Abstract:Mammography is one of the most widely used methods for breast cancer detection. The technique and theory of image processing and pattern recognition can be used for mammogram analysis. The analytical results can assist radiologist in finding missed tumors and identifying false positive tissues, leading to low false negative rate and false positive rate. The method using image processing should simulate the mammogram interpretation of the radiologist. Thus, the breast cancer detection and classification method based on multi-view is rather suitable in clinical practice. Determination of the correspondence between multi-view mammograms is the foundation for multi-view detection. Here, an overview of recent developments in determining correspondence between multi-view mammograms is presented. Nipple detection and pectoral muscle segmentation are first summarized. Both advantages and disadvantages of different methods are compared. Then two-view matching and bilateral matching are discussed. Finally, the problems in the existing matching methods are analyzed and improvements are suggested.

Abstract:Based on functional magnetic resonance imaging (fMRI) technique, exploring the brain plasticity has a great role in the decoding of the human brain activity and simulating the brain intelligence, and it is also a challenging task. However, the brain functional plasticity could be reflected by the variability of the functional connectivity, which closely depends on the effective analysis models. Thus, firstly,the key brain functional connectivity analysis methods are reviewed and the corresponding limitations of each method are also analyzed. Then, on the evidence from the researches of brain functional connectivity, the complex relationships between the brain functional plasticity and occupational factors are summarized and analyzed. Finally, the future research directions of the brain functional connectivity analysis models and the brain functional plasticity are recommended and discussed.

Abstract:Computer-aided diagnosis (CAD) system can detect, segment and diagnose pulmonary nodules from CT images, and improve the survival rate of early lung cancer, which has important clinical significance. As the appearance of pulmonary nodules varies with its type, size, location, internal structure, and malignancy, nodule detection and diagnosis have become a major challenge . Here the key techniques and challenges are analyzed in four main processing stages: segmentation of lungs from chest images, detection of pulmonary nodules inside the lung fields, pulmonary nodule segmentation, and diagnosis of pulmonary nodules as benign or malignant. Further research is needed to optimize the diagnosis algorithm sensitivity of nodules with different sizes and shapes, thus decreasing the number of false positives, and improving automation level of diagnosis. Finally, the picture archiving and communication systems (PACS) should be integrated with electronic medical record systems (EMRS) in order to be adopted in clinical practice. 

Abstract:Image hashing is a technique to map a digital image into a content-based and short binary code. It has the properties of robustness, security, compactness and one-wayness, which has been widely applied in the field of image authentication and identification. Here, a robust image hashing algorithm based on block compressed sensing is proposed, using the characteristics of secure computation and linear operation in the sampling stage. In the proposed algorithm, the input image is partitioned into sub-blocks. For each sub-block, random projection is applied to it based on the theory of compressed sensing, and a measurement vector can be obtained under the control of the secret key. Then each measurement vector is quantized as one bit and finally a binary hash value can be obtained whose length can be adjusted by the strategy of image partition. Experimental results show that the proposed algorithm has satisfied performance in robustness, security and speed.

Abstract:Under the background of global aging and empty nest family, the tumble of seniors has attracted a great deal of attention. To provide help for seniors and relieve the injury of tumble, a tumble recognition algorithm based on image processing and multi-features fusion is proposed. In view of the prospect of extraction, we propose an algorithm that combines three-frame difference method and background subtraction division of target by weight, then extract the height, ratio of width to height, the center of mass, the perimeter of a rectangle, Hu moments′ and Zernike moments of target contour, using five experimenters′ walking, siting down, squating down and tumbling as the experimental samples. The algorithm realizes tumble detection and recognition by training and predicting support vector machine (SVM) after parameter optimization. The experimental results show that the proposed algorithm is efficient and fast with easy implementation. The average recognition rate is more than 95%.

Abstract:Modeling and collision detection algorithm is the premise of the real-time in virtual surgery. Here, we extract the CT point cloud data from patients. Then, we model for soft tissues and organs based on the algorithm of octree subdivision and hierarchical structure of bounding sphere. In order to improve the real-time of the collision detection, the physical models of the surgical instruments are simplified to balls or lines. The geometry remains unchanged, and it thus does not affect the operation of the virtual visual while the speed of the collision detection is improved greatly. The experimental results show that the algorithm can accurately detect the point of virtual contact with the virtual mode l of surgical instruments, and collision detection in real-time have increased considerably. Simplified average collision detection time is 10% of the original.

