This paper proposed an automatic hypernasality grades classification algorithm in cleft palate speech based on discrete wavelet decomposition coefficients and cepstrum analysis.Currently,the widely used features to classify hypernasality grades include MFCC,Teager energy,Shannon energy and so on.However,the classification accuracy is low,and the computation amount is large.The speech data tested in this work include 1789 Mandarin syllables with the final \a\,which are spoken by cleft palate patients with four grades of hypernasality.The wavelet decomposition coefficientcepstrum was extracted as the acoustic feature,and then KNN classifier was applied to identify four grades of hyperna-sality.The classification performance was compared with five acoustic features:MFCC,LPCC,pitch period,formant and short-time energy.Meanwhile,the performance of KNN was compared with SVM classifier.The experimental results indicate that the recognition accuracy obtained by using wavelet decomposition coefficient cepstrum feature is higher than that obtained by using five classical acoustics features.The classification accuracy is higher when using KNN than SVM classifier.Recognition accuracy obtained by using wavelet decomposition coefficient cepstrum feature combined with KNN is 91.67%,and 87.60% combined with SVM.Recognition accuracy using classical acoustics features combined with KNN is only 21.69%~84.54%,and 30.61%~78.24% combined with SVM.

Due to the complexity of the urine sediment space environment,there is much redundant information of the collected tangible component image,and it also becomes difficult to extract effective image information.Therefore,the amount of data that need to be deal with is huge.Although the serial version DJ8000 system platform of BP neural network algorithm solves the problem of recognition accuracy of tangible components such as cells,it can’t meet the real-time requirement of urine sediment image medical examination.To solve this problem,this paper presented a system platform of parallel processing GPU framework based on BP neural network algorithm optimization.It uses parallel optimization framework to synchronize and accelerate processing of data efficiently.At the same time,it supports the hardware platform based on GPU computing and test platform.Whether from the hardware indicators,data transmission and bus technology or hardware and software compatibility,it will help solve the problems,which often occur in the uneven load irregularities.Experimental data show that BP neural network algorithm for urinary sediment identification improve the performance parameters such as speedup,aging ratio and running time on GPU platform processing platform.Compared with DJ8000 system platform,the parallel processing GPU framework system platforms of AMD HD7970 and NVIDAGTX680 are optimized,and their corresponding acceleration ratio parameter values are 10.82~21.35 and 7.63~15.28 standard equivalents respectively.The data show that optimizing the mapping relationship between logical data,address data and linear seeking function in the GPU processing system of parallel frame can dynamically allocate and optimize the algorithm structure and optimize the mapping between software and hardware system.Finally,it solves the problem of performance bottlenecks caused by load imbalance between threads.Thus,it effectively resolves the problem of real-time detection in urinary sediment environment.

In order to obtain the accurate information of atherothrombotic plaques in the cardiovasculars and assist the diagnosis and classification of the plaque tissues,this study applied apply a machine learning method to automatically characterize the atherothrombotic plaques in intravascular ultrasound(IVUS) grayscale images.In this study,207 plaque samples in the IVUS images were collected from 10 patients with cardiovascular disease in the hospital.Firstly,the size of a sliding patch is determined and then its centre pixel traverses in the plaque area.The values of the mean and entropy are calculated.Ten features of the patch along 4 directions are respectively obtained by using co-occurrence matrix method.Secondly,more texture features of the plaque region in the IVUS images are obtained by using Gabor filter and local binary pattern(LBP) methods.Finally,the classifiers of Liblinear,random forests and Harmonic to Minimum-Ge-neralized LVQ(H2M-GLVQ) are used to classify these pixels in the plaque tissues based on the features obtained through reducing dimension by using principal component analysis(PCA).The manual characterization by an experien-ced physician is considered as the gold standard.Results of the proposed automatic characterization method show the general identification rates of classifiers of random forests and H2M-GLVQ are over 80%.Compared with other two classifiers,the identification rate of random forests is relatively higher,i.e.89.04%,80.23% and 73.77% respectively for fibrotic,lipidic and calcified plaque tissues.

