• <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 ?

    Transfer Learning-based Computer-aided Diagnosis System for Predicting Grades of Diabetic Retinopathy

    2022-08-23 02:16:30QaisarAbbasMostafaIbrahim2andAbdulRaufBaig
    Computers Materials&Continua 2022年6期

    Qaisar Abbas,Mostafa E.A.Ibrahim2 and Abdul Rauf Baig

    1College of Computer and Information Sciences,Imam Mohammad Ibn Saud Islamic University(IMSIU),Riyadh,11432,Saudi Arabia

    2Department of Electrical Engineering,Benha Faculty of Engineering,Benha University,Qalubia,Benha,13518,Egypt

    Abstract:Diabetic retinopathy(DR)diagnosis through digital fundus images requires clinical experts to recognize the presence and importance of many intricate features.This task is very difficult for ophthalmologists and timeconsuming.Therefore,many computer-aided diagnosis(CAD)systems were developed to automate this screening process of DR.In this paper,a CAD-DR system is proposed based on preprocessing and a pre-train transfer learningbased convolutional neural network (PCNN) to recognize the five stages of DR through retinal fundus images.To develop this CAD-DR system,a preprocessing step is performed in a perceptual-oriented color space to enhance the DR-related lesions and then a standard pre-train PCNN model is improved to get high classification results.The architecture of the PCNN model is based on three main phases.Firstly, the training process of the proposed PCNN is accomplished by using the expected gradient length(EGL)to decrease the image labeling efforts during the training of the CNN model.Secondly,the most informative patches and images were automatically selected using a few pieces of training labeled samples.Thirdly, the PCNN method generated useful masks for prognostication and identified regions of interest.Fourthly, the DR-related lesions involved in the classification task such as micro-aneurysms,hemorrhages,and exudates were detected and then used for recognition of DR.The PCNN model is pre-trained using a high-end graphical processor unit(GPU)on the publicly available Kaggle benchmark.The obtained results demonstrate that the CAD-DR system outperforms compared to other state-of-the-art in terms of sensitivity(SE),specificity(SP),and accuracy(ACC).On the test set of 30,000 images,the CAD-DR system achieved an average SE of 93.20%,SP of 96.10%,and ACC of 98%.This result indicates that the proposed CAD-DR system is appropriate for the screening of the severity-level of DR.

    Keywords:Diabetic Retinopathy;retinal fundus images;computer-aided diagnosis system;deep learning;transfer learning;convolutional neural network

    1 Introduction

    One of the major causes of severe vision-loss among patients of diabetes is diabetic retinopathy(DR).DR is an asymptomatic disease, and it has no prior symptoms.However, many patients are suffered vision-loss without any proper diagnosis and treatment [1].According to the statistics of [2], 285 million population have diabetes, and one-third have signs of DR.In daily practice,ophthalmologists use non-mydriatic fundus images and Computer-aided diagnosis(CAD)programs for the early assessment/grade severity level of DR.The detection of lesions caused by DR is the basis of these earliest detection of DR.Those DR-related lesions in the fundus image are unhealthy objects appear on the retinal surface such as micro-aneurysms(MA’s),exudates(EX’s),hemorrhages(HEM’s)and cotton wool spots (CWS).The visual example of such DR-related lesions along with severitylevel is shown in Fig.1.Proliferative(PDR)and nonproliferative(NPDR)are two main types of DR,where PDR is advanced form of eye disease,and NPDR is an early sign of DR stage.As shown in this figure,there are five grades of severity-level of DR such as(severity 0:normal,severity 1:mild NPDR,severity 2: moderate NPDR, severity 3: severe NPDR and severity 4: PDR).Manual segmentation and count of DR-related lesions by clinicians is a difficult and repetitive task.Moreover,the manual grading of DR requires extensive domain-expert knowledge and reader inter-/intra-class variability experience[3–5].

    Figure 1:A visual example of five severity stages of DR and two main categories Proliferative(PDR)and nonproliferative(NPDR),where(a)Shows the(severity 0:normal),(b)Severity 1:mild NPDR,(c) Severity 2: moderate NPDR, (d) Severity 3: severe NPDR, and (e) Severity 4: PDR along with sample of DR-based lesions(f)

    Previously, several CAD systems have been developed to recognize grades of DR in the clinical setting by digital retinograph images.However,those CAD systems assist ophthalmologists to better screening of patients.As a result, the CAD systems helps clinical experts to identify the early signs of eye-related abnormality, which is difficult to identify by human naked eyes.It noticed that the DR is easily detected by a CAD system to grade the severity-level by using image processing and machine-learning techniques[6]on retinograph images.Nowadays,the CAD systems for recognition of grades of DR are affected by various factors such as (1) it is very much difficult to identify DRrelated lesions and anatomical structure of retinograph, (2) it is also difficult to detect accurate and early stage of retinal structure because it changes during the time,and(3)there is a dire need to develop the effective and automatic CAD system to accurate screening of DR-related diseases.Currently,many deep learning (DL) models especially deep convolution neural networks (CNN) have demonstrated outperform performance in the grading of DR severity-levels.Thus,we have used also deep transfer learning (TL) technique to recognize stages of DR with a pre-processing step to solve the abovementioned problems.In addition, the proposed system is capable to work on much larger datasets and to reduce the inter-reader variability.

    1.1 Research Highlights

    The main contribution of proposed CAD-DR classification system are as follows.

    1) A preprocessing step is developed in a perceptual-oriented CIEL*a*b*color space to enhance the contrast and adjust the light-illumination.

    2) A pretrain TL (DL) approach is used through expected gradient length (EGL) to eliminate the need of large number of labeled fundus images.This step reduces training efforts for CNN model.

    3) To develop PCNN, the 14-layer CNN network was pre-trained using fewer labeled fundus images.This can assist PCNN system to learn simple to complex fundus feature.

    4) Useful masks can be generated by the proposed PCNN system to predict and segment DRrelated regions.

    5) To the best of our knowledge, there is no previous CAD-DR model in the medical imaging field that works in harmony with CNN parameters to select the most informative patches and images.

    6) State-of-the-art comparisons are also performed to test and evaluate the performance of proposed CAD-DR system.

    1.2 Paper Organization

    The rest of the paper is organized as follows.Section 2 shows the literature review of the recent works related to the recognition of multistage of diabetic retinopathy(DR).In Section 3,the proposed methodology is described along with the acquisition dataset.Experimental results and comparisons with state-of-the-art methods are presented in Section 4.The discussion about the paper is described in Section 5 and finally,the paper concludes in Section 6.