Abstract:Due to incompletely extracted coronary vessels and discontinuous skeleton caused by image blurring and low contrast of coronary angiogram, a multi-scale coronary segmentation and skeleton extraction method is proposed based on the Hessian matrix. Vessel radius to be estimated also lays the foundation of three-dimensional reconstruction of coronary structure. By using the relations between Hessian matrix eigenvalues and line items, a novel vesselness measure is constructed and obtain a segmented result enhancing and thresholding coronary are segmented. The normal direction of the coronary vessels is determined by the Hessian matrix, and an initial set of coronary skeleton pixels is obtained by extreme points through solving the normal direction. Therefore, Euclidean skeleton of coronary vessels are extracted. The experiment result indicates that the algorithm is concise and effective. Also it can extract more tiny branches comparing with the existing algorithms. It can get a completed coronary skeleton and estimate radius accurately.

Abstract:Synthetic heart sound signals have certain application value in the teaching and scientific research. Here, a composite heart sound generator is designed. Firstly, the generation mechanism and chaos characteristics of heart sounds are analyzed, and the principle of composite heart sound is composed. Then based on chaotic characteristics of heart sound, a heart sound generator is constructed, which includes heart sound generating sub model on the left and right. With that, a set of amplitude and period adjustable synthetic heart sound signal is obtained by processing the output waveform. After the analysis of time-frequency and chaos characteristics of synthetic heart sound signal, it reveals that synthetic heart sound signal is highly similar to actual heart sound, and it can basically meet the need of teaching and scientific research.

Abstract:A nonlinear predictive control algorithm based on wavelet neural network (WNN) integrating optimal experimental design with manifold regularization is presented for the complex processes. Firstly, the wavelet hidden nodes are recursively selected from candidate node set to be added into WNN and the optimal parameters of selected nodes are obtained through extended Kalman filter (EKF). The optimum experimental design and Laplacian regularization are then integrated to select salient WNN hidden nodes, and minimum description length (MDL) is utilized to determine the number of hidden nodes. Initial WNN parameters and associated weight updating scheme are provided via an online Gustafson-kesscl(GK) based fuzzy satisfactory clustering algorithm with intuitive interpretation and physic meaning. Finally, a predictive functional control law is given by linearizing WNN. The simulation of industrial coking equipment shows the efficiency of the proposed algorithm.

Abstract:Aiming at upgrading the performance of face pose classification, we proposed an algorithm of face pose classification based on deep learning and gradient information fusion. First, the pixel gray intensity features and the features of gray intensity difference nearby each pixel from a face image are extracted. Then, these features of face images are processed with deep learning technique through a dedicated three-layer restricted Boltzmann machines network, which has been trained by a large number of samples. Finally, a corresponding relation between fusion deep learning features and the labels of face pose classifications is built through a Softmax classifier. The experiment results show that the proposed algorithm achieves a state of the art classification accuracy, generally higher than 95%, when learning and testing on CAS-PEAL-R1 face database

Abstract:The brain is the most complex tissue in the structure and function of the organism, which contains hundreds of neurons. As a basic unit of the structure of the brain, the structure and function of neurons contain many factors, among which the geometric feature is an important aspect. The morphology of the neurons in brain is so complicated and diversiform that it is a problem to recognize the category of them. Here, we first establish the fuzzy set model based on fuzzy clustering according to the geometry of neurons. We use the optimal classification model of multi-database classification model to improve the fuzzy clustering method and classify the neurons. Then we can obtain the optimal classification result. According to the evaluation method of clustering, we can verify that the improved fuzzy clustering method can get better clustering effect compared with other methods.

Abstract:A novel action recognition method based on general multiple kernel learning is proposed. Firstly, histogram of oriented gradients (HOG） based on edge of image and scale invariant feature transform（SIFT） based on dense sampling are extracted. Furthermore, spatial pyramid model is considered to obtain coarse spatial information. Then, the kernel matrix of each level in spatial model is computed by histogram intersection kernel function. With general multiple kernel learning, the weights of kernel matrixes are solved and the optimal kernel matrix is achieved by the linear combination of kernel matrixes. Finally, action recognition is realized by the decision function. The obtained impressive result shows that the proposed algorithm is more effective than some common methods in Willow-actions dataset.