Large cancer genome projects such as The Cancer Genome Atlas(TCGA) and International Cancer Genome Consortium(ICGC) have produced big amount of data collected from patients with different cancer types.The identification of mutated genes causing cancer is a significant challenge.Genovariation in cancer cells can be divided into two types:functional driver mutation and random passenger mutation.Identifcation of driver genes is benefit to understand the pathogenesis and progression of cancer,as well as research cancer drug and targeted therapy,and it is an essential problem in the field of bioinformatics.This paper proposed a driver pathway identification algorithm based on mutated gene networks for cancer(GNDP).In GNDP,a nonoverlap balance metric is defined to measure the possibility of two genes lying in the same driver pathway.To reduce the complexity of the constructed mutually exclusive gene networks,the nonoverlap balance metric,the exclusivity and the coverage of a gene pair are computed first,and then the edges with low nonoverlap balance metric,low exclusivity and low coverage are deleted.Then,all maximal cliques which might be potential driver pathways are found out.After that,the weight of each clique is assigned as the product of its exclusive degree and coverage degree and then every node of a clique will be checked to judge whether is’s deletion might obtain a larger weight.At last,the maximal weight cliques are obtained in mutually exclusive gene networks as the final driver pathways.This paper compared GNDP algorithm with classical algorithm Dendrix and Multi-Dendrix on both simulated data sets and somatic mutation data sets.The results show that GNDP can detect all artificial pathways in simulated data.For Lung adenocarcinoma and Glioblastoma data,GNDP shows higher efficiency and accuracy than the comparison algorithms.In addition,GNDP does not need any prior knowledge and does not need to set the number of genes in driver pathways in advance.

Accurate lung tumor segmentation is critical to the development of radiotherapy and surgical procedures.This paper proposed a new multimodal lung tumor image segmentation method by combining the advantages and disadvantages of PET and CT to solve the weakness of single-mode image segmentation,such as the unsatisfied segmentation accuracy.Firstly,the initial contour is obtained by the pre-segmentation of PET image through using region growing and mathematical morphology.The initial contour can be used to automatically obtain the seed points required for random walk of PET and CT images,at the same time,it can be also used as a constraint in the random walk of CT image to solve the shortcoming that the tumor area is not obvious if the CT image has not been enhanced.For the reason that CT provides essential details on anatomic structures,the anatomic structures of CT can be used to improve the weight of random walk on PET images.Finally,the similarity matrices obtained by random walk on PET and CT image are weighted to obtain an identical result on PET and CT images.Clinical PET-CT image segmentation of lung tumorshows that the proposed method has better performance than other traditional image segmentation methods.

Global features and local features are very important for texture visual perception.This paper investigated the influence of the global features and local features of texture on the eye movement pattern and the relationship between the eye movement pattern and visual similarity selection.Firstly,the texture images were synthesized by separately controlling global textural features and local textural features with primitive,grain,and point configuration (PGPC) texture model,which is a mathematical morphology based texture model.In the experiment,three scenes were utilized.For each scene,three textures (A,B,and S) were included.Secondly,an experiment was conducted on visual similarity selection to acquire eye movement data while the subjects were viewing the visual similarities of texture scenes under the task of “Which texture is more similar to texture S,texture A or texture B?”.Experimental data were obtained with an eye tracker Tobii T60 by conducting two tests on 89 subjects.The collected eye-tracking data were analyzed in terms of fi-xation point variance in each ROI and fixation transfer count between different ROIs.Analysis results indicate that the global features and local features of texture influence the eye-movement pattern,namely,the texture image that is glo-bally similar to the compared texture contains dispersed fixation points,and the texture image that is locally similar to the compared texture contains concentrated fixation points.Besides,the final visual similarity selection is related to the visual search between different ROIs.

In order to make the extracted vein feature have good clustering performance and thus be more conductive to correct identification,this paper proposed a recognition algorithm based on supervised Nonnegative Matrix Factorization (NMF).Firstly,vein image is divided into blocks,and the original vein feature can be acquired by fusing all sub image features.Secondly,the sparsity and clustering property of feature vectors areregarded as two regularization terms,and the original NMF model is improved.Then,gradient descent method is used to solve the improved NMF model,and feature optimization and dimension reduction can be achieved.Finally,by using nearest neighbor algorithm to match new vein features,the recognition results can be acquired.Experiment results show that the obtained false accept rate (FAR) and false reject rate (FRR) of the proposed recognition algorithm can be reached 0.02 and 0.03 respectively for three vein databases,in addition,the recognition time of 2.89 seconds can meet real-time requirement.

