BackgroundDeep learning (DL) has been increasingly applied to grade knee osteoarthritis (KOA) on radiographs, but reported diagnostic performance varies across Kellgren-Lawrence (K-L) grades.PurposeTo systematically evaluate the diagnostic performance of DL models for radiographic KOA grading.Material and MethodsPubMed, Embase, and Web of Science were searched through November 2024 for studies using DL algorithms to grade KOA on X-ray images. Sensitivity and precision were synthesized. Heterogeneity was assessed using the I2 statistic. Subgroup analyses and meta-regression were conducted according to transfer learning, external validation, multi-task learning, joint training strategy, and data splitting. Publication bias was assessed using funnel plots and Egger's test. Study quality was evaluated using the revised QUADAS-2 tool.ResultsOf 1004 records screened, 32 studies were included. Pooled sensitivity for K-L grades 0-4 was 0.90, 0.66, 0.80, 0.87, and 0.88, respectively, and pooled precision was 0.87, 0.71, 0.81, 0.86, and 0.91, respectively. Diagnostic performance was poorest for K-L grade 1, particularly in sensitivity, indicating limited reliability for early-stage KOA detection. Heterogeneity was high across outcomes and grades, particularly for sensitivity in K-L grades 1 and 2 and precision in K-L grades 0 and 1. Meta-regression identified transfer learning and data splitting as potential sources of heterogeneity. Egger's tests suggested no statistically significant small-study effects.ConclusionDL models showed better diagnostic performance for moderate-to-severe radiographic KOA than for early-stage disease. However, the poor sensitivity for K-L grade 1, substantial heterogeneity, and limited external validation suggest that current DL models are not yet reliable for early KOA detection or ready for routine clinical implementation. Further standardized reporting, robust validation, and multicenter external evaluation are required.

Background and Objective: Due to the underutilization of advanced deep learning techniques in the diagnosis and treatment of checkpoint inhibitor-related pneumonia (CIP), this study proposed a CIP prediction algorithm based on multimodal data fusion. Methods: Specifically, the algorithm constructed a model using patients' computed tomography (CT) imaging, electronic medical records, and physiological examination reports. First, feature extraction modules were developed to process each data modality. Subsequently, multimodal features were fused using a cross-attention approach. Finally, these features were input into a classifier for classification. Results: Experimental results demonstrated that multimodal cross-attention network (MMCA-Net) significantly outperformed single-modality models and traditional fusion methods. In10-fold patient-level cross-validation, the proposed model achieved an average accuracy of 87.12% (+0.83%) and an area under the curve (AUC) of 0.8981 (+0.007). Furthermore, the algorithm showed excellent reproducibility, with performance deviation of less than 0.2% in independent replication trials. Quantitative analysis of attention weights confirmed that the model effectively integrated clinical context to resolve ambiguous radiological patterns. Conclusions: The proposed deep learning-based multimodal method provides a stable and highly accurate tool for predicting CIP. By integrating information from imaging, textual data, and laboratory results, MMCA-Net offers a valuable clinical reference for physicians, with the potential to enhance patient safety and improve treatment outcomes in the management of cancer immunotherapy.

Despite years of methodological progress, how far AI has come in liver fibrosis staging has never been systematically evaluated under the heterogeneous, multi-center conditions that define clinical practice. To address this gap, we introduce LiFS, a large-scale dataset and benchmark derived from the MICCAI 2025 CARE-Liver challenge, comprising 610 patients across multiple centers and scanners with multi-sequence MRI. To the best of our knowledge, LiFS is the first benchmark providing complete gadoxetic acid-enhanced sequences with histopathology-confirmed annotations from diverse real-world scanners. Through systematic evaluation of 9 independently developed methods selected from 96 registered teams against in-cohort radiologist reference results, our findings address how far current AI has progressed toward clinical-level liver fibrosis staging from three complementary perspectives. First, against radiologists, the best AI methods were broadly comparable to the senior radiologist and significantly exceeded the junior radiologist in selected settings, while median AI performance generally approached junior-radiologist levels. Second, from a data perspective, cross-center heterogeneity, label imbalance, and contrast-enhanced sequence variability emerge as the dominant challenges for AI methods. Third, from a technical perspective, methodological design choices, including spatial registration, input dimensionality, multi-modal fusion strategy, and backbone architecture, appear to modulate cross-center robustness, although no single choice alone closes the gap. Overall, LiFS provides a rigorous real-world benchmark for positioning the current state of AI in liver fibrosis staging and for enabling future research on the key challenges that limit clinically reliable deployment.