Abstract:Differential expression analysis of genes and isoforms is important in obtaining the function of genes and isoforms, thus becoming an essential research focus of bioinformatics. RNA-seq is a new experimental technique based on high-throughput sequencing and is increasingly used in transcriptome research. Read-isoform multi-mappings make it difficult to detect differential expression of isoforms. Here, we proposed a new method, called PG_bayes, to detect differential expression for both genes and isoforms. PG_bayes, based on expressions estimation method PGseq, uses a Bayes factor model selection method to detect differential expression. We applied PG_bayes to three human datasets and one mouse dataset, and compared its performance with popular alternatives. Results show that PG_bayes performs favorably in sensitivity and specificity at both gene and isoform levels.

Abstract:To overcome the target boundary prone to be misclassification for an original image when the user-selected seed pixels become less in the graph cut algorithm. An interactive K-means and graph cut algorithm (KMGC) is proposed in the combination of the K-means with graph cut(GC) algorithm and the interactive segmentation with brain magnetic resonance image (MRI). The MRI intensity inhomogeneity is processed by K-means clustering algorithm. On this basis, the graph cut algorithm will further refine the MRI, so as to obtain effective segmentation of white matter and gray matter. We implement extensive segmentation experiments using both synthetic and real brain MRIs. Quantitative and qualitative analyses are carried out about the experimental results, and the results are compared with other segmentation algorithms. The experimental results show that the KMGC algorithm can effectively divide the brain MRI, and outperform others on the segmentation effect.

Abstract:Epilepsy is caused by abnormal synchronous discharge of neurons in the brain, which constructs the main basis of its diagnosis. The use of complexity theory to study the epileptic signal has become a hot spot. The symbolic transfer entropy as a reflection of the degree of chaos of nonlinear system of indicators can be used as a characteristic of epilepsy. It plays an increasingly important role in the study of epilepsy in electro encephalogram signals (EEG) feature extraction. But symbolic transfer entropy is generally used to measure the dynamic characteristics and directional information between two variables and ignores the interaction between multivariate. Epileptic EEG signals are analyzed based on multivariate symbol transfer entropy. By choosing the lead signal and the signal length to analyze the robustness, the method can be used to distinguish normal person and patients with epilepsy. It is proved that the algorithm is robust and reliable for clinical diagnosis.

Abstract:Image inpainting technique aims at infilling the images with missing or damaged portions in a way that they will be non-detectable for an observer.The Criminisi algorithm in the image inpainting is introduced. An improved algorithm is proposed to tackle the existing problem. These new improvements are in four aspects.Firstly, a new priority function is used to adjust the order of the selected pixel block, and therefore the incorrect filling order caused by rapid decay of data term is avoided. Secondly, a Sobel operator is introduced to improve computing method of isophote intensity,making the inpainting order along the isophotes. Thirdly, a new matching searching method to identify the samples in the neighborhood of the damaged region based on similarity is employed. Finally, to smooth the propagation of error in updating the confidence value, a new formula is defined. Experimental results show the improved algorithm can get a satisfied inpainting result and improve repairing efficiency.

Abstract:Convolutional neural networks are good at learning features, but not always optimal for classification, while extreme learning machines are good at producing decision surfaces from well-behaved feature vector, but cannot learn complicated invariances. Based on the advantages and disadvantages of convolutional neural networks and extreme learning machine, we present a hybrid system where a convolutional neural network is trained to extract features and an extreme learning machine is trained from the features learned by the convolutional neural networks to recognize faces. We also propose prefix part of the filters in the convolutional layers to reduce parameters for improving the recognition accuracy. The experimental results obtained on the ORL and XM2VTS databases show that the proposed method can effectively improve the performance of face recognition, and the method of prefixing part of the filters is better than the method of stochastic filters in small training data.

Abstract:Thyroid nodule is a kind of frequently-occurring disease. Ultrasound technology is the preferred examination method for the disease. Extracting the texture feature distinguishing the benign and malignancy in the ultrasound images and discriminate them has a wide prospect of clinical application. Dual-tree complex wavelet transform (DT-CWT) and Gabor wavelet are the important approaches to texture feature extraction. Here, we present an approach of thyroid nodules recognition by fusing multi-scale DT-CWT and Gabor wavelet features. Firstly, we use Gaussian pyramid to decompose the thyroid ultrasound image into multi-scale space. Followed by extracting DT-CWT and Gabor multi-scale features, the feature fusion is performed. Support vector machine (SVM) is applied to classify so as to verify the effectiveness of the proposed method. Experimental results show that the proposed method can achieve a high recognition rate.

