improved by applying the noise removal filter in the pre-processing stage.Further, the proposed model can be simulated using different types of optimizers, batch sizes, epochs and learning rates to enhance the results.Moreover,the processing time for the proposed model can be compared with the pre-trained models.

    Acknowledgement:This work was supported by Taif university Researchers Supporting Project Number(TURPS-2020/114),Taif University,Taif,Saudi Arabia.

    Funding Statement:This work was supported by Taif university Researchers Supporting Project Number(TURPS-2020/114),Taif University,Taif,Saudi Arabia.

    Conflicts of Interest:Authors declare that they have no conflicts of interest to report regarding the present study.

    国产女主播在线喷水免费视频网站| 一本久久精品| 欧美日韩亚洲高清精品| 亚洲,欧美,日韩| 男女下面插进去视频免费观看| 久久久久久伊人网av| 美女xxoo啪啪120秒动态图| 老司机亚洲免费影院| 97在线人人人人妻| 国产成人精品福利久久| 国产精品一国产av| 日韩成人av中文字幕在线观看| 黄色视频在线播放观看不卡| h视频一区二区三区| 欧美日韩综合久久久久久| 欧美人与性动交α欧美精品济南到 | 久久国内精品自在自线图片| 国产精品久久久av美女十八| 国产成人精品久久二区二区91 | av天堂久久9| 欧美日韩av久久| 新久久久久国产一级毛片| 看免费成人av毛片| 日韩精品有码人妻一区| 美女福利国产在线| 免费黄网站久久成人精品| 亚洲成av片中文字幕在线观看 | 美女国产高潮福利片在线看| 国产成人精品一,二区| 久久国产精品大桥未久av| 成人黄色视频免费在线看| 久久午夜综合久久蜜桃| 秋霞在线观看毛片| 久久精品aⅴ一区二区三区四区 | 亚洲四区av| av在线app专区| 日韩熟女老妇一区二区性免费视频| 欧美人与性动交α欧美精品济南到 | 妹子高潮喷水视频| 成年人免费黄色播放视频| 国产精品一国产av| 嫩草影院入口| 老女人水多毛片| 亚洲人成77777在线视频| 久久久久精品久久久久真实原创| videossex国产| 韩国精品一区二区三区| 99国产综合亚洲精品| 亚洲欧美成人综合另类久久久| 久久97久久精品| 国产无遮挡羞羞视频在线观看| 亚洲欧美成人精品一区二区| 日韩伦理黄色片| 啦啦啦中文免费视频观看日本| 韩国高清视频一区二区三区| 久久久久国产网址| 精品国产乱码久久久久久小说| 高清不卡的av网站| 国产乱人偷精品视频| 国产白丝娇喘喷水9色精品| 婷婷色综合大香蕉| 国产免费一区二区三区四区乱码| 免费黄网站久久成人精品| 久久精品人人爽人人爽视色| 亚洲国产av影院在线观看| 毛片一级片免费看久久久久| 少妇精品久久久久久久| 电影成人av| 一级毛片电影观看| 欧美亚洲日本最大视频资源| 熟女av电影| 一区二区三区精品91| 制服人妻中文乱码| 中文字幕制服av| 岛国毛片在线播放| 