Brain functional connectivity networks (FCNs) derived from resting-state functional magnetic resonance imaging (rs-fMRI) data have been widely used to identify altered brain network patterns in attention-deficit/hyperactivity disorder (ADHD). Current graph neural network (GNN) approaches using FCNs predominantly emphasize node features while underutilizing edge information. Moreover, these GNN-based methods also inadequately represent dynamic interdependencies among evolving node features across network layers, limiting their diagnostic performance. We present a graph convolutional network via joint feature learning between nodes and edges (JNEL-GCN) that integrates neuroimaging features for ADHD classification and biomarker discovery. Our framework constructs dual graph representations: 1) a node graph using amplitude of low-frequency fluctuations (ALFF) measures across multiple frequency bands as nodal features, along with functional connectivity (FC) and node feature relationship matrices as edge attributes; and 2) an edge graph derived through line graph theory, enabling the interchange of node and edge roles. By leveraging the dual-graph design, our model implements an alternating feature update mechanism with optimized graph convolution operations, facilitating feature hierarchical learning of node-edge relationships across network layers. Extensive experiments demonstrate remarkable performance, achieving 97.3% accuracy on ADHD200 and 97.1% on ABIDE-I datasets, significantly outperforming current benchmarks. Meanwhile, gradient-based biomarker analysis identifies significant regions in bilateral limbic and default mode networks associated with ADHD, aligning with the findings in existing literature. Therefore, this dual-graph approach advances neuroimaging-based diagnosis by comprehensively capturing dynamic network interactions, while providing interpretable biomarkers for clinical neuroscience applications.

In this study, we apply a graph convolutional network (GCN) in attention deficit hyperactivity disorder (ADHD) classification by using multimodal data. Here, multimodal data is integrated to construct a dual graph for leveraging the modality information. Then, a GCN learning model is performed within an existing binary hypothesis testing (BHT) framework to fulfill the classification task. To preserve specifical topological structure of dual graph, we propose a graph-based feature selection approach to choose typical data (nodes) on the graphs. With these typical data, we design our GCN model, effectively learning the high-level features for identifying ADHD subjects. The experiments show the average accuracy of our method reaches 95.2% on various ADHD-200 datasets. Importantly, this result is achieved only on the right limbic system rather than the whole brain. The limited brain regions used indicate their potential ability as biomarkers. Moreover, biomarkers are found with their special topological structures. On the Peking University dataset, the found biomarkers center around the hippocampus and parahippocampal gyri, which are highly correlated with ADHD disease. Differing from isolated biomarker findings in traditional methods, our method discloses the abnormal circuit for ADHD and is more meaningful for further learning ADHD mechanism.

This paper introduces a novel joint spatio-temporal filtering algorithm and investigate its impact on data alignment in transfer learning (TL) for motion imagery (MI) tasks to deal with the variability in subjects, trials, or tasks. While spatial filtering is an integral part of the common spatial pattern (CSP) algorithm, temporal filtering is introduced to deal with EEG signals of each channel. The proposed algorithm jointly optimizes the coefficients and the subsequent alignment of covariance matrices. Experimental results on various public datasets validate the effectiveness of our proposed algorithm.

Point cloud segmentation plays a crucial role in extracting unique attributes and separating various objects, thereby enabling semantic comprehension and analysis. In this paper, we introduce a novel point cloud segmentation approach based on Diffusion Probabilistic Network (DDPM). The proposed model treats points as particles undergoing diffusion towards a noise distribution, and a reverse diffusion process transforms this noise distribution into the desired shape. Leveraging a Markov diffusion model in the reverse process enables generating point clouds with more refined and specific topological structures. After the diffusion step, multi-scale sampled features are fused to enhance the discriminative representation of 3D shapes. Objective and subjective experimental results demonstrate that our segmentation method outperforms state-of-the-art techniques in terms of evaluation metrics.

We propose a subspace-enhanced hypergraph neural network (seHGNN) for classifying anxiety disorder (AD). By leveraging a learnable incidence matrix, seHGNN strengthens the influence of hyperedges in graphs and enhances feature extraction performance of HGNNs. Then, we conduct this model within an existing binary hypothesis testing framework, where multi-modal data on the limbic system is integrated into a hypergraph. Experiments show that the seHGNN achieves a high accuracy of 90.7% for AD classification, surpassing other deep-learning-based methods, especially GNN-based methods. Our seHGNN also successfully identifies discriminative AD biomarkers, consistent with existing reports. This provides strong evidence supporting the effectiveness and interpretability of our proposed method.

Deep learning methods have been extensively employed in the classification of attention deficit hyperactivity disorder (ADHD) in decades. However, these methods either are still with unsatisfactory accuracy or lack the ability to capture relevant biological markers. To address these challenges, we introduce a knowledge distillation architecture within a binary hypothesis testing framework. In detail, we employ an existing AENet as a teacher model and guide a student model in learning high-level features of ADHD disease. More importantly, an attention block is adopted in this student model to obtain attention weights and better quantify the contributions of used brain functional connections. Now, these weights become robust in coping with the individual diversity among ADHD and healthy control groups, making it convenient for biomarker detection. Experiments demonstrate that our method attains an average classification accuracy of 99.5%. ADHD biomarkers are effectively identified and solidly supported by recent reports. This further illustrates the validity of our knowledge distillation approach.