    2 Related Works

    Deep learning (DL) models especially deep convolution neural networks (CNN) have demonstrated outperform performance in the grading of DR severity levels in several settings and on several datasets when compared with traditional hand-designed methods[6].A data science platform named Kaggle launched a DR detection competition in 2015,the top participants used different settings of CNN models on approximately 35,000 high-resolution labeled fundus images.They achieved that successful training of such CNN networks was based on the large size of annotated samples.In a previous study [7], a CNN model was developed with data augmentation to classify DR into five stages such as normal, mild NPDR, moderate NPDR, severe NPDR, and proliferative PDR.Their model was trained on more than 100,000 labeled images, yielded comparable performance with a clinical expert.Similarly,Harry et al.in[8]trained a 13-layer CNN model on 80,000 labeled images and obtained significant results in a classification of five severity-level of DR.Another study in[9],a training procedure of CNN model was completed using 8,810 images and obtained comparable results with ophthalmologists.This presents a challenge in clinical practice,as the computational systems need thousands of labeled images to be feed to learn features,representing a time-consuming and expensive process.For a real-time scenario,a well-performing algorithm requires such as a fewer data-intensive CNN model that learns with a few labeled samples.Although, it observed that the previous CAD systems tried to detect DR-related lesions[10]to recognize diabetic retinopathy.Those CAD systems are briefly described in the upcoming paragraphs and compared in Tab.1.

    Reference[11]shows the Faster-RCNN deep-learning(DL)based method to classify five stages of DR lesions without using image preprocessing step for contrast enhancement and adjustment of light illumination.To extract features from retinograph images, the authors used the DenseNet-65 DL model, and then the Faster-RCNN model is finally utilized to recognize the severity-level of DR.For evaluation of the Faster-RCNN model, they used Kaggle and APTOS datasets to achieve 97.2% of accuracy.The reference [12] showed that the current CAD systems for DR are expensive in computation and lack the ability to extract highly nonlinear features that are needed to classify it into five stages.In that study, they utilized the lowest possible learnable parameters were used to speed up the training and have faster convergence.They developed a VGG-NiN model based on the VGG16 transfer learning and spatial pyramid pooling layer.On collected datasets,they showed 83.5%classification accuracy on five stages of DR in comparison to other systems.Whereas in[13],a pretrain based transfer learning algorithm(CNN)was used to detect five stages of DR from retinograph images.They showed that the CNN model based on pretrain strategy achieved higher performance compared to other systems.Similarly,in[14],they used transfer learning(TL)with representational learning to recognize multiple stages of DR.They utilized the Inception-v4 TL pretrain model with fine-tune step.To grade DR into five stages,they used fine-tune to achieve 96.6%accuracy.

    Table 1:Computer-aided diagnosis systems to recognize grades of diabetic retinopathy.Performance based on sensitivity(SE),specificity(SP),accuracy(ACC),and area under the curve(AUC)are shown for deep learning algorithms(DLA)

    Table 1:Continued

    In this research[15],the author’s developed preprocessing-based segmentation along They used saliency maps detection to highlight anatomical structures of lesions compared to the background.Afterward,the structure tensor technique was applied to enhance the edges of the lesions and active contours is performed to accurately segment DR-related lesions.Finally, they used the VGG-19 pretrain TL model to identify the level of severity of DR.The experiments were performed on the Kaggle dataset consists of 20,000 images.On average, they reported 82% of sensitivity and 96% of accuracy.In contradiction with the above-mentioned approaches,the researchers in[16]developed a recognition system of three stages-based DR instead of five severity-level of DR.In that study,they used semantic segmentation to detect microaneurysm along with CNN model to recognize three-stages of DR.

    3 Proposed Methodology

    The Fig.2 shows a systematic flow diagram of our proposed CAD-DR system through pretrain transfer-learning based model.The CAD-DR system is developed in different phases.In the first phase,the retinograph image is transferred to perceptual-oriented CIE L*a*b*uniform color space and preprocessed it to adjust light illumination and enhance the contrast.In the next phase,the fourteen layers of CNN architecture is proposed by a pretrain transfer learning strategy.In first 10 layers,the different convolutional filters are used along with ReLU,BN and max pool layers.Next,the BN,Max pool and dropout layers are integrated and lastly, the ReLU and SoftMax layers are integrated to recognize five stages of DR.

    3.1 Data Acquisition and Platform

    To pretrain and evaluate the proposed CAD-DR system,the retinograph images are obtained from the Kaggle platform [17].The images in this dataset are captured from various patients in different light illumination,many age groups and different people’s ethnicity.Due to these variations,it makes distortion of pixel intensity within the image and creates other variations that affect the classification results.To overcome these issues,the contrast enhancement and light adjustment of retinograph images are implemented through uniform color space and non-linear wavelet technique as stated in[18].After image normalization,the dataset resized to 48×48 pixels which retained the fundus feature to identify and thus reduced memory size of dataset the GeForce GTX TITAN X 1080GPU could handle.Each image of the patient can only have one label corresponding to a single group depending on the divisions outlined for the dataset.Moreover,while testing to measure the performance of PCNN network,only unseen patches of patient images are considered.

    Figure 2:The proposed CAD-DR system to recognize five severity-level of diabetic retinopathy shown as a systematic flow diagram

    The proposed CAD-DR based on PCNN model is trained using 80,000 images from the publicly available Kaggle dataset.Each image has a resolution of 6 M pixel and rated by the clinician for the presence of DR into five graded such as 0-normal, 1-mild DR, 2-moderate DR, 3-severe DR and 4-proliferative DR.These scales were used as labels to develop PCNN model.The training and testing procedures of PCNN are accomplished by using 50,000/30,000 images.All experimental codes were written in Python 3.6 and deep learning package Keras (http://keras.io/) with the TensorFlow(http://deeplearning.net/software/tensorflow/)backend.These platforms are used because of their low computational time, easy access to parameters and maturity level.The GeForce GTX TITAN X 1080GPU having a memory of 12 GB is the hardware used for the experiments.The proposed PCNN model classified image into each DR class in 0.04 s that shows the possibility of real-time feedback to the patient.

    3.2 Data Augmentation

    To avoid overfitting and to improve the localization power of the proposed CAD-DR system,a dropout of 0.5 value on dense layers 12 and 13 with data augmentation such as flipping(horizontal and vertical)and random rotation 0?–270?degrees are utilized.After data augmentation and cropping steps,the dataset splits are undertaken.For this purpose,patches of each class are chosen randomly as follows: 8,760/1,314 patches for each of the classes in the training and testing splits.This data augmentation step is implemented through Albumentations library functions.

    3.3 Preprocessing to Enhance Contrast and Illumination Adjustment

    Retinograph images are captured from different devices and environment conditions.A visual example is displayed in the Fig.3.As a result,the preprocessing step is trying to enhance the patterns,which are presented in the DR-related lesions at the same time to decrease the training efforts in the classification phase.The selected space is kept as close as possible to human perception because the enhancement algorithm aims to help doctors in their diagnosis of retinopathy.