日日撸夜夜添| 国产精品熟女久久久久浪| 午夜影院在线不卡| 欧美精品高潮呻吟av久久| 久久人妻熟女aⅴ| xxx大片免费视频| 成年人午夜在线观看视频| 国产xxxxx性猛交| 高清黄色对白视频在线免费看| 日韩中文字幕视频在线看片| 免费黄色在线免费观看| 国产在线免费精品| 七月丁香在线播放| 亚洲欧美日韩另类电影网站| 亚洲精华国产精华液的使用体验| 欧美人与性动交α欧美精品济南到 | 国产日韩欧美视频二区| 女性生殖器流出的白浆| 99热国产这里只有精品6| 天美传媒精品一区二区| 新久久久久国产一级毛片| 日本-黄色视频高清免费观看| 大码成人一级视频| 热re99久久精品国产66热6| 91精品三级在线观看| 婷婷色综合大香蕉| www日本在线高清视频| 精品视频人人做人人爽| 母亲3免费完整高清在线观看 | 日韩精品免费视频一区二区三区| av.在线天堂| 精品午夜福利在线看| av女优亚洲男人天堂| 精品人妻在线不人妻| 一级,二级,三级黄色视频| 亚洲欧洲精品一区二区精品久久久 | 精品久久蜜臀av无| 国产精品熟女久久久久浪| 超碰97精品在线观看| 欧美日本中文国产一区发布| 日本wwww免费看| av不卡在线播放| 国产色婷婷99| 亚洲精品视频女| 黑人欧美特级aaaaaa片| 女人高潮潮喷娇喘18禁视频| 狠狠婷婷综合久久久久久88av| 啦啦啦在线观看免费高清www| 久久久久精品性色| 欧美人与善性xxx| 久久精品国产自在天天线| 亚洲婷婷狠狠爱综合网| 赤兔流量卡办理| 国产精品成人在线| 国产成人a∨麻豆精品| 十八禁高潮呻吟视频| 免费观看av网站的网址| 久久热在线av| 久久久久久人妻| 丝袜美腿诱惑在线| 中国国产av一级| 飞空精品影院首页| 亚洲欧美精品自产自拍| 久久综合国产亚洲精品| 亚洲图色成人| 亚洲av中文av极速乱| 色视频在线一区二区三区| 乱人伦中国视频| 亚洲欧美精品自产自拍| 校园人妻丝袜中文字幕| 男女无遮挡免费网站观看| 亚洲一区二区三区欧美精品| 综合色丁香网| 在线观看一区二区三区激情| 久久精品亚洲av国产电影网| 亚洲综合精品二区| a级片在线免费高清观看视频| 婷婷色综合大香蕉| 下体分泌物呈黄色| 国产成人精品福利久久| 亚洲欧洲日产国产| 老熟女久久久| 1024视频免费在线观看| 午夜日韩欧美国产| 亚洲三级黄色毛片| 婷婷色综合大香蕉| 欧美国产精品va在线观看不卡| 国产午夜精品一二区理论片| 成人国产av品久久久| 丁香六月天网| 久久精品亚洲av国产电影网| 春色校园在线视频观看| 日韩在线高清观看一区二区三区| 亚洲国产av影院在线观看| 丝袜脚勾引网站| 中文字幕人妻丝袜一区二区 | av不卡在线播放| 欧美老熟妇乱子伦牲交| 18禁裸乳无遮挡动漫免费视频| 在线免费观看不下载黄p国产| 精品人妻熟女毛片av久久网站| 国产一区二区在线观看av| www日本在线高清视频| 欧美黄色片欧美黄色片| 久久久久视频综合| 免费黄色在线免费观看| 中文字幕人妻丝袜制服| 亚洲精品av麻豆狂野| 久久国产精品男人的天堂亚洲| 国产精品偷伦视频观看了| 咕卡用的链子| 亚洲,欧美,日韩| 成人国语在线视频| 国产精品久久久久久精品古装| 国产免费现黄频在线看| 久久ye,这里只有精品| 老熟女久久久| 制服诱惑二区| 黄色视频在线播放观看不卡| 精品酒店卫生间| 自拍欧美九色日韩亚洲蝌蚪91| 秋霞在线观看毛片| 寂寞人妻少妇视频99o| 国产精品女同一区二区软件| 美女午夜性视频免费| 欧美精品一区二区免费开放| 一区二区三区乱码不卡18| 欧美变态另类bdsm刘玥| 女人精品久久久久毛片| 久久久久久人人人人人| 91午夜精品亚洲一区二区三区| 看免费av毛片| 日韩在线高清观看一区二区三区| 亚洲av男天堂| 黄色 视频免费看| 中文字幕另类日韩欧美亚洲嫩草| 青春草亚洲视频在线观看| 丁香六月天网| 成人国产麻豆网| 