Checkpoint inhibitor pneumonitis (CIP) imposes substantial mortality risk in malignant tumor patients receiving immune checkpoint inhibitors (ICIs) therapy. Although computed tomography (CT) serves as the primary diagnostic modality for CIP, its heterogeneous imaging features pose critical challenges to clinical diagnosis. In this paper, we introduce DPVIT, a bitemporal CT analysis model designed to capture lung tissue changes. DPVIT features temporal phase embedding, encoding pre- and post-ICIs treatment scan timestamps for time-dependent pattern learning, and dynamic deformable convolution, which adapts sampling to detect subtle structural variations. Moreover, it replaces the Swin Transformer’s shifted window attention with a temporally extended grid attention module, enabling local-global feature interactions for multi-scale representation learning. In validation, DPVIT achieved an AUC score of 0.884, outperforming existing models and demonstrating its potential to improve the early and accurate diagnosis of CIP.

During the process of grain storage, aberrant temperatures in the grain can give rise to mold and pest infestation, posing a severe threat to food security. Therefore, accurately monitoring the trend of grain temperature changes is of great importance. We introduce a novel temperature forecasting approach for grain storage. It can accurately predict the temperature at different locations in the grain pile. By building spatial-temporal blocks utilizing the Graph Attention Network(GAT), attention model and the Gated Recurrent Unit(GRU), we culminate the entire procedure into a network as the Attention-based GNN Combine GRU Network, or A-GCRN for short. To fortify the practicality and validity of our model, we collected temperature data from more than 10 grain depots, and a total of 10,950 temperature data for model training and validation. Through the ablation experiments and comparison experiments conducted on this dataset, we have been able to convincingly validate the efficacy and rationality of the proposed method.

The pharmacokinetic (PK) parameters extracted from the DCE-MRI provide valuable information but suffer from many sources of variability. Thus, the efficient and fast estimation of the distributions of these ambiguous PK parameters caused by variabilities could significantly improve the robustness and repeatability of DCE-MRI. The estimation of the PK parameters’ distributions provides a way to quantify the PK parameters’ values and variabilities simultaneously. In this study, we demonstrated the feasibility of the normalizing flow-based distribution estimation network (FPDEN) for PK parameters’ distribution estimation in DCE-MRI.

Liver cirrhosis, a critical and potentially fatal liver condition, often progresses to severe complications, including liver failure and hepatocellular carcinoma. Early and accurate detection of this disease is imperative for the implementation of effective clinical strategies. Traditional diagnostic approaches, which rely heavily on invasive liver biopsies and are subject to interpretive variability, present significant limitations. The gut microbiome serves as a dynamic biomarker, reflecting the internal milieu of the body and exhibiting a strong correlation with the onset and progression of liver cirrhosis. To enhance the detection of cirrhosis, we introduce a novel multimodal predictive model termed VQ-MMCM, which leverages multimodal microbiome data. By amassing gut microbiome data and creating a comprehensive training set, we have engineered the VQ-MMCM architecture, grounded in the vector quantized variational autoencoder (VQ-VAE). This model employs a VQ-VAE as the encoder's foundational structure and integrates a trainable weighted sum module to synthesize features, thereby achieving superior representation of microbiome data. Through supervised learning on a dataset of liver cirrhosis cases, augmented by pre-training on unlabeled data, the VQ-MMCM effectively harnesses both the abundance profile and marker profile derived from gut microbiota. This dual-modality approach surpasses the limitations of existing models, such as poor generalization and convergence challenges, and has demonstrated a recognition AUC score of 0.931 for liver cirrhosis.