Deep learning methods have been widely applied in Attention Deficit Hyperactivity Disorder (ADHD) classification in the past decade due to their effective learned features. However, these features are lack of neurobiological meanings and hard to be biomarkers. Here, we proposed an attention auto-encoding network with triplet loss (Tri-Att-AENet) for both ADHD classification and biomarker identification. Taking brain functional connectivities (FCs) as material, we introduced an attention encoding subnetwork to obtain the weighted FCs with their weights as attention scores. A triplet loss function was further employed on these scores, providing sufficient evidence for biomarker selection. Meanwhile, the weighted FCs became discriminative to pursue a higher accuracy. Experiments show that our method achieves an average accuracy of 99.6% for classification. The selected FC biomarkers are in accord with reported neurobiological results and well fulfill the task of biomarker identification.

Abstract Deep‐learning methods have been developed in recent years and have achieved dramatic improvements for image denoising. The existing deep‐learning methods can be conducted using two major models: Encoder–decoder and high‐resolution, where the high‐resolution model has superior resolution ability for detail description and restoration. In this study, a high‐resolution‐based network called multiscale residual fusion network (MRF‐Net) is proposed, which employed the spatial and contextual information of images. In detail, dilated convolution layers are used to enlarge the network's receptive field and learned sufficient features in a multiscale feature extracting module. The function of dilated convolution is reinterpreted here and it is viewed as a complex downsampling operation. Therefore, multiscale feature analysis could be performed in the proposed network by dilated convolution. Multilevel feature maps are sequentially obtained through a residual projection module, where considerable contextual and spatial information was collected from the multiscale features. In a residual fusion module, all maps were aggregated to generate a residual image effectively for noise removal. Experiments demonstrated that the MRF‐Net outperformed several state‐of‐the‐art model‐based and deep‐learning methods in both blind and non‐blind image denoising tests. Meanwhile, ablation studies were executed to verify the denoising performance of each module. Moreover, this method exhibited high computational efficiency, thus demonstrating its practicability.

Attention Deficit Hyperactivity Disorder (ADHD) is a high incidence of neurobehavioral disease in school-age children. Its neurobiological classification is meaningful for clinicians. The existing ADHD classification methods suffer from two problems, i.e., insufficient data and noise disturbance. Here, a high-accuracy classification method is proposed, which uses brain Functional Connectivity (FC) as material for ADHD feature analysis. In detail, we introduce a binary hypothesis testing framework as the classification outline to cope with insufficient data of ADHD database. Under binary hypotheses, the FCs of test data are allowed to use for training and thus affect the subspace learning of training data. To overcome noise disturbance, an l2,1-norm LDA model is adopted to robustly learn ADHD features in subspaces. The subspace energies of training data under binary hypotheses are then calculated, and an energy-based comparison is finally performed to identify ADHD individuals. On the platform of ADHD-200 database, the experiments show our method outperforms other state-of-the-art methods with the significant average accuracy of 97.6%.

In this paper, we develop a novel deep-network architecture for semantic segmentation. In contrast to previous work that widely uses dilated convolutions, we employ the original ResNet as the backbone, and a multi-scale feature fusion module (MFFM) is introduced to extract long-range contextual information and upsample feature maps. Then, a graph reasoning module (GRM) based on graph-convolutional network (GCN) is developed to aggregate semantic information. Our graph reasoning network (GRNet) extracts global contexts of input features by modeling graph reasoning in a single framework. Experimental results demonstrate that our approach provides substantial benefits over a strong baseline and achieves superior segmentation performance on two benchmark datasets.

Because the spectrum folding phenomenon affects the down-sampling of graph signals, both critically sampled and oversampled graph filter banks with down-sampling operations can only be applied to bipartite graphs. However, general graph signals may not reside on bipartite graph structures. In this paper, we present a novel bipartite graph approximation algorithm, which aims to find a bipartite graph sufficiently close to the original graph. To tackle this problem, we first show that, if the non-negativity constraint of adjacency matrices is removed, closed-form solutions can be readily obtained by the eigenvalue decomposition. Based on this fact, an alternating direction method of multipliers (ADMM) is further developed to achieve a real adjacency matrix. Experimental results show that the proposed algorithm outperforms the other proposals in terms of approximation accuracy.

