Consistency learning is considered one of the most effective approaches for leveraging unlabeled data in semi-supervised medical image segmentation(SSMIS). However, existing methods often rely on a single perturbation strategy or only apply perturbations to unlabeled data, resulting in limited exploration of the perturbation space and an inability to learn precise decision boundaries. Moreover, excessive perturbations tend to cause training instability due to the accumulation of errors. To address the aforementioned issue, this paper proposes the BlenMatch method, which ensures the effect of consistency learning by using high perturbations initially, while striving to maintain the stability of the training. Specifically, BlenMatch applies random complementary interchange (RCI) and feature-level mixed perturbations on unlabeled data alongside labeled data, it ensures comprehensive exploration of the perturbation space while mitigating the negative effects of noisy pseudo-labels, leveraging ground truth guidance to counteract excessive perturbations. Additionally, SL loss is introduced to enhance training stability and prevent overfitting. Experimental results demonstrate that BlenMatch outperforms current state-of-the-art techniques on the ACDC and PROMISE12 datasets.

Semi-supervised learning methods combine a small amount of labeled data with a large amount of unlabeled data to achieve high-precision segmentation while reducing the annotation cost. However, existing semi-supervised learning methods usually focus on data-level perturbations or improvements in network structures, ignoring the problem of insufficient information interaction between different branches and regions in complex texture tasks. In addition, in scenarios that require high-resolution processing (such as 3D medical images), traditional Transformer-based methods are not only computationally expensive but also prone to overfitting problems. Therefore, to address the above problems, we propose a Mamba based Consistency Constraints network (MCC-Net) for semi-supervised 3D medical image segmentation. Specifically, the model is organized by a shared encoder and three-branch decoder architecture. First, a consistency regularization constraint mechanism is introduced to combine the segmentation map of one decoder with the pseudo-labels of other decoders, thereby capturing more valuable features in high-uncertainty areas and generating stable, low-entropy predictions; second, a new consistency loss function is designed to construct constraints between the signed distance map (SDM) of the two auxiliary decoders and the segmentation map of the main decoder to enhance the learning ability of the target geometric structure. In addition, a Fusion Mamba(FM) Block is proposed to improve the model’s capabilities in deep semantic feature extraction and computational efficiency by modeling long-distance dependent features. Experimental results on public dataset show that, compared with six state-of-the-art semi-supervised segmentation methods, our method achieves Dice scores of 89.48%, 91.46% and 91.96%, respectively, when 10%, 20% and 30% of labeled data are used for training, significantly outperforming the other methods. The experimental results show that the model has strong advantages in both segmentation accuracy and utilization of unlabeled data.

As a crucial component of autonomous vehicles and intelligent robots, human trajectory prediction provides route planning for intelligent agents. Recent studies have shown that human/pedestrian trajectory prediction models are vulnerable to adversarial attacks. However, previous studies failed to consider the practicality of perturbations, generating adversarial trajectory with the length equal to the input length required by the prediction model. It is hard for an attacker (signed as candidate agent) to precisely walk on such many nodes in adversarial trajectory to execute an attack. Moreover, when selecting a target agent to approach for candidate agent, the attack is based solely on the distance to the candidate agent, ignoring the target agents with high relative velocity. This paper proposes a two-stage attack called AGNS aimed at minimizing the number of perturbing trajectory nodes meanwhile keeping the attack effectiveness. In the first stage, we propose a node-selection method to select the "important" nodes to reduce the number of nodes that we need to perturb. In the second stage, the selected nodes are perturbed with the attention-guided loss to generate a partial adversarial trajectory. Experiments on four models demonstrate that AGNS causes an average collision rate of over 80% while perturbing only 56% of the trajectory nodes, and even achieves an average collision rate exceeding 70% when perturbing just a single node. Our code can be obtained on Github: https://anonymous.4open.science/r/AGNS.

Deep learning is widely used in the early diagnosis of Alzheimer ’s disease (AD) in recent years, and has achieved better performance than traditional machine learning methods. The current deep methods for early diagnosis of AD mainly adopt a fully supervised method, which relies too much on a large number of labeled high-quality medical image data, and has a large performance loss in under-labeled data scenarios. Meanwhile,the existing semi-supervised methods ignore the reliability of pseudo-labeling. Aiming at the above problems, a Dual Augmentation Semi-supervised Learning (DASSL) method is proposed. DASSL designs a feature enhancement method based on attention mechanism and combines it with a spatial enhancement method more suitable for medical images, which provides a new solution for early feature recognition of Alzheimer ’s disease. At the same time, a new loss function is designed, and combined with the threshold, high confidence samples are selected to cope with the challenge of high sample impurity rate. The DASSL method proposed in this paper was evaluated in 518 subjects on the ADNI-1 dataset. The experimental results show that DASSL achieves the best accuracy and stability in classifying AD, MCI, and NC (more than 95% accuracy for classifying AD and CN task, and more than 90% accuracy for classifying MCI and CN) compared to other semi-supervised methods.