    Figure 3:A visual example of the preprocessing step to enhance the original input retinograph images(See Fig.(a)),correct light illumination(See Fig.(b)L* image)and improve contrast(See Fig.(c))in a perceptual-oriented color space

    In practice, the color retinograph images can be characterized in different color spaces such as HSV (hue, saturation, value), RGB (red, green, blue), CIELUV, etc.The uniform color space is dependent on the application because the color space is very important for image enhancement.The HSV and RGB are not uniform color spaces so they cannot be adopted for image enhancement.As a result, if choose the right color space then the image enhancement method is helping the ophthalmologists in the diagnosis eye-screening process.Hence,it is required that the selected space must be as close as possible to human perception.The CIE L*a*b*and CIE L*u*v*color spaces are closed to human perception,but the CIE L*u*v*color space has a problem of white adaptation that can lead to poor image enhancement results.Therefore,in this paper,we have used CIEL*a*b*color space.To perform image enhancement on retinograph images,the first step is transformed from the ununiform RGB image into uniform perceptual-oriented color space(CIEL*a*b*).To perform this step,The only available color spaces close to human perception are CIE L*a*b*and CIE L*u*v*and both have been extensively used.It may be noted that the white adaptation in CIE L*u*v*can lead to poorer results as mentioned before.The white adaptation has a subtractive change that involves a vector displacement instead of the multiplicative normalization that will produce the desired proportional movement.Therefore, our proposed algorithm initially transform the images from the RGB color space used by the acquisition device to CIE L*a*b*color space.

    Contrast enhancement method that preserves the characteristically features of the images using multiscale discrete-shearlet transform (DST), the perceptual uniform color space CIE L*a*b*and a local-influence control function.The DST technique has provided in the past an efficient multiscale directional representation of the image in a discrete framework and it is, therefore, better suited for multi-scale edge enhancement than the traditional wavelet decomposition.The method follows three main steps: firstly, the DST coefficients of L*plane in the corresponding subbands are modified by a Ben Graham’s method[19]to enhance the conditions of illumination for the images so that we can augment the perceptions from eye images and contrast adjustments,respectively.Secondly,the inverse transform is applied to modify L*coefficients for a better reconstruction and visualization without generating artifacts.Thirdly,a*and b*planes are combined with this lightness component to perform the final enhancement.

    3.4 Architecture of Pretrain Transfer Learning

    To develop this CAD-DR system, the pretrain CNN model is utilized as a basis and the corresponding layers are selected as required to recognize five stages of DR.This research uses a 14-layer CNN architecture, shown in Fig.4, after studying the literature for other complex image recognition tasks.It is perceived that an increased number of layers allows the network to learn the deepest features.For instance, the initial convolutional layer learns basic features like edges, while the last convolutional layer performs a learning process of DR lesions.This PCNN model consists of an input patch layer followed by the convolutional layer,max pooling,and fully connected layers.The soft-max classifier is used in the last fully connected layer to perform five severity levels of DR classification.Leaky rectifier linear unit (ReLU) with 0.01 value and then batch normalization was used as hyper-parameters after each convolutional layer to stop over-reliance on nodes in a network and to control feature maps per block.A kernel size of 3×3 and 2×2 was applied to perform max pooling.Similarly,the initialization of network layers was performed using the weights and biases from the method stated in[20].Gaussian distribution technique was also applied to initialize the network to reduce training time and to randomly generate biases for the last fully connected layer.

    Figure 4: Proposed architecture of pretrain convolutional neural network (PCNN) model by using transfer learning-based CNN network

    3.5 Procedure of Pretrain Transfer Learning

    The training of our CNN was accomplished using expected gradient length (EGL) [21,22] to decrease image labelling efforts during CNN training and to make CNN intelligent to learn features from the relevant data.It trained PCNN system from scratch using a well-known optimization algorithm called Stochastic Gradient Descent(SGD)to optimize parameters by utilizing one instance or sample batches instead of complete training samples.Eq.(1) illustrates SGD cost function J optimization using model parametersω:

    wherexis an input matrix,xiis an element of that matrix,?ωis a weight of the convolution filter,b is the bias term,max(.)is the maximum function andxjrepresents the region ofxwhere the pooling operation is applied.

    In Eq.(1),Jc(?i)represents a cost function applied at ith training sample(xi,yi)in iteration i,σmeans learning rate and shows a gradient operator.A training sample ith with its label is then used to estimate the cost function Jc(?i)and its gradient length|| Jc(?i)||.To select the most relevant image patches,each batch of SGD depends on the highest gradient value of instance,having a probability of sample with ythlabel.A term l in Eq.(2)indicates total number of labels,whileΦrepresents sort values used by EGL algorithm from unlabeled data pool U.The selection mechanism of most informative samples is performed using the calculation of two terms of Eq.(1) such as, a probability of sample with jthlabel,forward propagation across the network is performed to get same probabilities from the last dense layer.Whereas gradient length is calculated using backward propagation to get frobenius norm of gradient parameter.This strategy is then repeated for all labels of each sample.At last, k sample with highest EGL value is chosen from the sort data poolΦ.Algorithm 1 shows the patch selection steps by our PCNN network.After calculation of the significant patches by Algorithm 1, it is straightforward to extend our experiments to choose the most informative images within training dataset.This was done by the calculation of interestingness of an image using image squares(patches) with the given stride and then densely calculatingΦ.Afterwards, image sortation by their top EGL value was done and then patches belong to the most relevant image added to the training data for further parameter updates by Algorithm 1 until convergence.These steps were described in Algorithm 2.

    Algorithm 1:A patch selection step to train PCNN model Requirement: Labeled(patch)dataset P,initial trained model M using patches in P∈Input,k samples of most informative patches 1.While no convergence do 2.Generate and mix sample batches from P 3.For each batch do 4.calculate Φ(x)using M,for all x €batch 5.end for 6.perform sorting process of Θ values and return highest k sample Pk 7.update M via P ∪Pk 8.end while Requirement: Labeled(patch)dataset P,initial trained model M using patches in P∈Input,k samples of most informative patches

    Algorithm 2:A patch selection step to train PCNN model Requirement: Labeled(patch)dataset P,training set T,number n of initial images to look at 1.Chose randomly an initial set Tn of images 2.An initial training of model M using the expert notions from the n images 3.While no convergence do 4.For each image in T/Tn do 5.Patched image and calculate ?image =∑patch ∈input_image ?(patch),using M 6.end for 7.Perform sorting process on all ?_image values and return an image with highest Imax 8.Tn =Tn ∪Imax 9.Pn =patch ∈Pi, for all i∈Tn 10.Update M with the patches Pn and k selected patches via Algorithm 1

    To calculate the loss function of PCNN model,the categorical cross-entropy(CCE)loss function is utilized and applied after SoftMax function for taking the final decision of five severity level of diabetic retinopathy.The CCE loss function is also known as SoftMax Loss.In practice, it is a combination of SoftMax activation with a cross-entropy (CE) loss.If the CCE loss is used, then the PCNN network model is trained to output a probability over the CC classes for each image.The CE Loss is defined as:

    where, the parameterSPis the PCNN score for the positive class.This loss function then able to compute the gradient with respect to the output neurons of the PCNN model to backpropagate it through the network and boost the defined loss function tuning the network parameters.As a result,it is necessary to calculate the gradient of CE Loss with respect to each PCNN class score in ss.The loss terms are zero for the negative classes.The Fig.5 shows the visual plots of lossvs.validation of the proposed PCCN model.