国产黄色视频一区二区在线观看| 国产av精品麻豆| 国产视频首页在线观看| 欧美97在线视频| 如日韩欧美国产精品一区二区三区| 久久国内精品自在自线图片| 国产成人精品在线电影| 亚洲国产精品一区三区| 一区福利在线观看| 国产成人91sexporn| 久久久久国产精品人妻一区二区| 夫妻性生交免费视频一级片| 午夜免费鲁丝| 国产成人午夜福利电影在线观看| 又粗又硬又长又爽又黄的视频| 一级黄片播放器| 欧美最新免费一区二区三区| 制服人妻中文乱码| 久久久精品免费免费高清| 久久久精品国产亚洲av高清涩受| 美女视频免费永久观看网站| 伊人亚洲综合成人网| av国产精品久久久久影院| 欧美在线黄色| 最近最新中文字幕免费大全7| 蜜桃国产av成人99| 电影成人av| 咕卡用的链子| 看十八女毛片水多多多| 国产综合精华液| 美女中出高潮动态图| 国产国语露脸激情在线看| 欧美日韩av久久| 黑丝袜美女国产一区| 久久综合国产亚洲精品| 热99久久久久精品小说推荐| av网站在线播放免费| 亚洲美女搞黄在线观看| 91国产中文字幕| 热99国产精品久久久久久7| 国产在线视频一区二区| 亚洲精品国产色婷婷电影| 免费黄频网站在线观看国产| 最黄视频免费看| 亚洲五月色婷婷综合| 国产精品一区二区在线观看99| 国产日韩一区二区三区精品不卡| 日韩视频在线欧美| 少妇人妻久久综合中文| 另类精品久久| 母亲3免费完整高清在线观看 | 五月开心婷婷网| 国产高清国产精品国产三级| 人妻 亚洲 视频| 美女国产视频在线观看| 精品99又大又爽又粗少妇毛片| 亚洲一区二区三区欧美精品| 国产精品99久久99久久久不卡 | 亚洲国产av新网站| av女优亚洲男人天堂| 人人妻人人澡人人爽人人夜夜| 国产精品免费大片| 国产免费又黄又爽又色| 可以免费在线观看a视频的电影网站 | 男人添女人高潮全过程视频| 精品一区二区三区四区五区乱码 | 日日摸夜夜添夜夜爱| 少妇的逼水好多| 精品久久蜜臀av无| 国产日韩欧美亚洲二区| 国产国语露脸激情在线看| 激情视频va一区二区三区| 亚洲av欧美aⅴ国产| 国产免费一区二区三区四区乱码| 99re6热这里在线精品视频| 日本猛色少妇xxxxx猛交久久| 国产在线一区二区三区精| 街头女战士在线观看网站| 国产老妇伦熟女老妇高清| 人人妻人人澡人人爽人人夜夜| 观看美女的网站| 亚洲人成网站在线观看播放| √禁漫天堂资源中文www| 老女人水多毛片| 少妇人妻 视频| 婷婷成人精品国产| 国产黄频视频在线观看| 国产日韩欧美视频二区| 欧美亚洲 丝袜 人妻 在线| 亚洲四区av| 欧美日韩综合久久久久久| 啦啦啦视频在线资源免费观看| 免费看不卡的av| 亚洲综合色网址| av线在线观看网站| 人人妻人人澡人人看| 亚洲欧美一区二区三区国产| a级毛片在线看网站| 欧美人与性动交α欧美软件| 国产熟女欧美一区二区| 婷婷色综合www| 女性生殖器流出的白浆| 国产成人a∨麻豆精品| 久久久久精品人妻al黑| 青青草视频在线视频观看| 日韩欧美一区视频在线观看| 一边亲一边摸免费视频| 久久狼人影院| 亚洲欧洲日产国产| 国产午夜精品一二区理论片| 搡老乐熟女国产| 日韩精品免费视频一区二区三区| 有码 亚洲区| 精品国产一区二区三区四区第35| 久久久久国产一级毛片高清牌| 熟妇人妻不卡中文字幕| 久久99蜜桃精品久久| 亚洲欧美一区二区三区国产| 亚洲av中文av极速乱| 黄色视频在线播放观看不卡| 日韩一本色道免费dvd| 下体分泌物呈黄色| 日韩av不卡免费在线播放| 亚洲国产精品成人久久小说| 天天操日日干夜夜撸| 精品卡一卡二卡四卡免费| 老汉色av国产亚洲站长工具| 高清不卡的av网站| 亚洲经典国产精华液单| 亚洲国产最新在线播放| 欧美日韩综合久久久久久| 日日摸夜夜添夜夜爱| 精品少妇黑人巨大在线播放| 亚洲精品av麻豆狂野| 国产一区二区三区综合在线观看| xxxhd国产人妻xxx| 