The existing deep learning framework used in diagnosing lung cancer still mainly focuses on lung Computed Tomography(CT) images,but it cannot obtain more higher diagnostic rate,when using only one images in the process of daily diagnosis.Therefore,in this paper,a new pulmonary nodule diagnosis method using dual-modal combined with CT and Positron Emission Tomography(PET) deep denoising autoencoder based on extreme learning machine (SDAE-ELM) was proposed to improve the diagnostic performance effectively.First of all,the method gets discriminative features information separate from the input data CT and PET.Secondly,it inputs CT and PET about candidate lung respectively in whole network.Thirdly,it extracts the high level discriminative features of nodules by alternating stack denoising autoencoder layers.Finally,it makes the fusion strategy of multi-feature fusion as the output of the whole framework.The experiment results show that classification accuracy of the proposed method can reach 92.81%,sensitivities up to 91.75% and specificity up to 1.58%.Meanwhile,the method achieves better discriminative results and is highly suited to be used for pulmonary nodule diagnosis.

Existing lung segmentation methods cannot fully segment all lung parenchyma images and have slow proces-sing speed.The position of the lung was used to obtain lung ROI sequences,and an algorithm of superpixel sequences segmentation was then proposed to segment the ROI image sequences.In addition,improved self-generating neural networks were utilized for superpixel clustering and the grey and geometric features were extracted to identify and segment lung image sequences.The experimental results show that our method’s average processing time is 0.61 second for a single slice and it can achieve average volume pixel overlap ratio of 92.09±1.52%.Compared with the existing me-thods,our method has higher segmentation precision and accuracy with less time.

Automatic recognition of the glands is critical in the histopathology diagnosis of breast cancer,as glandular density is an important factor in breast cancer grading.The gland is composed of a central lumen filled with cytoplasm and a ring of nuclei around the lumen.The spatial proximity of the lumen,cytoplasm,and nucleus may mean that it is a gland,but this method can lead to false-positive errors due to the presence of fat,bubbles and other lumen-like objects in the breast tissue section.In order to solve the above problems,this paper presented an automatic recognition method of breast gland based on two-step clustering and random forest.First,the images to be segmented are constructed by clustering and two-step clustering.A series of morphological operations are performed on the images to repair the objects.Then the segmentation is performed.After that,the method builds the candidate glands,and utilizes the spatial position relationship between central lumen and the nucleus around the lumen and some other features to describe glands.By using random forest classification algorithm,the experimental results show that more than 86% accuracy can be achieved.The result lays the foundation for breast cancer automatic grading.

Volume effect and artifact are important influence factors in MR image processing and single-modal methods can be easily affected.This paper proposed an improved method based on multi-modal local steering kernel to detect the multiple sclerosis in the brain.This method utilizes multi-modal brain MR images and the approximate symmetry of the brain for change detection of the brain.Local steering kernel can measure the similarity between pixels and their surroundings.The proposed method takes the local steering kernel as the feature and measures the dissimilarity by cosine similarity.The experimental results show that the introduction of multi-modal reduces the volume effect and artifact in the MRI,improving the detection effect.

Concerning the problems that many image segmentation algorithms based on fuzzy C mean (FCM) are sensitive to noise and contour segmentation is not clear,an improved algorithm based on wavelet image fusion and FCM clustering algorithm was proposed.And it is applied to MR medical image segmentation successfully.In the first stage of the image segmentation system,the Haar wavelet multi-resolution characteristics were used to maintain spatial information between pixels.In the second stage,wavelet image fusion algorithm was adopted to fuse the obtained multi-resolution image and original image,thus to enhance the clarity of processed images and to reduce noise.In the third stage,FCM technology was used for image segmentation.Experiments on BrainWeb datasets show that compared with the current algorithms,the proposed algorithm has higher segmentation accuracy and robustness to noise,and the processing time is not obviously increased.

Image registration is a key technology in medical image processing.This paper proposed a new multi-resolution medical image registration method based on self-adaptive difference algorithm (DE) and Powell algorithm.It can not only overcome the shortcomings of Powell algorithm depending on the initial,but also can reduce the possibility of getting into local extreme value.Firstly,the source image is processed by multi resolution,and the three layer image including the source image is obtained.Secondly,the adaptive DE algorithm is used to search the global transform parameters on the low resolution images.The transformation parameters are obtained as the initial points of the Powell algorithm.Finally,the Powell algorithm is used for registration in both high resolution images and source images.Compared with traditional experiment,this method has higher precision and can effectively avoid local convergence problem.