    Figure 5: Proposed architecture of pretrain convolutional neural network (PCNN) model by using transfer learning-based CNN network

    4 Experimental Results

    4.1 Hyperparameter Settings

    The proposed PCNN model to recognize severity-level of DR is formed by stacking the fourteen layers of the network including dropout and SoftMax layers.Transfer learning is used for the convolutional layers of the PCNN network.Fine tuning is also applied to the fully connected layers of the PCNN.The hyper-parameters along with their values are shown in Tab.2.An adaptive learning rate is used so that the learning process can be speeded up and over-fitting can be avoided.Initially,the learning rate is set to 0.01.Iterations are monitored and if the validation loss does not improve for five consecutive iterations, then the learning rate is decreased by a factor of 0.1.In addition, we used batch size of 8,minimum learning rate of 0.0001,initial learning rate of 0.01,momentum of 0.9,number of epochs of 24 and network layers of 14 to test and train the PCNN model.

    Table 2:Performance of the proposed PCNN with preprocessing step model for classification of fiveclass severity level of DR on 30,000 test images

    4.2 Statistical Metrics

    Five evaluation metrics have been used to evaluate and compare our model with other systems.These commonly used metrics are accuracy,F1-score,sensitivity,specificity,and ROC-AUC.The F1 score is the harmonic mean of the precision and recall.We have included F1 score as a metric because there is a large class imbalance.Since F1 score is the harmonic mean of precision and recall, it is considered a better metric than accuracy in such cases.A higher F1 score implies a better system.Sensitivity is used to measure true positive rate which in our case means the correct identification of vessel pixels.In contrast,specificity measures the true negative rate and that corresponds to the ability of identification of non-vessel pixels by our model.AUC is used to measure the ability of the model at discriminating between vessel pixels and non-vessel pixels.ROC graph plots the true positive rate against the true negative rate at various thresholds and AUC is the area under the ROC curve.The higher the AUC the better the model.The suitability of the proposed PCNN system for five severity level of DR was evaluated on 30,000 test images using the statistical metrics, i.e., sensitivity (SE),specificity(SP)and an accuracy.We define SE as the number of images correctly classified as having DR among the total amount with DR and SP as the number of images correctly identified as having no DR out of the total number with no DR.The accuracy is defined as the number of patients that are correctly classified by the system.

    4.3 Results Analysis and Comparisons

    This retinograph images’dataset is divided into 40%of the testing set and the rest as the training set.Also, we have split again the training set into 40% and assigned it to the validation set, and the rest is used for training purposes.On a total of 30,000 images in the dataset,the 60%is working as the training set,20%as the validation set,and 20%as the test set.There are 24 epochs performed based on the 10-fold cross-validation set.The Fig.5 displays the proposed model training and testing loss versus accuracy diagram.From this figure, it is noticed that this plot is displayed without performing any fine-tune of the proposed model.In addition,this figure shows that the change between the predicted stage of DR by our model and the true value in the form of the loss function,which is measured by a category cross-entropy.

    Fig.6 shows the confusion metrics to achieve the results for recognition of five stages of diabetes retinopathy.This confusion metric is calculated based on 20,000 retinograph images.This metric is calculated based on proposed architecture of pretrain convolutional neural network(PCNN)model by using transfer learning-based CNN network.On average,the 0.90 detection accuracy is obtained to predict five stages.However,if training and testing datasets are increased to 30,000 then the detection accuracy is increased too.The Tab.2 reports the highly acceptable SE values for normal (98.15%),mild (93.45%) and proliferative (90.45%) DR classes of the proposed CAD-DR system by using preprocessing step and pretrain PCNN architecture.While the values of SP and classification accuracy for five-classes of DR were found up-to-the-mark.The proposed PCNN is significantly improved in SE of 93.20%,SP of 96.10%and an accuracy of 98%on the 30,000 test samples.However,the Tab.3 shows the lower results because we did not use preprocessing step.We have also compared the PCNN transfer learning model with other transfer learning(TL)algorithms such as VGG16,VGG16noFC1,VGG16noFC2 and InceptionV3.On average,the results are mentioned in Tabs.4 and 5 describes the parameter used to compare the different TL models.The PCNN model is outperformed compared to all other TL algorithms because of use of effective layers with loss function.

    We have also performed comparisons of the proposed CAD-DR system with other state-of-theart systems such as CNN-Pratt-2016 [8], DenseNet-Albahli-2021 [11], and VGG-Khan-2021 [12] in terms of recognition of five stages of DR.We have implemented by ourselves other state-of-the-art DR systems such as CNN-Pratt-2016, DenseNet-Albahli-2021 and VGG-Khan-202.The authors are requested to read those papers for detailed implementation of these papers.We have selected these DR-related systems because those are closely related to our proposed CAD-DR system.Those comparisons are performed and evaluated based on different training and testing ratios on 30,000 test images.Tab.6 indicates the performance of the proposed PCNN model that is outperformed compared to other systems for recognition of five stages of DR.

    Figure 6:Confusion metric of pretrain convolutional neural network(PCNN)model by using transfer learning-based CNN network on 20,000 retinograph images

    Table 3: Performance of the proposed PCNN model for classification of five-class severity level without preprocessing step on DR of 30,000 test images

    Table 4: Comparison setup parameters to other state-of-the-art transfer learning algorithms

    Table 5:Performance comparisons of proposed PCNN art with different pre-trained CNN networks on 30,000 test images for five-class severity level of DR

    Table 6:Performance comparisons of proposed PCNN art with different pre-trained CNN networks on 30,000 test images for five-class severity level of DR