Monocular Depth Estimation (MDE) is vital for autonomous driving, enabling 3D perception without costly LiDAR, yet existing 2D white-box adversarial attacks against MDE lack robustness to viewpoint and real-world variations. We propose DepthHack, the first 3D black-box adversarial attack framework for MDE, using probabilistic sampling and score-based optimization to craft robust 3D adversarial textures. DepthHack ensures robustness across diverse weather conditions and viewpoints, achieving an average depth estimation error of 8.88 m on Monodepth2 (Carla), surpassing HardBeat by 2.1%, with only 60k queries. This work reveals MDE vulnerabilities, enhancing the safety of autonomous systems. Experiments on four MDE models and two datasets validate its superior performance, efficiency, and cross-dataset generalization.

With the proliferation of trajectory data generated by many Internet of Things (IoT) devices in the AI-driven transportation field, trajectory representation is crucial for extracting individual behavioral characteristics from intelligent IoT systems. Neural network-based trajectory representation learning methods excel at obtaining consistent representations of individual movements and mining spatio-temporal autocorrelation features of trajectories. However, existing methods achieve high accuracy in reliable experimental data, and their high performance is not always available in real-world wild scenarios, which is not available to reveal the performance bounds of trajectory representation learning models when confronted with real-world phenomena. Given this, we propose a universal trajectory adversarial attack framework that investigates the robustness vulnerabilities of trajectory representation learning models by generating adversarial trajectory examples. We reveal the impact of the attack surface in practical deployment and the adversarial perturbation budget on the trajectory adversarial attack performance. Guided by the new framework, we propose a universal black-box adversarial attack for trajectories, named Universal Trajectory Customized Iteration Attack (UTCIA). Specifically, we simulate trajectories through single-point perturbations to obtain a vulnerability-dependent set of victim points, selecting and perturbing only a small subset of points to degrade the entire model performance. Furthermore, we aggregate the discrete set of salient target trajectory points into a high-dimensional space for iterative direction estimation, and relax the constraint region to further compress noise. We generate adversarial trajectory examples for two trajectory representation downstream tasks with fundamentally different objectives using multiple large-scale real-world datasets. Extensive experiments demonstrate the feasibility and generalizability of our proposed framework in exploring the robustness vulnerabilities of trajectory representation learning models.

Automatic nuclei segmentation and classification (NSC) is a fundamental prerequisite in digital pathology analysis as it enables the quantification of biomarkers and histopathological features for precision medicine. Nuclei appear to be small, however, global spatial distribution and brightness contrast, or color correlation between the nucleus and background, have been recognized as key rationales for accurate nuclei segmentation in actual clinical practice. Although recent great breakthroughs in medical image segmentation have been achieved by Transformer-based methods, the adaptability of segmenting and classifying nuclei from histopathological images is rarely investigated. Also, the severe overlap of nuclei and the large intra-class variability are common in clinical wild data. Prevailing methods based on polygonal representations or distance maps are limited by empirically designed post-processing strategies, resulting in ineffective segmentation of large irregular nuclei instances. To address these challenges, we propose a keypoint-guided tri-decoder Transformer (PointFormer) for NSC simultaneously. Specifically, the overall NSC task is decoupled to a multi-task learning problem, where a tri-decoder structure is employed for decoding nuclei instance, edges, and types, respectively. The nuclei detection and classification (NDC) subtask is reformulated as a semantic keypoint estimation problem. Meanwhile, introduces a novel attention-guiding strategy to capture strong inter-branch correlations and mitigate inconsistencies between multi-decoder predictions. Finally, a multi-local perception module is designed as the base building block of PointFormer to achieve local and global trade-offs and reduce model complexity. Comprehensive quantitative and qualitative experimental results on three datasets of different volumes have demonstrated the superiority of the proposed method over prevalent methods, especially for the PanNuke dataset with an achievement of 70.6% on bPQ.

