Abstract:The 3D structure of protein determines its special biological function. The 3D structure of protein has important scientific significance for protein function research， disease diagnosis and treatment， and innovative drug research and development. It is an effective method to predict protein 3D structure from amino acid sequence by computer. Fragment assembly is a widely used technique for protein structure prediction， which can effectively reduce the conformational search space by converting continuous conformational space optimization into discrete experimental fragment combination optimization. This paper first introduces the technology of fragment assembly. Next， the development of protein structure prediction based on fragment assembly is summarized， and some typical prediction methods are briefly described. The commonly used databases and evaluation indexes in protein structure prediction are then demonstrated， and the performance of the representative prediction methods is compared. Finally， we analyse and point out the challenges of the current protein structure prediction methods based on fragment assembly， and look forward to the future research directions in this field.

Abstract:The 3D ground reaction force（GRF） is an important measurement parameter in human motion analysis， but its measurement is limited. The application status of machine-learning in predicting GRF is systematically reviewed. Searching with the combination of “ground reaction force”，“machine learning”and“neural network as the keyword， fourteen studies of predicting GRF by using machine learning model are screened. The motion tasks in these studies include walking， running and several special tasks in sports. Different learning algorithms are used to predict the GRF， whose input parameters include plantar pressure parameters， human motion parameters obtained by motion capture. The relative root mean square and cross-correlation coefficient are adopted as evaluation indicators. The results show that using the machine learning to predict GRF can obtain excellent prediction accuracy. The application of predicting GRF by using machine learning models in human motion needs more research， such as increasing the sizes of datasets for machine learning to further improve the prediction performance of learning models.

Abstract:The network is a popular way to model the interactions among elements in nature， and it has been widely used in many studies. The human brain can be considered as a complex network after defining nodes and edges. In such a network， human cognitive functions and some brain diseases are closely related to the structure of the network. Brian network analysis is a popular research area with important applications in a variety of disciplines. It provides a powerful approach to many works， including exploring working mechanisms of the brain， understanding pathological underpinnings of neurological disorders， and improving the efficiency of the therapeutic and diagnostic in clinical. This paper reviews the concepts， methods， and applications of brain network analysis， and it is divided into three parts with a introduction， i.e.， brain network construction， brain network representation， and brain network analysis. Finally， we summarize this paper and discuss some new directions and problems for future research.

Abstract:Liver cancer is one of the common malignant tumors with high clinical mortality. Discovering new targets and developing new drugs are essential for the treatment of the disease. This study analyzes the differentially expressed genes related to the occurrence and development of liver cancer from the perspective of biological function and network， aiming to screen out molecular markers and immunotherapy targets for early diagnosis of liver cancer. First， two sets of liver cancer-related gene expression data are screened for differential expression. Then through GO （Gene ontology） function and KEGG （The kyoto encyclopedia of genes and genomes） pathway analysis， 128 target genes that are significantly enriched in the main functions and signal pathways are obtained. Finally， through the gene regulation network to explore the interaction law of target genes， ten key genes are identified and survival analysis is performed to verify that four genes of CYP3A4， CYP3A5， CYP2C9 and CYP2C8 are suitable as liver cancer markers or targeted therapeutic targets. This study provids a reference for the mechanism study of the occurrence and development of liver cancer， the screening of tumor markers and the selection of drug targets， and provide a basis for further related functional research.

Abstract:Based on the eye movement data of subjects detected by eye tracker， a new eye movement index， the percentage change of eyelid spacing， is put forward. Then it is compared with such traditional eye movement indices as saccade count and pupil diameter. After significance and correlation analysis， the receiver operating characteristic （ROC） curve is used to evaluate the discriminant performance of every eye movement on the lens comfort. Experimental results show that the proposed percentage change of eyelid spacing is the best indicator to evaluate the lens comfort， followed by the pupil diameter， and the saccade index and the average fixation duration are the similar. It validates that the lens comfort can be evaluated accurately and quickly by detecting the percentage change of eyelid spacing of subjects. This research can help relevant practitioners to evaluate the lens comfort from an objective perspective， and also provides some technical and theoretical help for increasingly severe myopia prevention and control situation of China.