    5 Discussions

    Statistics show that many diabetic patients have a high probability of severe vision-loss by diabetic retinopathy (DR).In daily practice, ophthalmologists use non-mydriatic fundus images and Computer-aided diagnosis (CAD) programs for the early assessment/grade severity level of DR.Those earliest assessment of DR is based on the recognition of lesions related to DR.These DR-related lesions appear as unhealthy objects such as micro-aneurysms (MA’s), exudates (EX’s),hemorrhages (HEM’s) and cotton wool spots (CWS) on the retinal surface in the fundus image.Manual segmentation and count of DR-related lesions by clinicians is a difficult and repetitive task.Moreover, the manual grading of DR requires extensive domain-expert knowledge and reader inter-/intra-class variability experience.As a result,several CAD systems have been developed in the past to recognize grades of DR in the clinical setting by digital retinograph images.However, those CAD systems assist ophthalmologists to better screening of patients.As a result, the CAD systems helps clinical experts to identify the early signs of eye-related abnormality,which is difficult to identify by human naked eyes.It noticed that the DR is easily detected by a CAD system to grade the severitylevel by using image processing and machine-learning techniques[6]on retinograph images.Nowadays,the CAD systems for recognition of grades of DR are affected by various factors such as(1)it is very much difficult to identify DR-related lesions and anatomical structure of retinograph, (2) it is also difficult to detect accurate and early stage of retinal structure because it changes during the time,and(3) there is a dire need to develop the effective and automatic CAD system to accurate screening of DR-related diseases.Thus,there is a dire need to make DR diagnosis on much larger datasets and to reduce the inter-reader variability.

    Deep learning (DL) models especially deep convolution neural networks (CNN) have demonstrated outperform performance in the grading of DR severity levels in several datasets and settings compared to traditional hand-designed methods[6].A data science platform named Kaggle launched a DR detection competition in 2015,the top-most competitors used different settings of CNN models on approximately 35,000 high-resolution labeled fundus images.They achieved that successful training of such CNN networks was based on the large size of annotated samples.In a previous study[7],a CNN model was developed with data augmentation to classify DR into five stages such as normal, mild NPDR, moderate NPDR, severe NPDR, and proliferative PDR.Their model was trained on more than 100,000 labeled images,yielded comparable performance with a clinical expert.Similarly,Harry et al.in[8]trained a 13-layer CNN model on 80,000 labeled images and obtained significant results in a classification of five severity-level of DR.Another study in[9],a training procedure of CNN model was completed using 8,810 images and obtained comparable results with ophthalmologists.This presents a challenge in clinical practice,as the computational systems need thousands of labeled images to be feed to learn features,representing a time-consuming and expensive process.For a real-time scenario,a well-performing algorithm requires such as a fewer data-intensive CNN model that learns with a few labeled samples.Although,it observed that the previous CAD systems tried to detect DR-related lesions[10]to recognize diabetic retinopathy.Those CAD systems have already been briefly described in the Tab.1.

    To develop this CAD-DR system,a preprocessing step is performed in a perceptual-oriented color space to enhance the DR-related lesions and then a standard pre-train PCNN model is improved to get high classification results.The architecture of the PCNN model has three main phases.Firstly,the training process of the proposed PCNN is accomplished by using the expected gradient length(EGL) to decrease the image labeling efforts during the training of the CNN model.Secondly, the most informative patches and images were automatically selected using a few pieces of training labeled samples.Thirdly,the PCNN method generated useful masks for prognostication and identified regions of interest.Fourthly,the DR-related lesions related to the classification task such as micro-aneurysms,hemorrhages,and exudates were detected and then used for recognition of DR.The PCNN model is pre-trained on the publicly available Kaggle benchmark making use of a high-end graphical processor unit(GPU).The obtained results demonstrate that the CAD-DR system outperforms when compared to other latest, state-of-the-art systems in terms of sensitivity (SE), specificity (SP), and accuracy(ACC).On the test set of 30,000 images, the CAD-DR system obtained an average SE of 93.20%,SP of 96.10%,and ACC of 98%.Some of the example images are described in Fig.7 that are correctly classified by proposed PCNN model.We have achieved good results based on several improvements to the CAD-DR system such as a preprocessing step is developed in a perceptual-oriented CIEL*a*b*color space to enhance the contrast and adjust the light-illumination.A pretrain TL(DL)approach is used through expected gradient length(EGL)to eliminate the need of large number of labeled fundus images.This step reduces training efforts for CNN model.This result indicates that the proposed CAD-DR system is appropriate for the screening of the DR severity-levels.

    Figure 7:Fundus images depicting the five stages of diabetic retinopathy:(a)Without DR,(b)Mild,(c)Moderate,(d)Severe and(e)PDR

    6 Conclusions

    In this paper, a new pre-train scheme of the CNN model (PCNN) is presented to develop a label efficient training mechanism in the domain of retinal fundus images for diagnosis of DR.In addition,we have also developed a preprocessing step in a perceptual-oriented color space to enhance the contrast and adjust the light illumination.This proposed CAD-DR system is outperformed compared to other state-of-the-art systems on 30,000 retinograph images.The DR-related lesion patterns are identified by the proposed PCNN system into five classes.Moreover, an additional interpretation layer is utilized to identify those image areas that should be labeled by the clinical expert.In this paper,an improved computer-aided diagnosis(CAD)system to assist ophthalmologists was developed.The presented PCNN architecture was evaluated on 80 thousand fundus images,and the achieved results illustrate the feasibility of the presented grading system for DR-related lesion detection and classification of the five-class severity level of DR.PCNN system was found useful for human evaluation such as the high value of SE and SP rates.To complete this PCNN system, fast image analysis methods with stable interpretation are utilized.Upon completion of PCNN model training, the classification of DR stages was accomplished in 0.04 s.Keeping these conditions, the obtained SE and SP rates indicate that a vast majority of the images were accurately classified into one of the five-stages of DR.To the best of our knowledge,there is no previous CAD-DR model in the medical imaging field that works in harmony with CNN parameters to select the most informative patches and images.The proposed PCNN method is computationally challenged when dealing with large-scale data.This issue can be resolved with standard sampling techniques.As future work, the other existing deep-learning (DL) methods scheme can be utilized to train a CNN model detecting and classifying diabetic maculopathy on the large-scale annotated datasets.

    Acknowledgement:The authors extend their appreciation to the Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University for funding this work through Research Group no.RG-21-07-01.

    Consultant Works:We would also like to thank Dr.M.Arfan Jaffar for serving as a consultant to critically reviewed the study proposal and participated in technical editing of the manuscript.

    Funding Statement:Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University for funding this work through Research Group no.RG-21-07-01.