Gastrointestinal stromal tumors(GIST) are the most common mesenchymal tumors of the gastrointestinal tract with non-directional differentiation,varying malignancy potential and deficient specificity.Therefore,it is a more concerned issue to diagnosis benign or malignant of GIST.However,it is relatively difficult to use pathological biopsy and CT imaging to study solid tumors heterogeneity.This paper proposed a noninvasive method based on a large number of quantitative radiomics features extracted from CT images and SVM classifier to discriminate benign or malignant of GIST.120 patients with GISTs were enrolled in this retrospective study.Firstly,four non-texture features (shape features) and forty-three texture features were extracted from the tumour region of CT images of each patiant.For the initial feature set,ReliefF and forward selection were executed sequentially to feature selection.Then,SVM classifier was trained by the optimal feature subset for benign or malignant discrimination of GIST.14 texture features were selected for the optimal feature subset from the original feature set.The AUC,accuracy,sensitivity and specificity of the model were 0.9949,0.9277,0.9537 and 0.9018 in the training set,and 0.8524,0.8313,0.8197 and 0.8420 in the test set.The model established by the radiomics method provides a noninvasive detection method for predicting the benign or malignant of GIST,and this mothed maybe as an auxiliary diagnosis tool to improve the accuracy efficiently for malignant and benign discrimination of GIST.

It is difficult to exploit the existing methods to achieve efficient classification and identification of tongue ima-ge’ labels in parallel,and it is also difficult to utilize the correlation between labels for comprehensive analysis.Aiming at the problems above,this paper proposed a classification method of tongue image based on multi-task deep convolutional neural network and constructed a multi-task joint learning model based on deep convolutional neural network to realize the simultaneous identification of tongue color,moss color,tongue crack and tooth marks in tongue diagnosis of Chinese medicine.First,the shared network layer is used to learn all labels,and the correlation between the tags is extracted and utilized automatically from the perspective of feature extraction.Then,the learning tasks of specific labels are completed in different sub-network layers to eliminate the ambiguity in the multi-label classification problem.Finally,multiple Softmax classifiers are trained to achieve parallel prediction of all labels.Experimental results suggest that the proposed method can simultaneous extract multiple features of tongue image and implement classification by means of end to end.The lowest value is about 0.96 in several evaluation indexes and the multi-task recognition rate is about 34ms.Therefore,this algorithm has obvious advantages in accuracy and speed.

Low-frequency and high-frequency ultrasound probes used in intravascular ultrasound (IVUS) image acquisition have their own characteristics.Doctors have to choose ultrasound probes with different frequencies according to clinical needs during the diagnosis of Coronary atherosclerosis and other diseases.Therefore,a Cycle Generative Adversarial Networks (CycleGANs) based on the Wasserstein distance for intravascular ultrasound images enhancement was presented to combine high-frequency ultrasonic details and overcome the problems of edge blur and low resolution of low-frequency ultrasound image,assisting doctors in the diagnosis of cardiovascular disease.Firstly, according to the shape characteristics of coronary artery,several approaches used for data augmentationsuch as rotating,scaling up or down and implementing gamma transformation,are applied to increase the number of IVUS samples in training set,in order to reduce the risk of over-fitting during the training stage.Then,in the spirit of adversarial training,a joint loss function based on adversarial loss and cycle-consistent loss is constructed.Finally,the Wasserstein distance is added to the loss function as a regular term to stabilize the training and accelerated the convergence process.The input of this model is a low-frequency IVUS image and the output is an enhanced IVUS image containing high frequency detail information.An international standard IVUS image database was used for verification in the experiment.Clarity,contrast and edge energy were used as evaluation criteria to quantify.It is verified that the convergence speed of this model is twice of the original CycleGANs model.Three evaluation criteria are increased by 15.8%,11.4% and 46.6%,respectively.The experimental results show that the W-CycleGANs model can learn the feature information of the image domain effectively.Based on the original CycleGANs algorithm,it can further enrich the details of image edges and enhance the diagnostic information,also improve the sensitivity of doctors to diagnosis cardiovascular disease.In addition,100 pieces of clinical IVUS images are used for verification and well enhancement results are gotten.