Abstract:Rhinitis is a common chronic inflammation of the upper respiratory tract with a variety of symptoms and signs. The clinical classification of rhinitis is characterized by different types of instances and class imbalance， and belongs to multiple output classification. Low recognition rate and poor generalization performance often occur for minority class instances. Therefore， this article proposes a novel classification model based on heterogeneous integrated frame， which translates the multi-output classification of rhinitis to multi-label and multi-class classification， then builds a heterogeneous integrated classifier by ensemble learning algorithm. The proposed model can automatically adjust the number and depth of integrated forest learners according to the imbalance ratio of single class label in a subset. As a result， it can effectively reduce influence of class imbalance and improve classification performance of majority and minority class concurrently， further to enhance generalization of integrated classifiers. We conduct cross-validation classification experiments on 461 cases of clinical rhinitis. The outcomes show that the evaluation indicators of the proposed model， such as sensitivity， specificity， accuracy， F1 and AUC， are 74.9%，86.5%，92.0%，0.783 and 0.953， respectively. In comparison to other baseline methods， it achieves better evaluation performance and is more suitable for the early clinical diagnosis of rhinitis.

Abstract:Blood flow and blood oxygen are important physiological parameters of the organism， which reflect the functional state of the organism. For brain injury with rapid onset and pathological changes， taking cerebral blood flow and blood oxygen as parameters for monitoring head injury is conducive to the diagnosis， treatment and evaluation of head injury. The paper uses a combined measurement system constructed by spectroscopy analysis and laser speckle imaging technology to monitor the blood flow and blood oxygen in the process of lipopolysaccharide induced brain injury in mice， and further analyzes the consistency and difference of the treatment effects of brain injury with hypertonic saline and mannitol. Experimental results show that compared with the blood oxygen and blood flow value of the control group （1.09±0.075， 0.75±0.019）， lipopolysaccharide could increase the cerebral blood flow （1.36±0.080） and decrease the cerebral blood oxygen （0.62±0.021） in mice， which has significant difference. Blood oxygen and blood flow gradually recover after treatment， and the recovery is more obvious after hypertonic saline treatment. It is found that the blood flow and blood oxygen reveal an opposite trend after brain injury， which reflects the imbalance of blood brain maintenance caused by the change of the blood-brain barrier after brain injury. However， the blood flow and blood oxygen recover after the improvement of therapeutic agents， indicating that the blood flow and blood oxygen could be used as two important parameters for therapeutic evaluation. Therefore， the combined monitoring of blood oxygen and blood flow using spectroscopy analysis technology and laser speckle imaging technology provides a technical solution for real-time monitoring of brain injury and therapeutic evaluation.

Abstract:Endoscopy is the main method in the clinical diagnosis and treatment of gastrointestinal diseases. Its function is to obtain the structural information of the tissue. The occurrence and development of digestive diseases are usually accompanied by the changes of functional information， such as blood oxygen and metabolic parameters. This paper develops a fiber optic endoscopy system based on endogenous optical signal imaging technology， which can measure changes of six types of functional information including biological tissue oxy-hemoglobin， deoxy-hemoglobin， cytochrome C， cytochrome oxidase， FAD and scattering characteristics for biological tissues. In this study， we designed an in-vivo animal experiment with injection of excessive sodium nitrite. Then， the designed fiber optic endoscopy system was utilized to obtain the blood oxygen and metabolism related parameters in the same field at 3， 6， 9， 12， 15 and 18 min time points during drug injection. The concentration of oxyhemoglobin in the experimental model continued to decrease， while that of deoxyhemoglobin continued to increase. Research shows that the fiber optic endoscopy system based on intrinsic optical signal can accurately reflect the change of biological tissue function information.

Abstract:To solve the problem of speckle noise in optical coherence tomography （OCT） system， an optical coherence tomography denoising algorithm based on bilateral random projection is proposed. Based on the high similarity of biological tissue structure between adjacent frames of 3D OCT image and the high resolution of image， the original OCT image signal is decomposed into noiseless low rank matrix， sparse matrix and noise matrix. The bilateral random projection algorithm is then used to solve the problem and extract the low rank matrix， so as to remove the noise and restore the noiseless image. The proposed algorithm is tested on the clinical data set， and the noise reduction effect is evaluated by signal to noise ratio （SNR）， contrast to noise ratio （CNR） and equivalent number of looks （ENL）. Experimental results show that compared with the robust principal component analysis algorithm， the proposed algorithm improves the SNR， CNR and ENL by 1.22 dB， 0.84 dB and 59.5， respectively. It can suppress speckle noise more effectively and has lower computational complexity.