女性生殖器流出的白浆| 亚洲国产欧美网| 国产欧美亚洲国产| 国产成人免费无遮挡视频| 欧美激情极品国产一区二区三区| 啦啦啦在线免费观看视频4| 伦精品一区二区三区| av天堂久久9| 亚洲精品一二三| 男女高潮啪啪啪动态图| 午夜免费男女啪啪视频观看| 国产精品女同一区二区软件| 成年人免费黄色播放视频| 免费观看性生交大片5| 捣出白浆h1v1| 啦啦啦在线免费观看视频4| 在线天堂中文资源库| 久久97久久精品| 国产在线免费精品| av国产久精品久网站免费入址| 色94色欧美一区二区| 日日摸夜夜添夜夜爱| 黄色一级大片看看| 国产精品久久久久久av不卡| 在线天堂最新版资源| av国产久精品久网站免费入址| 黑丝袜美女国产一区| 咕卡用的链子| 亚洲av日韩在线播放| 免费在线观看视频国产中文字幕亚洲 | 免费观看性生交大片5| 老汉色av国产亚洲站长工具| av卡一久久| 黑丝袜美女国产一区| 国语对白做爰xxxⅹ性视频网站| 亚洲人成77777在线视频| 午夜91福利影院| 你懂的网址亚洲精品在线观看| 人妻人人澡人人爽人人| 国产成人精品久久二区二区91 | 天天躁夜夜躁狠狠躁躁| 精品一区二区三区四区五区乱码 | 欧美 亚洲 国产 日韩一| 欧美日韩综合久久久久久| 国产在视频线精品| 国产成人精品久久久久久| 亚洲av免费高清在线观看| 国产av一区二区精品久久| 欧美中文综合在线视频| 午夜免费男女啪啪视频观看| 校园人妻丝袜中文字幕| 伦理电影大哥的女人| videossex国产| 狠狠婷婷综合久久久久久88av| 国产精品嫩草影院av在线观看| 男女下面插进去视频免费观看| 国产 精品1| 日韩人妻精品一区2区三区| 最近的中文字幕免费完整| 亚洲精品日韩在线中文字幕| 黄色 视频免费看| 人人妻人人澡人人看| av在线app专区| av又黄又爽大尺度在线免费看| 国产淫语在线视频| 国产人伦9x9x在线观看 | 亚洲欧美清纯卡通| 母亲3免费完整高清在线观看 | 国产女主播在线喷水免费视频网站| 777久久人妻少妇嫩草av网站| 久久久亚洲精品成人影院| 亚洲三区欧美一区| 中文字幕人妻熟女乱码| 国产野战对白在线观看| 2018国产大陆天天弄谢| 欧美黄色片欧美黄色片| 国产精品久久久久久av不卡| 久久精品国产亚洲av高清一级| 男人爽女人下面视频在线观看| 久久久久久久久久久久大奶| 韩国高清视频一区二区三区| 人人妻人人澡人人看| 日本91视频免费播放| xxx大片免费视频| 国产av一区二区精品久久| 亚洲情色 制服丝袜| 国产欧美亚洲国产| 一本大道久久a久久精品| 交换朋友夫妻互换小说| 亚洲成人一二三区av| 亚洲欧美成人综合另类久久久| 免费黄频网站在线观看国产| www.熟女人妻精品国产| 免费大片黄手机在线观看| 亚洲欧美一区二区三区久久| 天天操日日干夜夜撸| 三级国产精品片| 欧美人与性动交α欧美精品济南到 | 在现免费观看毛片| 免费观看av网站的网址| 亚洲精品成人av观看孕妇| 免费观看a级毛片全部| 久久久久久久大尺度免费视频| 99热国产这里只有精品6| 日本91视频免费播放| 男女边吃奶边做爰视频| 欧美日本中文国产一区发布| 春色校园在线视频观看| 亚洲av国产av综合av卡| 成年av动漫网址| 久久久久国产一级毛片高清牌| 少妇的逼水好多| 免费观看性生交大片5| 色网站视频免费| 丝袜脚勾引网站| 日本av免费视频播放| 精品亚洲乱码少妇综合久久| 高清av免费在线| 一本一本久久a久久精品综合妖精 国产伦在线观看视频一区 | 精品视频人人做人人爽| 日韩成人av中文字幕在线观看| 亚洲人成77777在线视频| 男人爽女人下面视频在线观看| 亚洲精品,欧美精品| 自线自在国产av| 中文精品一卡2卡3卡4更新| a级片在线免费高清观看视频| 啦啦啦视频在线资源免费观看| 免费观看在线日韩| 韩国高清视频一区二区三区| 