Multi-contrast high-resolution magnetic resonance imaging(MRI) technology can non-invasively display the wall structure and plaque composition,providing an effective method for diagnosis and analysis of carotid atherosclerotic plaque.The registration of vessels in multi-contrast images becomes a critical task for plaque identification.This paper proposed a three-dimensional registration algorithm based on spatial position alignment and lumen contour matching.With multi-contrast carotid MRI,a coarse-to-fine strategy was adopted.Firstly,the physical coordinates are found to perform the spatial alignment.Then,the ostu algorithm and active contour model are used to complete the semi-automatic continuous segmentation of the blood vessel lumens.Finally,the lumen contour point clouds are utilized to perform three-dimensional rigid registration based on an improved iterative closest point algorithm.The results indicate that the three-dimensional average lumen inclusion rate between TOF and T₁Gd sequence reaches 92.79%,and the average lumen inclusion rate between T₁WI and T₁Gd sequence reaches 94.66%.The proposed algorithm achieves three-dimensional accurate registration of multi-contrast MRI,which lays the foundation for the subsequent analysis of vulnerable atherosclerotic plaque.

To solve the classification problems of melanoma,such as low contrast,indistinguishable by the naked eyes,mass information interference,small dataset and data imbalance,this paper proposed an integrated classification method based on mask data augment and deep convolutional residual network.Firstly,according to the characteristics of skin lesion image and the previous researches,two data augmentation methods by masking the partial area of the trainingima-ges were proposed.Secondly,on the basis of these two data augmentation methods,some features were extracted by using deep convolutional residual 50-layer network.Thirdly,two different classification models were constructed and integrated based on these features.Finally,a series of experiments were conducted based on the datasets of Internal Skin Imaging Collaboration (ISIC) 2016 Challenge competition.The experimental results show that the integrated classification structure model can overcome the deficiencies of a single convolution residual network in melanoma classification tasks,and can achieve better classification results than other methods on skin lesion dataset with less training examples,and multiple evaluation indicators in the proposed method are better than the top-5 results in the ISIC2016 Challenge competition.

In view of the limitation of most traditional multi-modal medical image fusion methods that cannot deal with the medical volumetric images,this paper presented a multi-modal medical volumetric image fusion method based on 3-D shearlet transform (3DST) and generalized gaussian model.Firstly,the preregistered medical volumetric images are decomposed into low frequency parts and high frequency parts by using the 3DST.Next,a novel fusion rule with the local energy is performed on the low frequency subbands.Moreover,an effective fusion rule based on Generalized Gaussian Model (GGD) and fuzzy logic is proposed for integrating the high frequency subbands.Finally,the fused image is obtained by the inverse 3DST.Through subjective and objective performance comparison,experiments on medical volumetric images show thatthe proposed method can obtain better fusion results.

Detecting and tracking cancer cells under phase contrast microscopic images plays a critical role for analyzing the life cycle of cancer cells and developing new anti-cancer drugs.Traditional target tracking methods are mostly applied to rigid target tracking or single target tracking,while cancer cells are non-rigid multiple targets with constant fission,and it makes tracking more challenging.Taking bladder cancer cells in the sequence of phase contrast micrographs images as research object,this paper proposed a multi-target tracking method of cancer cells based on convolutional neural network.Firstly,through making use of detection-based multi-target method,the proposed algorithm adopted the deep learning detection framework Faster R-CNN to detect the bladder cancer cells and preliminarily obtain the cancer cells to be tracked.Then CSA (circle scanning algorithm) was utilized to further optimize the detection of adhesion cancer cells,and further improve the detection accuracy of cells in adhesion area.Finally,it integrated the features of convolution,size and position into a synthetic feature descriptor by using weighting methods,thus tracking multiple cancer cells by achieving the efficient correlation and matching of different frames of cancer cells.The results of a series of experiments show that this method can not only improve the accuracy of detecting and tracking cancer cell,but also deal with the occlusion problem effectively.