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

    一本一本综合久久| 在线免费观看不下载黄p国产| 婷婷色av中文字幕| 人妻 亚洲 视频| 肉色欧美久久久久久久蜜桃 | av线在线观看网站| 免费大片黄手机在线观看| 久久热精品热| 亚洲内射少妇av| 亚洲精品国产av成人精品| 亚洲成人一二三区av| 少妇的逼水好多| 国产成人精品久久久久久| 精品久久久久久电影网| 91久久精品国产一区二区成人| 少妇人妻久久综合中文| 国产高清有码在线观看视频| 国产亚洲一区二区精品| 99re6热这里在线精品视频| 天天躁夜夜躁狠狠久久av| 国产男人的电影天堂91| 99久久精品国产国产毛片| 极品教师在线视频| 人妻制服诱惑在线中文字幕| 又大又黄又爽视频免费| 18禁在线无遮挡免费观看视频| 91精品国产九色| 久久这里有精品视频免费| 亚洲av中文字字幕乱码综合| 国内少妇人妻偷人精品xxx网站| 国产乱人视频| 久久人人爽人人爽人人片va| 精品人妻熟女av久视频| 午夜免费男女啪啪视频观看| 国产精品福利在线免费观看| 欧美性感艳星| 极品教师在线视频| 国产毛片a区久久久久| 97人妻精品一区二区三区麻豆| 99九九线精品视频在线观看视频| 女人被狂操c到高潮| 久久亚洲国产成人精品v| 日韩一本色道免费dvd| 国产黄a三级三级三级人| 国产精品熟女久久久久浪| 久久久a久久爽久久v久久| 美女脱内裤让男人舔精品视频| 久久久成人免费电影| 少妇猛男粗大的猛烈进出视频 | 99热网站在线观看| 国产精品嫩草影院av在线观看| 涩涩av久久男人的天堂| 大话2 男鬼变身卡| 国国产精品蜜臀av免费| 狂野欧美白嫩少妇大欣赏| 天堂网av新在线| 看黄色毛片网站| 国产免费一区二区三区四区乱码| 51国产日韩欧美| 久久久久久久久久久丰满| 国产成人freesex在线| 午夜日本视频在线| 精品午夜福利在线看| 国产精品精品国产色婷婷| 少妇猛男粗大的猛烈进出视频 | 国产成人freesex在线| 亚洲成人av在线免费| h日本视频在线播放| 夫妻性生交免费视频一级片| 特级一级黄色大片| 亚洲精华国产精华液的使用体验| 亚洲成色77777| 亚洲精品aⅴ在线观看| 亚洲熟女精品中文字幕| 深夜a级毛片| 麻豆久久精品国产亚洲av| 欧美xxxx黑人xx丫x性爽| 美女主播在线视频| 2018国产大陆天天弄谢| 偷拍熟女少妇极品色| 精品久久久久久久久av| 一级a做视频免费观看| 久久人人爽av亚洲精品天堂 | 一个人观看的视频www高清免费观看| 午夜激情福利司机影院| 亚洲精品中文字幕在线视频 | 亚洲欧洲国产日韩| 精品午夜福利在线看| 女人十人毛片免费观看3o分钟| 最近手机中文字幕大全| av女优亚洲男人天堂| 中文字幕亚洲精品专区| 国产伦在线观看视频一区| 成人毛片a级毛片在线播放| 国产精品av视频在线免费观看| xxx大片免费视频| 一级二级三级毛片免费看| 69人妻影院| 校园人妻丝袜中文字幕| 毛片一级片免费看久久久久| 男插女下体视频免费在线播放| 亚洲欧美清纯卡通| 精品人妻熟女av久视频| 成人鲁丝片一二三区免费| 国产爽快片一区二区三区| 国产精品麻豆人妻色哟哟久久| 亚洲在线观看片| 赤兔流量卡办理| 久久久欧美国产精品| 在线精品无人区一区二区三 | 国产熟女欧美一区二区| 亚洲电影在线观看av| 日韩av不卡免费在线播放| 97人妻精品一区二区三区麻豆| 少妇丰满av| 777米奇影视久久| 一本久久精品| av国产久精品久网站免费入址| 一级毛片我不卡| 少妇人妻精品综合一区二区| 韩国av在线不卡| 久久ye,这里只有精品| 日日摸夜夜添夜夜爱| 国产免费一区二区三区四区乱码| 色婷婷久久久亚洲欧美| 亚洲国产最新在线播放| 丝袜美腿在线中文| 日韩欧美精品免费久久| 日韩免费高清中文字幕av| 亚洲国产精品成人久久小说| 国产视频内射| 内地一区二区视频在线| 国产探花在线观看一区二区| 国产伦理片在线播放av一区| 久久久久国产网址| 免费观看av网站的网址| 日本-黄色视频高清免费观看| 国产熟女欧美一区二区| 欧美精品人与动牲交sv欧美| av播播在线观看一区| 日韩精品有码人妻一区| 免费av毛片视频| 少妇人妻久久综合中文| 一级a做视频免费观看| 国产黄片视频在线免费观看| av国产久精品久网站免费入址| 97精品久久久久久久久久精品| 免费看日本二区| 久久精品国产亚洲av天美| 老女人水多毛片| 成人欧美大片| 国产伦在线观看视频一区| 日本黄大片高清| av国产免费在线观看| 亚州av有码| 卡戴珊不雅视频在线播放| 国产成人aa在线观看| 久热久热在线精品观看| 久久精品国产鲁丝片午夜精品| 天堂网av新在线| 亚洲最大成人中文| 国产av码专区亚洲av| 国产乱人视频| 亚洲美女视频黄频| 亚洲av一区综合| 亚洲综合精品二区| 亚洲精品成人久久久久久| 一级片'在线观看视频| 免费观看性生交大片5| 久久久亚洲精品成人影院| 久久久久久久大尺度免费视频| 成年av动漫网址| 中国美白少妇内射xxxbb| 欧美精品一区二区大全| 精品久久久久久电影网| 精品久久国产蜜桃| 国内精品宾馆在线| 狂野欧美白嫩少妇大欣赏| 久久精品国产亚洲av天美| 男人添女人高潮全过程视频| 国产精品不卡视频一区二区| 久久精品国产亚洲av涩爱| 日本免费在线观看一区| 国产精品成人在线| 深爱激情五月婷婷| 欧美成人一区二区免费高清观看| 国产v大片淫在线免费观看| 男人添女人高潮全过程视频| 国产午夜精品一二区理论片| 在线播放无遮挡| 国产毛片a区久久久久| 波野结衣二区三区在线| 精品国产露脸久久av麻豆| 我的老师免费观看完整版| 国产高潮美女av| 国产黄频视频在线观看| 亚洲成人精品中文字幕电影| 久久精品久久久久久噜噜老黄| 国产精品秋霞免费鲁丝片| 亚洲自偷自拍三级| 亚洲欧洲日产国产| 久久影院123| 日本av手机在线免费观看| 嫩草影院新地址| 嫩草影院入口| 欧美xxxx黑人xx丫x性爽| 久久久色成人| a级毛色黄片| 晚上一个人看的免费电影| 丝瓜视频免费看黄片| 一级爰片在线观看| 久热这里只有精品99| 日本熟妇午夜| 在线a可以看的网站| 夜夜看夜夜爽夜夜摸| 22中文网久久字幕| 日韩制服骚丝袜av| 国产男人的电影天堂91| 日韩成人av中文字幕在线观看| 免费观看无遮挡的男女| 亚洲综合精品二区| 国产老妇伦熟女老妇高清| 国产亚洲一区二区精品| 直男gayav资源| 欧美潮喷喷水| 国产精品.