Multimodal cancer survival prediction draws on whole-slide pathology and transcriptomics, but it often treats predictive importance as evidence reliability, allowing poorly supported or conflicting observations to disproportionately affect patient-level estimates. This study presents Knowledge-enhanced Pathology–Pathway Survival (KP2Surv), a pathology–genomics framework that models patient-specific trust in evidence. Biological knowledge defines structured evidence units: spatial–phenotypic tissue groups form a pathology hypergraph, and curated molecular pathways form a pathway hypergraph. Separate token-importance and quality-risk estimates guide bidirectional pathology–pathway interaction. Pathology, pathway, and cross-modal branches predict continuous-time survival distributions, which are fused using predictive precision, learned reliability, and cross-branch disagreement. Across 4,263 patients in eight cohorts from The Cancer Genome Atlas (TCGA), KP2Surv achieved the highest numerical mean concordance index (C-index) among evaluated methods in every cohort; the unweighted mean was 0.7369 versus 0.7159 for the strongest comparator on average. Exact paired tests provided nominal evidence of improvement in four cohorts, while four comparisons were not statistically significant. Ablation and sensitivity analyses supported the selected evidence-trust components and default settings within the evaluated cohorts. Evaluation in two Clinical Proteomic Tumor Analysis Consortium (CPTAC) cohorts produced mean C-indices of 0.5987 for lung adenocarcinoma and 0.6143 for uterine corpus endometrial carcinoma, showing partial transportability but substantial attenuation under domain shift. Pathway-conditioned spatial attribution generated hypotheses linking molecular programs to localized tissue patterns. The findings support explicit modeling of evidence meaning and trustworthiness. Prospective validation and dedicated uncertainty evaluation remain necessary before clinical use.
We introduce -Net++, a novel three-stage architecture designed to address the persistent challenge of occlusion removal in light-field (LF) images by leveraging adaptive attention-guided feature integration and progressive hierarchical reconstruction. Unlike existing methods that struggle to fully exploit spatial hierarchies and adaptively restore occluded regions across scales, our model incorporates a ResASPP-AttFPN feature extractor, which integrates Residual Atrous Spatial Pyramid Pooling (ResASPP) with a spatial attention-enhanced Feature Pyramid Network (AttFPN) to selectively fuse multiscale features while emphasizing salient spatial cues essential for occlusion localization. The core of our framework is a tri-stage -Net++ reconstruction module, which performs progressive restoration through three hierarchically connected encoder-decoder stages of decreasing depth (4-level, 3-level, and 2-level), each built on VGG-based blocks and dense skip connections to recover increasingly refined background content. To further enhance detail preservation and structural consistency, we introduce a residual feature refiner (RFR) that consolidates residual cues and sharpens the boundaries of objects. Extensive experimental evaluations demonstrate that the proposed method surpasses recent state-of-the-art (SOTA) LF occlusion removal approaches-representing the most advanced and best-performing techniques reported in the literature-in both quantitative metrics and visual reconstruction quality. Specifically, our model achieves average improvements of 0.86 dB in PSNR and 0.016 in SSIM across real-world (CD scene) and synthetic LF datasets, including sparse (4-Syn, 9-Syn) and dense (Single-Occ, Double-Occ) settings. This capability is particularly relevant to the Big Data paradigm, where large-scale visual datasets demand robust preprocessing to remove occlusions and ensure reliable downstream analytics. By improving LF data fidelity while remaining efficient, our model supports scalable pipelines for high-volume visual data processing.
The simultaneous presence of URLLC and eMBB services in 5G brings about a problem of resource allocation. On the one hand, URLLC requires guaranteed low latency and reliability, whereas on the other hand, eMBB needs guaranteed high throughput and fairness. This paper addresses the problem of dynamic scheduling of orthogonal resource blocks in a single-cell environment, in order to decrease URLLC violations and increase eMBB throughput and fairness under the conditions of time-varying channels, bursty traffic and traffic distribution changes. Attention-enhanced meta-graph reinforcement learning algorithm is proposed. Dynamic contention graph represents competition for the resources at the MAC layer level; permutation-equivariant GATv2 encoder learns users’ interactions, and model-agnostic meta-learning allows to adapt the model quickly to new traffic. The scheduler selects the most prioritized user, and the rest of the resources are allocated using proportional-fair scheduling algorithm. Under the nominal evaluation condition, the proposed method achieves a URLLC violation rate of 63.11
Video surveillance systems have witnessed rapid advancements over the past two decades, fueled by progress in computer vision, artificial intelligence, and edge computing. However, despite extensive research, there remains a need for a review that synthesizes the architectures, methodologies, and evolving trends of intelligent video surveillance systems from a system-level perspective. Most existing literature reviews focus narrowly on individual tasks such as object detection, motion tracking, or anomaly recognition, without addressing how these components integrate within a complete system. This paper addresses this gap by conducting a systematic review of research published between January 2000 and July 2025, drawing from major digital repositories. The main contributions of this work are threefold: (1) it proposes a structured taxonomy of intelligent surveillance systems based on key functionalities, including object detection, motion tracking, abnormal event detection, vehicle and traffic monitoring, object counting, and multi-camera coordination; (2) it provides a comparative analysis of methodologies, datasets, and platforms used across studies; and (3) it explores the interconnections among these core tasks while identifying critical challenges such as real-time processing, privacy preservation, data scalability, and system interoperability. Key outcomes of this review include the identification of underexplored research areas—such as federated learning and cross-camera integration—and practical guidelines for designing next-generation surveillance systems. The literature analyzed was sourced from major digital repositories including IEEE Xplore, ScienceDirect, Google Scholar, arXiv, Springer, and the ACM Digital Library.
Accurate wind speed forecasting is increasingly important for supporting reliable wind power grid integration and enhancing early warning of extreme wind events. Existing spatiotemporal models can fuse multiple meteorological variables and perform well in medium-resolution wind field forecasts. However, when handling high-resolution data, they often produce overly smooth wind fields, miss local details, and exhibit large errors. To address these problems, we propose SMWPF, an edge-cloud collaborative framework for high-resolution wind field prediction. A lightweight edge-side predictor first generates low-latency forecasts. Its latent representations are refined by soft vector quantization using a two-layer multilayer perceptron (MLP) and an extended codebook, which helps alleviate over-smoothing. An extreme wind speed factor is fused with elevation information to guide the attention module toward regions historically susceptible to strong winds. The framework then estimates the confidence of each edge forecast. Low-confidence samples are sent to a Qwen-based cloud predictor, which maps the original normalized historical meteorological tensor to the future wind field for re-prediction. Extensive experimental results on both private and public datasets indicate that SMWPF reduces root mean square error (RMSE) by 5