The sex determination of skulls is one of the hot research topics in forensic anthropology.It has important research value in the field of criminal investigation and archaeological anthropology.Skull sex recognition is determined by anthropologists through empirical observation of morphology or measurement and analysis of characteristics with gender dimorphism differences.Which is with strong subjective.This aper proposed an automatic gender recognition me-thod for three-dimensional digital cranium.Firstly,a statistical shape model for skulls is built.Then,the feature of high-dimensional skull is projected into low-dimensional shape space.Finally,Fisher discriminant analysis is used to classify the skull in low-dimensional shape space.This method combines the advantages of the measurement and morphological methods.It is easy to operate with no professional and tedious manual measurements.In the experiment,267 Uygur cranium models were selected,including 114 male and 153 female.Of these,76 male and 102 female skulls were used to establish gender discrimination models,and the rest were used to verify.The results show that the accuracy rate in Uygur males and females was 94.7% and 92.1%,respectively.Leave-one-out cross validation shows that this method has high accuracy.

The evaluation method of renal function in clinical renal dynamic imaging depends too much on manual acquisition of ROI (Region of Interest)and has low time efficiency.In order to solve this problem,this paper proposed anautomatic quantitative assessment method for medical renal dynamic imaging.Firstly,the images of renal dynamic imaging at different stages are pretreated.Secondly,an improved level set model is utilized to obtain the ROI of the renal function imaging.The ROI is obtained by morphological methods,then the ROI of the aorta in the renal perfusion imaging is located and obtained.Finally,GFR(Glomerular Filtration Rate) is calculated according to the Gates method,and the time-radioactivity curve is plotted based on the radioactivity counts in ROI,so as to achieve integrated and automated assessment for renal function.The results of clinical trials show that the proposed automatic assessment method can improve the automation level in a short period of time and raise the assessment accuracy,which provide effective help for clinical diagnosis and adjuvant treatment

In view that the existing DCA (dendritic cell algorithm) relies heavily on domain knowledge and artificial experience defining antigen signals in fault detection application,and a single antigen anomaly evaluation method can’t reflect the overall health condition of system,this paper proposed a fault detection method based on immune homeostasis mechanism－IHDC-FD.First of all,in order to solve problem that the danger signal definition is not explicit in actual application,by introducing body’s immune homeostasis mechanism,the change that breaks the homeostasis is consi-dered to be the danger source of system.Therefore,the method of antigen signal of DC adaptive extraction from the change of system state by numerical differential method is proposed.Secondly,the concentration of specific cells within the tissue is the critical factor that can reflect the health of body,and in order to keep healthy,the body’s immune homeostasis has to be maintained.So,by reference to the activation and suppression mechanism of body’s immune homeostasis,the Th and Ts cell concentration which maintain the immune homeostasis is regarded as the evaluation indicators of system imbalance,and once the system lose balance,a fault occurs.Finally,the performance of our method is tested by using step,random and slow drift faults on TE benchmark.Compared with the original DCA,the results show that IHDC-FD not only improves the adaptability of DCA,but also increases the average of fault detection rate by 9.93%,decreases false alarm rate by 230.4% and decreases delay time by 101.2% on the three types of faults testing.Therefore,the IHDC-FD method based on immune homeostasis mechanism has a large improvement than the original DCA on detection performance and adaptability,and it is effective and generality

In order to imrove the accuracy of iris center localization of eyes in gaze estimation,this paper proposed a novel iris center localization method based on 3D eyeball model and Snakuscule.Firstly,a facial alignment method is employed to get the facial feature points and the obtained points are used to obtain the initial iris center.Then,the eye status is judged to reduce the error brought by the low-quality image.To further obtain the accurate iris center location,the energy model of Snakuscule is improved and the iris contour is updated iteratively by a fixed size of Snakuscule.The energy value is obtained by combining the Snakuscule model and 3D eye model.The 3D eyeball model reflects the geometric relationship between the iris center,the eyeball center and the iris contours.According to the energy value,the final iris center can be obtained by updating the iris contours iteratively.Finally,experiments conducted on BioID face database validate the effectiveness and superiority of the proposed method.The maximum standard errors of the algorithm reach at 85.0%,97.8% and 99.8% respectively when e≤0.05,e≤0.1,and e≤0.25.