Deep neural networks are extensively employed in CAD (computer-aided diagnosis) to enhance diagnostic efficiency, accuracy, and reliability. However, relying on manually annotation, labeling medical images is costly and time-consuming. Consequently, obtaining labeled data of sufficient scale and high quality poses a challenge. In the field of data science, semi-supervised learning (SSL) methods have significantly reduced the burden of dataset labeling. However, two primary issues exist: (1) the multi-classification challenge involving simultaneous classification of multiple lesions and diagnosis of multiple diseases; (2) data imbalance results in significant variations in disease prevalence rates. In this paper, the design of Curriculum Pseudo-labels (CPL) methods is adopted to overcome the limitations of applying SSL methods across data science and CAD. Additionally, a novel strategy called MedAugment is proposed for enhancing the model's robustness and generalization ability. Experimental results show that the design of MedMatch model reaches 94.21% in AUC score compared with state-of-the-art SSL methods on CAD task.

Nuclei detection and segmentation are indispensable prerequisites in digital pathology research, whereas the precise segmentation of nuclei by domain experts relies heavily on global spatial information and the inter-nuclei correlation. However, previous automatic nuclei segmentation works are mostly built on convolutional neural networks, which are unable to capture long-range global context with their inherent convolutional operations. Additionally, window-based design in few transformer-based approaches limits remote token interactions. The present study introduces a novel Multi-Local Perception (MLP) network, MLPSeg, which is proposed to address the aforementioned challenging issues. Specifically, the parallel computation of depthwise separable convolution and local window attention is designed to extract local information. Then, the parallel module of local horizontal attention and local vertical attention is designed to establish the global dependency. Moreover, to model the cross-scale dependencies and narrow the contextual semantic gap, the Context Cross Attention (CCA) is introduced for optimising skip connections. A tri-decoder structure is adopted to generate nuclei instance masks, normal edge masks and clustered edge masks. The superior performance of MLPSeg for nuclei segmentation is demonstrated across two datasets with different modalities, resulting in a 2.29% - 5.82% improvement compared to the state-of-the-art methods.

With the deep integration of computing and network development, as a new type of network infrastructure, computing power network (CPN) has become a research hotspot in the industry. Computing resource metrics integrates the computing resources connected to the CPN, realizes the collaborative management of heterogeneous resources through the measurement of multi-dimensional computing resource, and provides an accurate resource view for resource matching, which has become an important part of the CPN. The traditional measurement methods are too single to measure computing resources from a single dimension, which is difficult to adapt to the development of CPN. The existing methods of computing resource metrics need to be improved in the accuracy of resource matching and cannot reflect the comprehensive performance of computing resources. In this paper, a multi-dimensional computing resource measurement method based on entropy weight TOPSIS is designed to score the comprehensive performance of computing resources, storage resources and communication resources of computing nodes, then the nodes are divided into different categories of comprehensive performance according to the score, so as to narrow the scope of resource matching for different user requirements. At the same time, a multi-dimensional resource matching algorithm based on deep reinforcement learning is proposed. The resource matching process is constructed as a Markov decision process to realize the matching of tasks and nodes. The simulation results show that the proposed algorithm can better solve the matching problem of multi-dimensional resources, and the utilization rate of all kinds of resources reaches more than 90%.

Computing Power Network (CPN) overcome the performance limitations of computing power islands by integrating computing and network resources, dynamically scheduling business traffic to optimal nodes. However, effectively and collaboratively utilizing computing resources to reduce the delay of computing tasks has become a challenging issue in CPN due to the heterogeneity of resources and dynamic load. Existing works often treat workflow tasks as atomic tasks for offloading, disregarding subtask dependencies within a task. This approach leads to increased overall waiting time due to varying execution delays of each subtask. To address this problem, this paper proposes an optimization algorithm for workflow tasks based on slack quantity (CSA_WTO). The algorithm optimizes the arrival order of workflow task graphs before offloading, considering subtask dependencies and arranging them appropriately for subsequent offloading. This effectively reduces waiting delays for subtasks. Additionally, we utilize the Dependent Task Offloading algorithm (DTO) based on A3C to offload optimized workflow tasks, thereby improving execution efficiency in CPN. Simulation results demonstrate that compared with other algorithms, CSA_WTO significantly reduces task waiting delays by up to 85%, and DTO achieves a request acceptance rate of up to 96% while reducing the average completion time by 71%.