Abstract:Gene splicing as a tightly regulated process,is a pivotal process between transcription and translation during gene expression. Splice sites are the kernel regulatory elements for gene splicing. Here, based on the sequential features minded from splice site sequences, we develop a score system for splice site sequences. Through this score system, splice site sequence can be measured quantitatively. The experimental results show that the canonical and pseudo splice site sequences can be discriminated effectively. Moreover,this model outperforms the maximum information entropy model with a great robustness, and the pathogenic splice site sequence mutations can be detected efficiently by the model

Abstract:A noise-robust fingerprint-factor-based audio feature and a semi-supervised audio dictionary training algorithm are proposed to fill up the deficiency caused by noise in content-based audio retrieval. The proposed method extracts audio fingerprint from Mel spectra and utilizes non-negative matrix factorization to factorize fingerprint into noise-robust spectral factor and temporal factor as features. Also an semi-supervised audio dictionary training algorithm is proposed. It uses an audio effect set to calculate the distribution of basic sound effects as initialized dictionary. The quantization is conducted while the dictionary is dynamically updated at the same time to better characterize data. The experimental results show that under low signal-to-noise ratio (SNR), the proposed method significantly improves the average precision compared with other algorithms.

Abstract:To improve the low accuracy of breast X-ray medical image multi-class classification， a new medical image multi-class classification method based on edge detection is proposed. The breast X-ray medical image is firstly preprocessed, including image denoising and enhancement. Tumor region in X-ray medical image can be acquired through edge detection algorithm. Feature selection on tumor is implemented using gray level co-occurrence matrix. This method uses support vector machine (SVM) to classify the medical image according to selected features. The X-ray medical image without any detected edge can be directly classified into the normal without breast cancer. The experiment result shows that the new medical image multi-class classification method based on edge detection has a higher precision than the traditional SVM multi-class classification algoritm on breast X-ray medical image. 

Abstract:A method based on partial phase synchronization was employed which can distinguish the direct coupling and indirect coupling effectively in a multivariate nonlinear system compared to the bivariate phase synchronization. Applying this method to three coupled stochastic Rossler oscillators, the result reveals that the analysis of partial phase synchronization can infer the correct interactions among oscillators. Using the method to analyze the teenage EEG and middle-aged EEG, the result shows that the coupling degrees of teenage EEG and middle-aged EEG are significantly different. The method of partial phase synchronization can be a reference to distinguish teenage and middle-age EEGs.

Abstract:Face recognition is still challenging due to the large variations of facial appearance, caused by lighting, partial occlusions, head pose, etc. The feature extraction is a key step for face recognition. In order to improve the recognition rate of face recognition,we introduce a novel feature extraction technique for face recognition, which is a combination of compressed sensing and spatial pyramid model method. The scale invariant feature transform is first used to be a feature extractor to obtain facial features.Then by using sparse coding in the randomly generated dictionary, dimensionalities of those features are reduced. After the spatial pyramid is used to be a feature extractor to obtain different spatial scales, the max pool is used to integrate the features. Finally, the kernel sparse representation classifier is proposed to classify the features to complete the face recognition. The experimental results based on the Extended Yale B, AR and CMU PIE databases demonstrate that the method has a strong rustness in the illumination, pose and disguise variation with a faster running speed.

Abstract:An improved method is proposed to deal with the low estimation accuracy and convergence rate in fixed parameters estimation of mixed digital signals using particle filtering. By modification of the traditional random walk model, and modeling the posterior probability of the parameters as the BETA distribution, the estimation accuracy and convergence rate are enhanced, the separation performance is improved either. In order to evaluate the performance of the proposed algorithm, the joint Cramer-Rao bound for parameter estimation is derived under the condition of known transmitted symbols. Simulation results show that the new algorithm has better performance than that of the traditional methods.

Abstract:Most of blind channel identification algorithms cannot estimate the channel with common zeros and they are sensitive to the channel order error. Here, this paper proposes a new cross-relation-based semi-blind channel identification method. The algorithm uses the output data structure correlation matrix Wand builds a linear system of equations based on the orthogonal relationship between matrix W and channel vector. Some known symbols are utilized based on MLS criterion to build other equations. The closed form solution of channel response is derived by the least-square method. The proposed algorithm effectively overcomes many limitations of blind channel identification algorithms, avoids the selection of optimal weighted parameter that commonly appears in the traditional semi-blind methods with strong robustness to channel noise and channel order. Simulation results verify the effectiveness and superiority of the proposed algorithm.