OBJECTIVE:The pharmacokinetic (PK) parameters estimated from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provide valuable information for clinical research and diagnosis. However, these estimated PK parameters suffer from many sources of variability. Thus, the estimation of the posterior distributions of these PK parameters could provide a way to simultaneously quantify the values and uncertainties of the PK parameters. Our objective is to develop an efficient and flexible method to more closely approximate and estimate the underlying posterior distributions of the PK parameters. METHODS:The normalizing flow model-based parameters distribution estimation neural network (FPDEN) is proposed to adaptively learn and estimate the posterior distributions of the PK parameters. The maximum likelihood estimation (MLE) loss is directly constructed based on the parameter distributions learned by the normalizing flow model, rather than pre-defined distributions. RESULTS:Experimental analysis shows that the proposed method can improve parameter estimation accuracy. Moreover, the uncertainty derived from the parameter distribution constitutes an effective indicator to exclude unreliable parametric results. A successful demonstration is the improved classification performance of the glioma World Health Organization (WHO) grading task, specifically in terms of distinguishing between low and high grades, as well as between Grade III and Grade IV. CONCLUSION:The FPDEN method offers improved accuracy for estimation of PK parameters and boosts the performance of the glioma grading task. SIGNIFICANCE:By enhancing the precision and reliability of DCE-MRI, the proposed method promotes its further applications in clinical practice.

Real-time detection of traffic accidents holds crucial significance for emergency management in Industrial Internet of Things (IIoT) systems. Deep learning-based object detection techniques offer effective means to reduce manual monitoring costs and have shown promising results in accident detection research. However, existing methods often fail to consider learning the overall characteristics of road accidents and exhibit weak feature extraction capabilities, particularly in detecting distant small targets, thereby making it challenging to accurately identify traffic accidents from the perspective of road surveillance. To address these challenges, in this paper, we designed a data collection and preprocessing system based on road surveillance and established a dataset with abundant samples of road accidents for training and validation. Additionally, we propose a real-time, accurate road accident detection method named YOLOSED, which is an improvement over YOLOv8. We introduce the SPD-Conv and SE attention block to enhance the model’s ability to retain more details in feature extraction and improve inter-channel modeling capability. YOLO-SED compensates for the lack of feature extraction for small targets at long distances and enhances the ability to distinguish the overall features of accidents from road scenes in the surveillance view. Experiments on real datasets demonstrate the effectiveness of our approach, with YOLO-SED yielding significant improvements compared to the baseline model.

介绍了浙江大学信息与电子工程学院针对电子信息技术专业人才培养的特点,以金课建设为目标,依据国家级一流线上线下混合课程的评审指标,设计符合专业学位研究生实践类课程要求的教学模式及探索实践.重点讨论"校企共建实训平台结合PBL教学法"的实践类课程群教学模式,在工程训练平台上,培养学生解决复杂工程问题时的自主学习能力、问题分析能力、应用创新能力.再通过课程思政实现授业解惑和传道育人的双重育人目标.该教学模式也为今后专业学位研究生案例教学的深入探索实践提供了经验和借鉴.

Unmanned Aerial Vehicle (UAV) has very wide application prospect in aiding terrestrial cellular network communication, but it remains a challenge to optimize UAV locations and maximize user service rate during deployment. In this paper, a novel network optimization scheme based on anti-flocking model and improved Nash Equilibrium (NE) algorithm is proposed by studying the problem of dynamic UAV deployment and backhaul transmission. Firstly, the UAV-adaptive algorithm based on gray wolf optimization (U-GWO) is used to predeploy UAVs with limited number of UAVs. Secondly, ground mobile users are tracked by building a UAV-based anti-flocking (U-AF) model. Then, during the tracking of ground users by UAV, an improved NE strategy is used to establish the backhaul transmission links between UAVs, ground BSs and other UAVs to ensure that deployed UAVs can maximize the service rate and effective backhaul transmission rate of ground users. Simulation results show that the average service rate of User Equipment (UE) with U-GWO algorithm is improved from 1 % to 5.77 % compared to other different swarm intelligence optimization algorithms. And the service rate obtained with U-AF algorithm is 43.2 % improved compared to the baseline scenario without U-AF algorithm. For UAV backhaul transmission link construction, the simulation results show that the proposed improved NE strategy improves the average effective backhaul transmission rate by 12 %, the minimum backhaul transmission rate by 84 % and the overall iteration number by 5 % on average compared to a pure NE strategy.