Recent advances in single-cell RNA sequencing(scRNA-seq) technologies have shown significant improvements in cellular-level biological research. Clustering individual cells into subpopulations is a pivotal procedure of scRNA-seq data analysis. The high sparsity and dimensionality of scRNA-seq data still limit the performance of cell clustering. Recent graph-based scRNA-seq clustering methods overlook the expression of cells, leading to suboptimal representation of cell embeddings. In this study, we propose single-cell Combined Graph Attentional Clustering (scCAT), a novel unsupervised clustering method for scRNA-seq data. scCAT first learns cell embeddings through a joint dimensionality reduction (JDR). Then, a graph attention autoencoder (GATE) is designed to analyze the cell topological information. Finally, the clustering operation is executed through a self-optimizing unit. Experimental results on five scRNA-seq datasets show that scCAT significantly improves clustering accuracy over existing methods with an average increase of 11.2% in the Adjusted Rand Index (ARI) and 9.6% in the Normalized Mutual Information (NMI) parameters.

Traffic flow prediction assumes a pivotal role in aiding governments and companies accurately forecast changes in vehicle volume, consequently enhancing transportation efficiency and facilitating vehicle travel. Presently, the majority of traffic flow prediction methods rely on centralized learning strategies, which entail the transmission of substantial data and may jeopardize user privacy. To address this issue, we propose a Federated Learning-based Attention Graph Convolutional Network (FLSTAGCN) algorithm for traffic flow prediction. Firstly, we develop a Spatial-Temporal Attention Graph Convolutional Network (STAGCN) method that employs attention mechanism to proficiently extract spatial-temporal features from traffic flow data, augmenting the model's learning capabilities. Subsequently, within the aggregation mechanism of Federated learning, we devise a bespoke optimal selection to enhance training accuracy and reduce communication costs in traffic flow prediction scenarios. Finally, we integrate Federated Learning with STAGCN and utilize the optimal selection protocol to designate participants for transmitting optimal parameters. The Experimental results substantiate that our approach outperforms advanced deep learning approaches in terms of traffic flow prediction performance while ensuring the privacy and security of traffic data.

Recent advances in large model and neuroscience have enabled exploration of the mechanism of brain activity by using neuroimaging data. Brain decoding is one of the most promising researches to further understand the human cognitive function. However, current methods excessively depends on high-quality labeled data, which brings enormous expense of collection and annotation of neural images by experts. Besides, the performance of cross-individual decoding suffers from inconsistency in data distribution caused by individual variation and different collection equipments. To address mentioned above issues, a Join Domain Adapative Decoding (JDAD) framework is proposed for unsupervised decoding specific brain cognitive state related to behavioral task. Based on the volumetric feature extraction from task-based functional Magnetic Resonance Imaging (tfMRI) data, a novel objective loss function is designed by the combination of joint distribution regularizer, which aims to restrict the distance of both the conditional and marginal probability distribution of labeled and unlabeled samples. Experimental results on the public Human Connectome Project (HCP) S1200 dataset show that JDAD achieves superior performance than other prevalent methods, especially for fine-grained task with 11.5%-21.6% improvements of decoding accuracy. The learned 3D features are visualized by Grad-CAM to build a combination with brain functional regions, which provides a novel path to learn the function of brain cortex regions related to specific cognitive task in group level.

With the development of deep learning in medical image analysis, decoding brain states from functional magnetic resonance imaging (fMRI) signals has made significant progress. Previous studies often utilized deep neural networks to automatically classify brain activity patterns related to diverse cognitive states. However, due to the individual differences between subjects and the variation in acquisition parameters across devices, the inconsistency in data distributions degrades the performance of cross-subject decoding. Besides, most current networks were trained in a supervised way, which is not suitable for the actual scenarios in which massive amounts of data are unlabeled. To address these problems, we proposed the deep cross-subject adaptation decoding (DCAD) framework to decipher the brain states. The proposed volume-based 3D feature extraction architecture can automatically learn the common spatiotemporal features of labeled source data to generate a distinct descriptor. Then, the distance between the source and target distributions is minimized via an unsupervised domain adaptation (UDA) method, which can help to accurately decode the cognitive states across subjects. The performance of the DCAD was evaluated on task-fMRI (tfMRI) dataset from the Human Connectome Project (HCP). Experimental results showed that the proposed method achieved the state-of-the-art decoding performance with mean 81.9% and 84.9% accuracies under two conditions (4 brain states and 9 brain states respectively) of working memory task. Our findings also demonstrated that UDA can mitigate the impact of the data distribution shift, thereby providing a superior choice for increasing the performance of cross-subject decoding without depending on annotations.