久久久| 国产高清有码在线观看视频| 婷婷色综合www| 亚洲欧美中文字幕日韩二区| 天堂俺去俺来也www色官网| 中文字幕人妻熟人妻熟丝袜美| 国产色爽女视频免费观看| 看黄色毛片网站| 亚洲最大成人中文| 亚洲欧美日韩东京热| 精品一区二区三卡| 久久久a久久爽久久v久久| av在线app专区| 99久久精品一区二区三区| 精品久久久久久久人妻蜜臀av| 国产一区二区三区综合在线观看 | 三级国产精品欧美在线观看| 久久久久久九九精品二区国产| 黄片wwwwww| 亚洲av福利一区| 欧美激情在线99| 精品久久久久久久末码| 麻豆乱淫一区二区| 国产高清三级在线| 亚洲在久久综合| 看十八女毛片水多多多| av网站免费在线观看视频| 欧美日韩精品成人综合77777| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 亚洲国产最新在线播放| 亚洲国产最新在线播放| 简卡轻食公司| 天堂俺去俺来也www色官网| 久久久a久久爽久久v久久| 亚洲国产日韩一区二区| 欧美精品人与动牲交sv欧美| 久久国产乱子免费精品| 777米奇影视久久| 亚洲成人久久爱视频| 尾随美女入室| 欧美三级亚洲精品| 成人黄色视频免费在线看| 黄片wwwwww| a级毛片免费高清观看在线播放| 六月丁香七月| 韩国高清视频一区二区三区| 精品一区二区三卡| www.av在线官网国产| 亚洲精品aⅴ在线观看| 视频区图区小说| 男女边吃奶边做爰视频| 最近中文字幕高清免费大全6| 午夜激情福利司机影院| 精品一区二区三卡| .国产精品久久| 国产精品偷伦视频观看了| 久久久国产一区二区| 丰满少妇做爰视频| 国产爽快片一区二区三区| 91精品一卡2卡3卡4卡| 国内精品美女久久久久久| 久久久成人免费电影| 欧美日韩视频高清一区二区三区二| 日韩亚洲欧美综合| 涩涩av久久男人的天堂| 白带黄色成豆腐渣| 国产真实伦视频高清在线观看| 亚洲精品国产色婷婷电影| 国产亚洲av嫩草精品影院| 亚洲不卡免费看| 亚洲欧美成人精品一区二区| 国产午夜福利久久久久久| 久久久久精品久久久久真实原创| 美女xxoo啪啪120秒动态图| 日韩 亚洲 欧美在线| 真实男女啪啪啪动态图| 波野结衣二区三区在线| 亚洲天堂av无毛| 精品一区在线观看国产| 国产淫片久久久久久久久| av在线老鸭窝| av在线观看视频网站免费| 亚洲欧美日韩卡通动漫| 国产成人精品一,二区| 国产黄色免费在线视频| 欧美成人a在线观看| 性色av一级| 久久久欧美国产精品| 天天躁夜夜躁狠狠久久av| 一级片'在线观看视频| 性色avwww在线观看| 亚洲久久久久久中文字幕| 国产成人freesex在线| 亚洲自拍偷在线| 狂野欧美激情性xxxx在线观看| 国产精品99久久99久久久不卡 | 亚洲欧美成人综合另类久久久| 成人漫画全彩无遮挡| 亚洲精品第二区| 久久久欧美国产精品| 午夜日本视频在线| 精品久久国产蜜桃| 成年人午夜在线观看视频| 国产成人a区在线观看| 制服丝袜香蕉在线| 亚洲av男天堂| 国产黄片视频在线免费观看| 久久人人爽人人片av| 在线观看一区二区三区激情| 精品国产一区二区三区久久久樱花 | 3wmmmm亚洲av在线观看| 少妇人妻 视频| 欧美性感艳星| 国产视频内射| 在线播放无遮挡| 2022亚洲国产成人精品| 国产精品爽爽va在线观看网站| 亚洲精品日本国产第一区| 真实男女啪啪啪动态图| 夜夜看夜夜爽夜夜摸| 亚洲自偷自拍三级| 国产片特级美女逼逼视频| 国产精品人妻久久久久久| 男人舔奶头视频| 一级a做视频免费观看| av女优亚洲男人天堂| av专区在线播放| 在线观看av片永久免费下载| 精品午夜福利在线看| 日本欧美国产在线视频| 夜夜爽夜夜爽视频| 91久久精品电影网| 亚洲精品乱久久久久久| 婷婷色综合www| 国产免费又黄又爽又色| 国产成人精品福利久久| 亚洲欧美成人综合另类久久久| 69av精品久久久久久| 国产一级毛片在线| 国产成人精品福利久久| 草草在线视频免费看| 一级毛片我不卡| 国产高清三级在线| 日本黄大片高清| 精品久久国产蜜桃| 一级黄片播放器| 日韩成人伦理影院| 亚洲真实伦在线观看| 美女cb高潮喷水在线观看| 在线观看一区二区三区激情| 国产精品无大码| 少妇猛男粗大的猛烈进出视频 | 亚洲欧美日韩无卡精品| 国精品久久久久久国模美| 亚洲av在线观看美女高潮| 欧美三级亚洲精品| 中文精品一卡2卡3卡4更新| 在现免费观看毛片| 成人午夜精彩视频在线观看| 亚洲精品久久久久久婷婷小说| 自拍欧美九色日韩亚洲蝌蚪91 | 人体艺术视频欧美日本| 97超视频在线观看视频| 午夜福利视频1000在线观看| 国产成人精品一,二区| 永久免费av网站大全| 激情五月婷婷亚洲| 免费黄色在线免费观看| 日韩一区二区视频免费看| 亚洲精品aⅴ在线观看| a级毛片免费高清观看在线播放| 国内精品宾馆在线| 国产在线男女| 国产精品国产三级专区第一集| 听说在线观看完整版免费高清| 免费看av在线观看网站| 伦理电影大哥的女人| 七月丁香在线播放| 久久久久网色| 中文资源天堂在线| 国产午夜精品久久久久久一区二区三区| av国产精品久久久久影院| 