A unified hardware architecture was proposed in order to reduce the hardware implementation area and the power of the 2D transform in H.266/VVC. The architecture supported the full-size discrete cosine transform (DCT-II, DCT-VIII) and the discrete sine transform (DST-VII). The architecture consisted of two parallel 1D transform modules and one transpose memory. The 1D transform module was designed based on the multiple constant multiplication (MCM), and a reusable MCM computing unit was designed for all transform types and sizes. The transpose memory was proposed in order to support the pipeline input of the mixed blocks. And the transpose memory was implemented based on static random-access memory (SRAM), used a diagonal storage method with read and write pointers, and used first input first output (FIFO) to cache block information. Experimental results showed that the unified computing unit reduced the area of the transform architecture by 1.3% and the power consumption by 49.5%, and the transpose memory reduced the SRAM storage space by half with the high-frequency zeroing feature of VVC.

自"新工科"建设等提出以来,如何强化学生解决复杂工程问题的能力培养,成为中国高等教育的重要课题.以多学科技术融合软硬件平台为基础,设计了贯穿本科4年的工程教育新模式.平台集成传感、通信、智能识别、机电一体等要素,配合企业项目制引导的课程群建设,形成了由浅入深、逐步高阶的工程能力培养机制.多年来的教学实践显示,该模式能有效提高学生的专业兴趣、创新能力和复杂工程问题的解决能力.

以Cisco NetSpace及"学在浙大"为线上教学平台,以浙江大学—思科网院实验室及浙江大学TP-LINK无线网络实验室为线下教学平台,采用线上智慧学习与线下探究实验相融合的教学模式,将Cisco CCNA7网络技术在线培训用于"无线网络应用"实验课程教学,设计实现了一套线上线下融合教学系统.教学实践表明,该系统能明显提高学生解决实际问题的动手实践能力,教学效果显著,取得了省一流课程等教学成果,促进了浙江大学实验课程线上线下融合教学水平的提升.

OBJECTIVE:Gadolinium-based contrast agents (GBCAs) have been widely used to better visualize disease in brain magnetic resonance imaging (MRI). However, gadolinium deposition within the brain and body has raised safety concerns about the use of GBCAs. Therefore, the development of novel approaches that can decrease or even eliminate GBCA exposure while providing similar contrast information would be of significant use clinically.METHODS:In this work, we present a deep learning based approach for contrast-enhanced T1 synthesis on brain tumor patients. A 3D high-resolution fully convolutional network (FCN), which maintains high resolution information through processing and aggregates multi-scale information in parallel, is designed to map pre-contrast MRI sequences to contrast-enhanced MRI sequences. Specifically, three pre-contrast MRI sequences, T1, T2 and apparent diffusion coefficient map (ADC), are utilized as inputs and the post-contrast T1 sequences are utilized as target output. To alleviate the data imbalance problem between normal tissues and the tumor regions, we introduce a local loss to improve the contribution of the tumor regions, which leads to better enhancement results on tumors.RESULTS:Extensive quantitative and visual assessments are performed, with our proposed model achieving a PSNR of 28.24 dB in the brain and 21.2 dB in tumor regions.CONCLUSION AND SIGNIFICANCE:Our results suggest the potential of substituting GBCAs with synthetic contrast images generated via deep learning.

Real-time monitoring and quantitative measurement of molecular exchange between different microdomains are useful to characterize the local dynamics in porous media and biomedical applications of magnetic resonance. Diffusion exchange spectroscopy (DEXSY) is a noninvasive technique for such measurements. However, its application is largely limited by the involved long acquisition time and complex parameter estimation. In this study, we introduce a physics-guided deep neural network that accelerates DEXSY acquisition in a data-driven manner. The proposed method combines sampling pattern optimization and physical parameter estimation into a unified framework. Comprehensive simulations and experiments based on a two-site exchange system are conducted to demonstrate this new sampling optimization method in terms of accuracy, repeatability, and efficiency. This general framework can be adapted for other molecular exchange magnetic resonance measurements.