In order to realize the accurate identification of Web fingerprints in the Web fingerprint database,it is necessary to study the Web fingerprint identification algorithm.When the current fingerprint recognition algorithm is used to identify the Web fingerprint in the Web fingerprint database,there is an error between the recognition result and the actual result,and the recognition takes a long time,which result in low recognition accuracy and recognition efficiency.Based on the cosine measure,a Web fingerprint identification algorithm was proposed.The source fingerprint method is used to select the Web fingerprint in the four aspects of structural features,static files,cookie design and keywords,and a Web fingerprint database is established.Firstly,the characteristics of the data in the Web fingerprint database are extracted,and the abnormal data existing in the Web fingerprint database are removed according to the feature extraction result.Then,the cosine distance function is used as the similarity measurement function,and the K-means algorithm is used to cluster the Web fingerprints in the Web fingerprint database.Finally,the identification of the web fingerprint is completed according to the clustering result.The experimental results show that the proposed method can accurately complete the Web fingerprint identification in the Web fingerprint database in a short time,and has the advantages of high recognition accuracy and high recognition efficiency.

At present,cardiovascular diseases have become the main cause of global non-communicable death,death toll accounts for about one third of the total toll of death in the world,and the number of patients is increasing year by year.Wearable devices is used to automaticaly classify electrocardiogram to facilitate the early monitoring and prevention of cardiovascular diseases for patients.With the rise of edge machine lear-ning and federated learning ,small machine learning models have become a hot issue.According to the characteristics of wearable electrocardiogram equipment such as low configuration,low power consumption and personalization,this paper studied a lightweight network model based on LSTM,and used adaptive algorithm to optimize the ECG classification model of individual patients.The experiment is conducted by using the MIT-BIH open dataset.And compared with the current studies on the detection performance of VEB and SVEB,the experiment results show that the proposed algorithm has simple model structure and high classification performance,which can meet the requirement of ECG monitoring for patients by wearable devices.

In order to realize the automatic classification and detection of palate pharyngeal fricative and normal speech, an automatic pharyngeal fricative detection algorithm based on multi-delay fourth-order cumulant one-third octave spectral line (FTSL) was proposed by studying the pronunciation characteristics of cleft palate patients with pharyngeal fricative.Currently,most researches involved with the detection of pharyngeal fricatives are based on the length of consonants and the energy distribution of speech in frequency-domain.There exist few researches which have achieved automatic classification of pharyngeal fricatives and normal speech.This experiment is based on the pronunciation characteristics of pharyngeal fricative.Each frame’s multi-delay fourth-ordercumulant is computed,and then one-third octave is used to extract the FTSL.Automatic classification of pharyngeal fricative and normal speech is realized by FTSL.In this experiment,the FTSL of 200 normal consonants and 194 consonants of pharyngeal fricative are extracted,and the SVM classifier is used to classify.Besides,comparative experiments were conducted on FTSL feature and traditional acoustic features,and the results were fully analyzed and discussed in this paper.The experimental results show that the proposed FTSL has an accurate rate of 92.7% for the automatic classification of pharyngeal speeches,and it has excellent performance and can provide an effective,objective and non-invasive auxiliary basis for clinical pharyngeal state assessment.

Nasal leak is a typical symptom of patients with velopharyngeal insufficiency.This paper studied the characteristics of nasal leak in cleft palate speech.Recursive plot based on the nonlinear dynamics method is used to explore the features.Combined with the recursive trend analysis method and the region distribution processing based on the recursive plot, quantitative parameters and minimum regions of the recursive plot analysis are extracted as characteristic matrix.Combined with classifier,automatic detection of nasal leak in cleft palate speech is achieved.The experiment analyzes the detection effect for factors such as downsampling point,delay time,critical distance,speech unit and classifier type then comprehensively weighs the influence of each factor on the detection accuracy in order to select the optimal value.The experimental results show that when the KNN classifier is used,the downsampling point is 30000 points,the delay time is 3ms,the critical distance is 5 units,and the speech unit is 4 frames,the detection accuracy of nasal leak in cleft palate speech is 84.63%.The automatic detection algorithm of nasal leak in cleft palate speech is aimed at providing an effective and objective auxiliary diagnosis basis for clinical pharyngeal function assessment.