Abstract:To improve the treatment precision of focused ultrasound， two traveling wave focusing （TWF） transducers with the same frequency of 0.6 MHz are placed in opposite direction coaxially and confocally to realize standing wave focusing （SWF）. Combining with the numerical simulation of sound field， cavitation and nonlinear research， lesion changes in tissue-mimicking phantom and the formation mechanism of lesion during SWF and TWF ultrasonic irradiation are compared under the same peak focus acoustic pressure about 17 MPa. The results demonstrate that the size of initial lesion formed by SWF ultrasonic irradiation is 0.18λ×0.25λ， which is significant smaller than that of TWF， 0.91λ×0.3λ， under the same peak focus acoustic pressure. But the surface sound pressure of the SWF transducers is only 0.46 times that of TWF. It shows that SWF can significantly compress the lesion size in the transverse direction and can protect organs and tissues by reducing the amplitude of acoustic pressure. And the lesion caused by TWF ultrasound appeared faster than SWF， which is related to the strong nonlinear effect of TWF. The results in this paper show that SWF can achieve more precise lesion than TWF， and the acoustic path is safer， which provides a theoretical reference for the application of SWF in clinic.

Abstract:Aiming at the problem of low recognition accuracy of human lower limb motion state， a lower limb motion state recognition prediction method based on plantar pressure is put forward. The plantar pressure data of 40 groups at different walking speeds collected by EMED plantar pressure plate are used as test samples. By analyzing the characteristic parameters of plantar pressure， the gait phase is constructed， and the gait period relationship and the gait displacement model are established. The lower limb movement is nonlinear， and the plantar pressure prediction is realized by gait cycle model combined with particle filter algorithm. Firstly， the prior probability density function is obtained by particle swarm initialization， and the predicted pressure is estimated. Secondly， the state vector is tested， and the predicted plantar pressure is deduced by multiple linear regression. Experimental results show that the particle filter algorithm has good performance at different gait speeds with the accuracy of more than 97%， and the plantar pressure prediction is effective. The plantar pressure data of participants of different ages， genders and weights are added for analysis and the prediction accuracy is all over 97.5%， verifying the stability and accuracy of the prediction algorithm.

Abstract:With the rapid development of medical imaging technology， medical images have been widely used in clinical detection and scientific research. In view of the insufficient clinical image data set， this paper proposes a generation model based on dense connection self-inverse generative adversarial network （GAN） to realize the mutual generation of T1- and T2-weighted MR images. Especially， the dense block is introduced into the generator module of self-inverse GAN model， and the multi-scale fusion framework of U-net is adopted to realize the mutual generation of T1 and T2 weighted MR images. The BraTS 2018 data set is used for validation and the peak signal-to-noise ratio and structure similarity of the generated images could reach 22.78 and 0.8， respectively. Contrast experimental results of different generators show that the model with the generator based on dense block has better performance than the model with the generator based on U-net or ResNet. The MR image generation method based on dense connection self-inverse GAN proposed in this paper can reduce the negative influence brought from missing T1 or T2 weighted images and provide more information for clinical judgment.

Abstract:In order to improve the recognition accuracy and operation speed of the traditional fall detection system and reduce the false alarm rate and the missing alarm rate， a real-time fall detection algorithm based on fuzzy C-means （FCM） clustering algorithm and convolutional neural network algorithm is proposed. The algorithm takes the depth vision sensor as the data acquisition source， extracts the velocity， the height， the acceleration， and the angle of the cluster center point as the fall recognition feature vector， and uses the combination of threshold analysis and machine algorithm to realize human fall recognition. The experimental results show that the recognition accuracy of the algorithm reaches 99% and the operation speed is 0.178 s， which is higher than those of the traditional algorithm.

Abstract:To solve the problem that deep convolutional neural network （DCNN） models with heavy weights are difficult to be effectively applied on AI edge devices with weak computing power and high storage costs， a flower image classification system equipped with a lightweight DCNN is proposed with the help of a heavyweight DCNN during training process. First， an extended flower data set suitable for lightweight DCNN training is constructed by using a heavyweight DCNN combined with transferring learning， the crawler technology and the maximum connected region segmentation method. Then， two lightweight DCNN models， Tiny-Darknet and Darknet-Reference， oriented for devices with weak computer power are trained based on the specially built flower image gallery. Experimental results show that the two optimized models obtained can achieve 98.07% and 98.83% average classification accuracy respectively on Oxford102 flower dataset while keeping the model size as 4 MB and 28 MB， which have promising application potentials for AI edge computer devices.