Recently the development of deep learning has been propelling the sheer growth of vision and speech applications on lightweight embedded and mobile systems. However, the limitation of computation resource and power delivery capability in embedded platforms is recognized as a significant bottleneck that prevents the systems from providing real-time deep learning ability, since the inference of deep convolutional neural networks (CNNs) and recurrent neural networks (RNNs) involves large quantities of weights and operations. Particularly, how to provide quality-of-services (QoS)-guaranteed neural network inference ability in the multitask execution environment of multicore SoCs is even more complicated due to the existence of resource contention. In this article, we present a novel deep neural network architecture, MV-Net, which provides performance elasticity and contention-aware self-scheduling ability for QoS enhancement in mobile computing systems. When the constraints of QoS, output accuracy, and resource contention status of the system change, MV-Net can dynamically reconfigure the corresponding neural network propagation paths and thus achieves an effective tradeoff between neural network computational complexity and prediction accuracy via approximate computing. The experimental results show that (1) MV-Net significantly improves the performance flexibility of current CNN models and makes it possible to provide always-guaranteed QoS in a multitask environment, and (2) it satisfies the quality-of-results (QoR) requirement, outperforming the baseline implementation significantly, and improves the system energy efficiency at the same time.

Processing-in-Memory (PIM), has recently been revisited as one of the most promising solutions to deal with the issue of bandwidth and power wall between processor and memory. In this paper, we propose a light-weight PIM architecture, approxPIM, which leverages approximate computing techniques to enable InMemory Processing in a realistic 3D-stacked DRAM, Micron's Hybrid Memory Cube (HMC). Using the newly-released atomic instruction support of the HMC, approxPIM can process a wide range of data-intensive applications without adding any logic resources into the memory devices. Furthermore, we propose to approximate those accuracy-insensitive applications with the limited functioning set of HMC commands so that they could be smoothly mapped to the HMCs without the inference from processors, therefore enabling energy-efficient Processing-in-Memory and greatly expanding the scope of target PIM applications with HMC. In general, approxPIM gives a comprehensive study on HMC's potential and weakness in the application of Processing-in-Memory. Evaluation results show that our approxPIM significantly boosts the energy-efficiency and performance of the whole system.

Restoring underwater image from a single image is known to be an ill-posed problem. Some assumptions made in previous methods are not suitable in many situations. In this paper, an effective method is proposed to restore underwater images. Using the quad-tree subdivision and graph-based segmentation, the global background light can be robustly estimated. The medium transmission map is estimated based on minimum information loss principle and optical properties of underwater imaging. Qualitative experiments show that our results are characterized by relatively genuine color, natural appearance, and improved contrast and visibility. Quantitative comparisons demonstrate that the proposed method can achieve better quality of underwater images when compared with several other methods.

A novel stability condition is developed in this paper. It is both necessary and sufficient, which ensures that optimal design cannot be excluded from the admissible solutions. Compared to other necessary and sufficient stability conditions, the proposed one can be expressed as a quadratic constraint in terms of denominator coefficients, which facilitates its combination with other widely used IIR filter design strategies. In this paper, we adopt the Steiglitz-McBride scheme to design IIR filters. In each iteration, an approximation version of the proposed stability condition is further expressed as a set of linear inequality constraints, such that the resulting design problem becomes a quadratic program that can be efficiently and reliably solved. Simulations demonstrate the effectiveness of the proposed stability condition.

In this paper, a sparse approximation algorithm using eigenvectors of the graph Laplacian is proposed for image denoising, in which the eigenvectors of the graph Laplacian of images are incorporated in the sparse model as basis functions. Here, an eigenvector-based sparse approximation problem is presented under a set of residual error constraints. The corresponding relaxed iterative solution is also provided to efficiently solve such problem in the framework of the double sparsity model. Experiments show that the proposed algorithm can achieve a better performance than some state-of-art denoising methods, especially measured with the SSIM index.

A large number of experiments have demonstrated that for an FIR filter the sparsity of filter coefficients is highly related to its filter order. However, traditional sparse FIR filter design methods focus on how to increase the number of zero-valued coefficients, but overlook the impact of filter orders on design performance. As an attempt to jointly optimize filter length and sparsity of an FIR filter, a novel method is proposed in this paper to design sparse linear-phase FIR filters. With peak error constraints, the objective function of the design problem is formulated as a combination of the sparsity of filter coefficients and a measure of the effective filter order. Then, the design problem is then recast as a weighted l0-norm optimization problem, which is solved by an efficient numerical method based on the iterative-reweighted-least-squares (IRLS) algorithms. Experimental results illustrate that the proposed method can efficiently reduce the effective filter order while enhancing the sparsity of an FIR filter.

Sparse representation theory has been well developed in recent years. In this paper, we consider an image denoising problem which can be efficiently solved under the framework of the sparse representation theory. The traditional image denoising methods based on the sparse representation seldom take into account the special structure of the data. As an attempt to overcome such problem, the Graph regularized K-means singular value decomposition (Graph K-SVD) algorithm is proposed with the manifold learning. The local geometrical structure of the image is considered in the sparse optimization model with the graph Laplacian. This manifold-based optimization problem is well solved in the framework of the traditional K-SVD algorithm. Since the novel strategy adds a graph regularizer to the sparse representation model in order to emphasize the correlations among image blocks, the Graph K-SVD can achieve better denoising performance than the traditional K-SVD.