Analyzing the components of urine in the microscopic image such as red and white blood cells can help doctors evaluate patients with kidney and urinary diseases.According to the characteristics such as low contrast,fuzzy edge of red and white cells in the non-stained and unlabeled urine image,a recognition method based on improved BP neural network was proposed in this paper.Firstly,the method combines genetic algorithm with BP neural network to optimize the weights and thresholds,to solve the problems of local extremum in the training process and improve the recognition accuracy of the BP neural network.Secondly,it uses the method of momentum gradient descent to eliminate the oscillation of network in gradient descent,to accelerate the convergence of the network and improve the learning rate of BP neural network.Compared with basic BP neural network,the improved algorithm improves the recognition rate of red and white blood cells by 6.9% and 9.5%,and the recognition speed has increased by 19.3s and 42.1s.Compared with the CNN recognition algorithm,the improved algorithm improves the recognition rate of white blood cells by 1.7%.Compared with the SVM recognition algorithm,the improved algorithm improves the recognition rate of red and white blood cells by 12.9% and 12.7%.The results of verification test and control test show that the improved method can realize the recognition of red and white cells with higher accuracy and faster recognition speed.

In order to improve the accuracy of eye movement directions identification based on electro-oculogram (EOG) signals,this paper utilized the electrooculogram (EEG) signals containing EOG artifacts and proposed a new approach to classify eye movement directions.Firstly,EEG signals from the 8 channels in the frontal lobe of the human brain are collected,and EEG data pre-processing is made ,including data normalization and least squares based denoising.Then support vector machine based methodis applied to perform multiple binary-classification,and finally voting strategy is used to solve four-classification problems,thus achieving eye movement directions identification.The experiment results show when using the approach of this paper to classify eye movement directions,the classification accuracy rates in the upper,lower,left and right directions are 78.47%,72.22%,84.03%,79.86% respectively,and the average classification accuracy rates reach 78.65%.In addition,compared with the existed classification methods,the classification accuracy rate of this paper is higher,and the classification algorithm is simpler.It is validated the feasibility and effectiveness of using EEG signals to identify eye movement directions.

The segmentation of the retinal vessels in fundus image is important for the diagnosis of ophthalmic diseases such as diabetes,retinopathy and glaucoma.Aiming at the difficulties of extracting blood vessels from retinal blood vessel images and the lack of data samples,a retinal vessel segmentation method combining attention module with encoder-decoder structure is proposed.To improve the segmentation effect of retinal blood vessels,a spatial and channel attention module is added to each convolutional layer of the encoder-decoder convolutional neural network to enhance the utilization of the spatial and channel information of the image features (such as the size,shape,and connectivity of the blood vessels),where the spatial attention focuses on the topological characteristics of blood vessels,and the channel attention focuses on the correct classification of blood vessel pixels.Moreover,the Dice loss function is used to solve the imbalance of positive and negative samples in retinal blood vessel images.The proposed method has been applied on three public fundus image databases DRIVE,STARE and CHASE_DB1.The experimental data show that the accuracy,sensitivity,specificity and AUC values are superior to the existing retinal vessel segmentation me-thods,with AUC values of 0.9889,0.9812 and 0.9831,respectively.The experimental results show that the proposed method can effectively extract the vascular network in healthy retinal images and diseased retinal images,and can segment small blood vessels well.

For the growing demand of ECG data analysis,a new ECG classification algorithm is proposed.Firstly,the original data are truncated by fixed length,sample equilibrium is obtained,and the pre-processing operations such as instantaneous frequency and spectral entropy of the signal are obtained.After the data is preprocessed,the model can better extract features from the data for learning.In training progress,a two-way LSTM network stacking model is adopted.The stacked two-way LSTM model is an improved cyclic neural network model.Compared with convolutional neural networks,the cyclic neural network is more sui-table for processing sequence data such as electrocardiogram.The experiment is conducted using MATLAB2018b under Windows for training and testing.The CUDA version is 9.0.The classification accuracy rate is used as an indicator to measure the performance of the model.The model is tested on two data sets,one is the data of the 2017 Physiological Signal Challenge(hereinafter referred to as the 2017 dataset),the final classification accuracy rate is 97.4%;the other is the data of the 2018 Physiological Signal Challenge (hereinafter referred to as the 2018 dataset),and the final classification accuracy rate is 77.6% on this dataset.The MATLAB group to which it belongs has achieved the third place.This algorithm improves the accuracy of 5.6% in the 2017 dataset and 7.6% in the 2018 dataset compared to the results of the traditional LSTM network.Compared to the results of a single-layer bidirectional LSTM network,in the 2017 data set,the accuracy rate improves 4.2%,and the accuracy rate improves 5.7％ in the 2018 data set,which fully verifies the feasibility and advantages of the algorithm.