久久精品国产a三级三级三级| 精品国产一区二区三区久久久樱花 | 日本欧美国产在线视频| 国产伦精品一区二区三区四那| 日本与韩国留学比较| 我要看日韩黄色一级片| 久热久热在线精品观看| 在线a可以看的网站| 一边亲一边摸免费视频| 国产乱来视频区| 国产探花极品一区二区| 99热网站在线观看| 免费在线观看成人毛片| 大香蕉久久网| 国产亚洲91精品色在线| 亚洲精品第二区| 少妇高潮的动态图| 国产男人的电影天堂91| 干丝袜人妻中文字幕| 免费av不卡在线播放| 亚洲自偷自拍三级| 国产视频内射| 亚洲av二区三区四区| 中国三级夫妇交换| 最近的中文字幕免费完整| 十八禁网站网址无遮挡 | 日本黄色片子视频| 波野结衣二区三区在线| 午夜激情久久久久久久| 麻豆乱淫一区二区| www.av在线官网国产| 大码成人一级视频| 欧美丝袜亚洲另类| 亚洲国产日韩一区二区| 一本色道久久久久久精品综合| 少妇 在线观看| 精品酒店卫生间| 亚洲国产精品专区欧美| 国产黄色视频一区二区在线观看| 91狼人影院| 亚洲av成人精品一二三区| 男插女下体视频免费在线播放| 成人欧美大片| 日韩中字成人| 在现免费观看毛片| 欧美变态另类bdsm刘玥| 黄片无遮挡物在线观看| 最后的刺客免费高清国语| 人人妻人人看人人澡| 精品国产乱码久久久久久小说| 禁无遮挡网站| 18禁动态无遮挡网站| 亚洲av不卡在线观看| 麻豆乱淫一区二区| 久久国内精品自在自线图片| 黄色怎么调成土黄色| 天堂网av新在线| 亚洲欧美精品自产自拍| 18禁在线播放成人免费| videos熟女内射| 中文欧美无线码| 禁无遮挡网站| 亚洲欧美日韩东京热| 亚洲欧洲日产国产| 身体一侧抽搐| 免费电影在线观看免费观看| 最后的刺客免费高清国语| 亚洲美女搞黄在线观看| 在线观看av片永久免费下载| 少妇裸体淫交视频免费看高清| 中文字幕制服av| 一个人看视频在线观看www免费| 日本午夜av视频| 天天躁日日操中文字幕| 欧美成人午夜免费资源| 国内精品美女久久久久久| 久久综合国产亚洲精品| 尤物成人国产欧美一区二区三区| 久久久午夜欧美精品| 性色avwww在线观看| 日日摸夜夜添夜夜添av毛片| 爱豆传媒免费全集在线观看| 国产精品一二三区在线看| 欧美日韩一区二区视频在线观看视频在线 | 亚洲精品成人久久久久久| 免费在线观看成人毛片| 永久免费av网站大全| 嫩草影院入口| 大又大粗又爽又黄少妇毛片口| 2021天堂中文幕一二区在线观| 亚洲综合精品二区| 亚洲欧美日韩东京热| 久久久精品94久久精品| 又爽又黄a免费视频| 欧美日韩在线观看h| 精品一区二区三区视频在线| 免费看光身美女| 国产av不卡久久| 丝瓜视频免费看黄片| freevideosex欧美| 五月天丁香电影| 人人妻人人爽人人添夜夜欢视频 | 最近中文字幕高清免费大全6| 赤兔流量卡办理| 丰满人妻一区二区三区视频av| 免费看av在线观看网站| 国产熟女欧美一区二区| 欧美zozozo另类| 免费看不卡的av| 国产一区亚洲一区在线观看| 中文字幕免费在线视频6| 久久99精品国语久久久| 国产 精品1| 精品国产一区二区三区久久久樱花 | 午夜精品一区二区三区免费看| 日韩制服骚丝袜av| 晚上一个人看的免费电影| 久久99精品国语久久久| 人妻少妇偷人精品九色| 九九在线视频观看精品| 国产免费视频播放在线视频| 亚洲性久久影院| 日韩av免费高清视频| 国产高清三级在线| 草草在线视频免费看| 欧美 日韩 精品 国产| 日产精品乱码卡一卡2卡三| 亚洲精品日本国产第一区| 你懂的网址亚洲精品在线观看| 日日摸夜夜添夜夜爱| 91狼人影院| 亚洲国产欧美人成| 午夜精品一区二区三区免费看| 亚洲av男天堂| 高清毛片免费看| 3wmmmm亚洲av在线观看| 成人鲁丝片一二三区免费| 日产精品乱码卡一卡2卡三| 国产黄色视频一区二区在线观看| 我的老师免费观看完整版| 午夜福利视频1000在线观看| 免费av不卡在线播放| 久久韩国三级中文字幕| www.av在线官网国产| 少妇裸体淫交视频免费看高清| 另类亚洲欧美激情| 欧美少妇被猛烈插入视频| 一级毛片aaaaaa免费看小| 国产精品一及| 别揉我奶头 嗯啊视频| 最新中文字幕久久久久| 亚洲精品中文字幕在线视频 | 噜噜噜噜噜久久久久久91| 成年女人在线观看亚洲视频 | 国产日韩欧美亚洲二区| 九九久久精品国产亚洲av麻豆| 久久久久久久国产电影| 亚洲欧美日韩无卡精品| 久久精品久久久久久久性| 中文字幕av成人在线电影| 大香蕉97超碰在线| 你懂的网址亚洲精品在线观看| 能在线免费看毛片的网站| 黄色日韩在线| av国产久精品久网站免费入址| 亚洲av二区三区四区| 亚洲精品影视一区二区三区av| 成年版毛片免费区| av卡一久久| 三级男女做爰猛烈吃奶摸视频| 好男人视频免费观看在线| 久久女婷五月综合色啪小说 | 又爽又黄无遮挡网站| 日本午夜av视频| 在线亚洲精品国产二区图片欧美 | 国产黄频视频在线观看| 国产精品一二三区在线看| 国产精品福利在线免费观看| 美女高潮的动态| 亚洲精品,欧美精品| 成人鲁丝片一二三区免费| 日韩欧美精品v在线| 嘟嘟电影网在线观看| 日韩中字成人| 亚洲不卡免费看| 国产成人免费观看mmmm| 又黄又爽又刺激的免费视频.| 亚洲国产欧美在线一区| 青春草视频在线免费观看| 少妇的逼好多水| 小蜜桃在线观看免费完整版高清| 亚洲欧洲日产国产| 免费看不卡的av| 深爱激情五月婷婷| 看免费成人av毛片| 六月丁香七月| 最近中文字幕2019免费版| 美女cb高潮喷水在线观看| 夜夜看夜夜爽夜夜摸| 中文乱码字字幕精品一区二区三区| 国产白丝娇喘喷水9色精品| 天美传媒精品一区二区| 欧美变态另类bdsm刘玥| 激情五月婷婷亚洲| 精品久久久噜噜| 亚洲欧美成人综合另类久久久| 亚洲精品乱码久久久v下载方式| 自拍欧美九色日韩亚洲蝌蚪91 |