Abstract:By constructing a model combining affine-invariant discrete hashing （AIDH） and confidential random field （CRF） ， the object detection of remote sensing image is achieved. Firstly， the remote sensing image is reconstructed by superpixel segmentation， and the undirected graph structure with superpixel block as vertex is constructed for CRF. Then， the superpixel block is used as the test sample for AIDH learning which is used as CRF unary potential function to generate the initial category label. Then， the pairwise potential function of CRF is constructed by using Potts model for label re-learning， while the object neighborhood information is smoothed and the missing area of object detection is resolved. Finally， the convex hull boundary method is used for generating minimum external rectangular frame as object detection result. Experimental results demonstrate that the proposed method achieves the tradeoff of accuracy and efficiency for objection detection tasks.

Abstract:Network alignment is a key way to integrate data from different platforms. Obtaining node representations by using network representation learning and establishing node matching strategies is one of the current mainstream technologies for alignment of heterogeneous networks. In this kind of research， network representation model and computational complexity are two key problems. This paper proposes an unsupervised network alignment method based on multi-scale feature modeling and improved sampling strategy. Firstly， a node feature representation with different scales is proposed to extract node features. Then， a network embedding model is used to obtain the initial representation of the network. On this basis， a sampling strategy based on node importance is designed to select landmark nodes and improve the random sampling strategy. The similarity matrix of network nodes based on landmark nodes is established， and the low rank matrix approximation is introduced. Finally， the two networks are aligned according to the similarity of node representation. Experimental results on three data sets show that the proposed model is better than other baselines.

Abstract:With the rapid development of surface mount technology， higher requirements are put forward for mark positioning technology based on machine vision. In this paper， a fast and high accuracy positioning algorithm for triangular ring mark is proposed. By using geometric features of convex hull and constructing the concept of deviation histogram， the negative factors which affect the positioning accuracy are eliminated from coarse to fine， such as arc corners， convex points， subtle bumps and burrs. Then， a high-quality data set for fitting sides of triangle ring is obtained. Finally， the fitted equations of each side of the mark are obtained through the linear fitting with the constraint of minimum distance， and the geometric center of the triangle ring mark is calculated to realize the high accuracy positioning. The proposed algorithm provides a state-of-the-art idea for linear marks positioning with high accuracy and fast speed requirement.

Abstract:Existing age estimation methods of performance measurement are mainly based on the training set and testing set of independent identically distributed hypothesis. In order to better conform to the actual scene and better assess the age estimation method of generalization performance， a kind of heterogeneous data sets to evaluate agreement is put forward， i.e. paying more attention to the training set and test set with different distribution and characteristics. In addition， in order to improve the age estimation method based on convolution neural network fitting ability， on the basis of fully considering the adjacent age characteristics， a new theory of loss function analysis is proposed through modeling the age estimation problem as the label distribution study based on Gauss model. Theoretical analysis and experimental results show the effectiveness and robustness of the proposed method.

Abstract:The maintenance of power system is an important guarantee for the stable operation of power system. The power inspection based on intelligent algorithm provides convenience for the maintenance of power system. Extracting power line is a key technology for autonomous power inspection and ensures the safety of aircraft at low altitudes. At the same time， extracting power line using deep learning is an important breakthrough of power inspection. Based on these， deep learning is applied to extract power line. Combined with the characteristics of power line， the improved strategy of inputting image and attention module are embedded. The model of extracting power line based on stage attention mechanism （SA-Unet） is proposed. In the coding stage， the stage input fusion strategy （SIFS） is adopted to make full use of the multi-scale information of the image to reduce the loss of spatial information. In the decoding stage， the features of power line are focused by the embedded stage attention module （SAM）， and high-value information is screened out from a large amount of information quickly. Experimental results show that the method has good performance in multiple scenes with complex backgrounds.

Abstract:Due to the increasing complexity of aerospace exploration， integrated circuits are applied in many aerospace electronic systems such as deep space communication and attitude control. With the further shrinking of integrated circuit technology， the probability of errors in circuit due to single event effects has become higher. Evaluating the sensitivity of integrated circuits to single event upset （SEU） is of great significance to the development of aerospace. The continuous increase of circuit scale and the improvement of system function integration pose severe challenges to the speed of evaluation. For this reason， this paper proposes a fast fault injection method suitable for very large scale integration （VLSI）. This method can automatically analyze the circuit through scripts， and modify the logic to make the circuit available for fault injection. Experiment results show that the fault injection speed can reach nanosecond level，which can alleviate the contradiction between circuit scale and evaluation time. Consequently， it can meet the evaluation requirements of VLSI.