Search

Magnetic resonance imaging (MRI), a non-ionizing radiation imaging method, is widely utilized in diagnosing Alzheimer's disease (AD) due to its excellent imaging ability for brain soft tissues and high-resolution slice imaging. However, the enormous size of 3D MRI makes it difficult to process and analyze it. Therefore, a challenge is to accurately mine encephalatrophy patches from 3D MRI and build a patch-based diagnostic model. The existing patch-based methods mainly extract and fuse features of each patch in isolation, ignoring mutual information between patches, which makes the diagnosis performance unsatisfactory. We propose a novel dual-branch AD diagnostic model based on distinguishing atrophic patch localization. (1) We propose a Distinguishing Atrophic Patch Localization (DAPL) algorithm based on Distinguishing Index (DI) and Spatial Contact Ratio (SCR) to extract lesion areas that have significant impacts on diagnosis from 3D MRI. Meanwhile, we proposed a Discontinuity-voxel-based Dynamic Voxel Wrapping algorithm (DV 2 W) to calculate DI for each patch. (2) A dual-branch diagnostic network (DBDN) is constructed to obtain intra-patch and inter-patch features synchronously. The intra-feature extraction branch extracts feature from each patch through a parallel multi-channel network and fuse them. The Inter-feature extraction branch defines the Spatial Context Mixing matrix (SCM) and performs feature extraction on SCM to obtain mutual information. The performance evaluation demonstrates that our DBDN model has adequate diagnostic performance compared to state-of-the-art methods. In addition, the distinguishing pathological locations identified by our DAPL algorithm can effectively guide inexperienced clinical doctors to identify lesion areas and guide doctors in diagnosis quickly. [ABSTRACT FROM AUTHOR]

In recent years, graph neural networks (GNNs) have developed rapidly. However, GNNs are difficult to deepen because of over-smoothing. This limits their applications. Starting from the relationship between graph sparsification and over-smoothing, for the problems existing in current graph sparsification methods, we propose a novel graph sparsification method SimSparse based on vertex feature similarity and theoretically prove that it can help to relieve the over-smoothing. Furthermore, we also propose its derivatives SimSparse-G (SimSparse with G umbel-Softmax) and SimSparse-GC (SimSparse-G with memory occupancy reduC tion), which can achieve end-to-end learning. Moreover, SimSparse-GC can apply to larger graphs. Based on the three graph sparsification methods, we further propose a general sparse-convolution block SparseConvBlock with a sparsification layer and a graph convolutional layer to construct deep GNNs in which over-smoothing can be better alleviated. Extensive experiments show that the GNNs constructed by SparseConvBlock have better performance when the numbers of the network layers increase (i.e., in deep GNNs). Furthermore, we verify that our graph sparsification methods help to relieve over-smoothing in deep GNNs. [ABSTRACT FROM AUTHOR]

Graph convolutional networks have developed rapidly these years. Over-smoothing is an important factor that makes it difficult to deepen the networks, affecting the further development of graph convolutional networks. There have been some studies to solve the over-smoothing issue. However, the shapes of the filters in the existing graph convolutional networks are fixed for various data, and the convolution operations in some networks require eigendecomposition. Starting from the relationship between filters and over-smoothing, we propose a novel way to overcome over-smoothing in this paper. In this way, a novel quadratic polynomial filter (QPF) is proposed, and then a quadratic polynomial graph convolutional network (QPGCN) is derived. Without increasing the complexity of the network, QPGCN can adaptively learn the shape of QPF and does not require eigendecomposition, which can better alleviate over-smoothing. The extensive experiments on Cora, Citeseer, Pubmed, DBLP and Ogbn-arxiv datasets show that QPGCN achieves state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Intelligent Transportation System (ITS) is now being widely built all over the world. Traffic congestion prediction, as a major part of ITS, not only provides reliable traffic information for travelers to save their time, but also assists traffic management agencies to manage the traffic system. We find that the existing papers, including neural networks, do not perform well due to the complex association features between the road sections. In addition, the lack of a better evaluation standard for traffic congestion also makes the effect of traffic congestion prediction worse. In this paper, we propose a method to generate traffic congestion index by mining free-stream speed and free-stream flow. Considering the association features of road segments in the road network, we propose a road segment grouping method based on association subgraph to pre-train deep learning model and realize information sharing between road segments. Combining the features of traffic data and CNN model, we propose a traffic congestion prediction model named SG-CNN which training process is optimized by road segment grouping algorithm. The experiments demonstrate that our proposed model has higher accuracy rate than other models. [ABSTRACT FROM AUTHOR]

Robust object tracking is still a challenging task in the field of computer vision and has application value in many fields such as automatic driving, human–computer interaction and robot visual navigation. More and more researchers are devoted to researching more accurate object tracking methods. How to better deal with occlusion and deformation has always been the difficult challenges in the object tracking field, and the existing methods cannot solve these problems well. In this regard, we propose a novel, effective and portable module called part-based tracking and assembly (PTA), which is added to the fully convolutional siamese networks to divide the exemplar feature map into several parts. Each part is separately tracked, and then the tracking results of all parts are assembled to obtain the final tracking results. And the experiments on several popular tracking benchmarks show our variant trackers with the PTA module that operate at almost the same tracking speed with the original trackers and achieve superior tracking performance. Moreover, the tracking accuracy is significantly improved on the data with occlusion, deformation and background clutter. Compared with some real-time tracking methods, our variant trackers with the PTA module can achieve the state-of-the-art performance. [ABSTRACT FROM AUTHOR]

In this paper, we present a robust fault-tolerant control scheme to achieve attitude control of flexible spacecraft with disturbances and actuator failures. It is shown that the control algorithms are not only attenuate exogenous bounded disturbances with attenuation level, but also able to tolerate partial loss of actuator effectiveness. The proposed controller design is simple and can guarantee the faulty closed-loop system to be quadratically stable with a prescribed upper bound of the cost function. The design algorithms are obtained by combining free weighting matrices method with linear matrix inequality technique. The effectiveness of the proposed design method is demonstrated in a spacecraft attitude control system subject to loss of actuator effectiveness. [ABSTRACT FROM AUTHOR]

With the wide use of EDGEs (electronic data gathering equipments) such as sensors and RFID (radio frequency identification) devices, unprecedented volumes of event streams have been generated. Mining frequent episodes within the latest time windows over event streams plays a significant role in event monitoring. It helps to generate episode rules, which can reflect the latest change, and predict future events effectively. The paper proposes how to mine MinEpi (minimal occurrence based frequent episode) within the latest time windows. The existing MinEpi mining methods are all Apriori-like, which need to scan data time after time and generate quantities of candidate episodes. This results in high time and space cost. Moreover, Apriori-like methods cannot be applied to event streams. For these problems, the paper proposes the episode matrix and frequent episode tree based mining method (EM&FET), which can generate frequent 2-episodes by constructing an episode matrix and generate higher-level frequent episodes directly by extending lower-level ones gradually, only scanning data once without candidate generation. Moreover, the paper further improves EM&FET, which enhances efficiency and saves space greatly. The experiments on different types of real data sets show the effectiveness and high efficiency of EM&FET and its improvement. [ABSTRACT FROM AUTHOR]