色婷婷av一区二区三区视频| 亚洲人成网站在线观看播放| 久久韩国三级中文字幕| 制服丝袜香蕉在线| 超碰97精品在线观看| 女人精品久久久久毛片| 国产不卡av网站在线观看| 在线观看免费高清a一片| 成人毛片60女人毛片免费| 亚洲精品在线美女| 十分钟在线观看高清视频www| av视频免费观看在线观看| 国精品久久久久久国模美| 美女脱内裤让男人舔精品视频| 亚洲国产精品成人久久小说| 五月伊人婷婷丁香| 精品一区二区三卡| 观看美女的网站| 中文字幕另类日韩欧美亚洲嫩草| 青春草国产在线视频| 少妇人妻久久综合中文| 亚洲人成网站在线观看播放| 丝袜美足系列| 午夜日本视频在线| 亚洲图色成人| 91久久精品国产一区二区三区| 欧美日本中文国产一区发布| 老司机影院成人| 亚洲色图 男人天堂 中文字幕| 老女人水多毛片| 青春草亚洲视频在线观看| 亚洲成人av在线免费| 欧美日韩精品网址| 欧美日韩一级在线毛片| 久久婷婷青草| 在线观看免费高清a一片| 交换朋友夫妻互换小说| 999久久久国产精品视频| 大片电影免费在线观看免费| 91精品三级在线观看| 我要看黄色一级片免费的| 青草久久国产| 久久久久久久亚洲中文字幕| 我的亚洲天堂| 中文字幕另类日韩欧美亚洲嫩草| 热re99久久精品国产66热6| 亚洲一级一片aⅴ在线观看| 久久av网站| 一级毛片我不卡| 亚洲精品日韩在线中文字幕| av在线app专区| 青春草亚洲视频在线观看| 国产精品av久久久久免费| 99热全是精品| 国产国语露脸激情在线看| 精品人妻在线不人妻| av卡一久久| 男女边摸边吃奶| 国产欧美日韩一区二区三区在线| 成年女人毛片免费观看观看9 | 热re99久久精品国产66热6| 中国三级夫妇交换| 一本一本久久a久久精品综合妖精 国产伦在线观看视频一区 | 亚洲成av片中文字幕在线观看 | 最近中文字幕2019免费版| 99久国产av精品国产电影| 波多野结衣一区麻豆| 美女午夜性视频免费| 免费观看无遮挡的男女| 天天操日日干夜夜撸| 成年av动漫网址| 精品一区在线观看国产| 精品国产露脸久久av麻豆| 热re99久久精品国产66热6| 大码成人一级视频| 国产一区二区激情短视频 | 菩萨蛮人人尽说江南好唐韦庄| 精品人妻在线不人妻| 亚洲婷婷狠狠爱综合网| 免费高清在线观看视频在线观看| 午夜福利视频在线观看免费| 亚洲精品久久午夜乱码| av网站免费在线观看视频| 欧美 日韩 精品 国产| 大片电影免费在线观看免费| 久久久国产精品麻豆| 99热全是精品| 中文字幕色久视频| 黄色一级大片看看| 在线免费观看不下载黄p国产| av免费在线看不卡| 老汉色av国产亚洲站长工具| 欧美日韩精品网址| 美女国产视频在线观看| 在线精品无人区一区二区三| 色婷婷久久久亚洲欧美| 少妇人妻 视频| 欧美激情 高清一区二区三区| 免费在线观看视频国产中文字幕亚洲 | 97精品久久久久久久久久精品| 亚洲国产av影院在线观看| 人妻一区二区av| 99久国产av精品国产电影| 日韩伦理黄色片| kizo精华| 又黄又粗又硬又大视频| 久久精品国产自在天天线| 国产成人午夜福利电影在线观看| 我要看黄色一级片免费的| 欧美bdsm另类| 久久毛片免费看一区二区三区| 一本久久精品| 最黄视频免费看| 久久久欧美国产精品| 亚洲精品国产av成人精品| 两性夫妻黄色片| h视频一区二区三区| 天天躁日日躁夜夜躁夜夜| 两性夫妻黄色片| 午夜免费观看性视频| 亚洲欧美一区二区三区久久| 男人操女人黄网站| 欧美变态另类bdsm刘玥| 欧美人与性动交α欧美软件| 久久精品久久久久久噜噜老黄| 欧美老熟妇乱子伦牲交| 国产老妇伦熟女老妇高清| 亚洲伊人久久精品综合| 99国产精品免费福利视频| 最近2019中文字幕mv第一页| 免费不卡的大黄色大毛片视频在线观看| 国产一区二区激情短视频 | 国产97色在线日韩免费| 在线天堂中文资源库| 麻豆乱淫一区二区|