“Electronic system design” is an important course closely related to electronic design competition. A wide range of knowledge modules are involved in the class, but some modules are not studied by students. Due to the constraints of classroom time, they need to preview independently before class under the guidance of teachers. We design a recommendation system based on improved collaborative filtering algorithm, so that students can learn by themselves according to the learning resources pushed by teachers. To increase the accuracy of collaborative filtering algorithm, the user's attribute similarity is combined with the traditional collaborative filtering recommendation algorithm to improve the cold start problem and data sparsity of the algorithm. Then we build an experimental teaching platform of "Electronic System Design" curriculum to achieve automatic recommendation of experimental learning resources under teachers’ guidance.

With 3D magnetic resonance imaging (MRI), a tradeoff exists between higher image quality and shorter scan time. One way to solve this problem is to reconstruct high-quality MRI images from undersampled k-space. There have been many recent studies exploring effective k-space undersampling patterns and designing MRI reconstruction methods from undersampled k-space, which are two necessary steps. Most studies separately considered these two steps, although in theory, their performance is dependent on each other. In this study, we propose a joint optimization model, trained end-to-end, to simultaneously optimize the undersampling pattern in the Fourier domain and the reconstruction model in the image domain. A 2D probabilistic undersampling layer was designed to optimize the undersampling pattern and probability distribution in a differentiable manner. A 2D inverse Fourier transform layer was implemented to connect the Fourier domain and the image domain during the forward and back propagation. Finally, we discovered an optimized relationship between the probability distribution of the undersampling pattern and its corresponding sampling rate. Further testing was performed using 3D T1-weighted MR images of the brain from the MICCAI 2013 Grand Challenge on Multi-Atlas Labeling dataset and locally acquired brain 3D T1-weighted MR images of healthy volunteers and contrast-enhanced 3D T1-weighted MR images of high-grade glioma patients. The results showed that the recovered MR images using our 2D probabilistic undersampling pattern (with or without the reconstruction network) significantly outperformed those using the existing start-of-the-art undersampling strategies for both qualitative and quantitative comparison, suggesting the advantages and some extent of the generalization of our proposed method.

Implementing the Internet plus grain action plan and upgrading the level of information industry is one of the important national policies of the advanced countries in the world. To vigorously develop the informatization of grain enterprises and promote the construction of smart grain depots, we need to focus on strengthening the intelligent upgrading of grain depots and the cultivation of scientific and technological talents. At present, most of the existing ecological measurement and control systems of grain depots still use the traditional manual trigger IoT sensing and acquisition mode. The cable array has problems such as single point of failure, insufficient density, and a long time for one measurement. In addition, the heat capacity of the warehouse wall is easy to cause heat accumulation in the outer ring of the grain pile, and the temperature change lags behind. In the IoT course experiment of engineering degree graduate students in colleges and universities, there is a lack of experimental teaching links combined with engineering projects. This paper designs and implements a 5G edge computing grain depot ecological accurate measurement and control system and engineering experimental teaching platform, which can not only solve the technical difficulties to be overcome in modern grain depots, but also realize the teaching of graduate engineering experiment courses. The experimental teaching effect in the past three years shows that 35 sets of engineering experimental teaching software and hardware devices developed have played an important role in graduate course experiment and engineering training.

“IoT information security” is an introductory course in network security. Safety offensive and defense experiments are the means that students can get in contact with network attacks. Students’ network security protection awareness and information protection capabilities can be improved by offensive and defense experiments. At present, there is a problem with high update costs, insufficient comprehensive, complicated operation, and insufficient operation. A new way is presented in this paper to improve the current offensive exercise experiments, providing training platform for postgraduate industrial Internet security courses. It achieves this by using Generative Adversarial Networks (GAN) in honeypot to generate a virtual data for industrial equipment simulation, and using support vector machine (SVM), decision tree and other networks to detect intrusion data. And then a visualization platform is established for students to do experiment.