Industrial anomaly detection aims to identify surface defects (e.g., in texture and color) and structural defects in products. A significant obstacle in this field is the necessity for large, high-quality training datasets. While zero-shot and few-shot learning methods have emerged to address this need by operating with minimal labeled data, they are often plagued by slow inference speeds that hinder real-time industrial deployment. To bridge this gap between data efficiency and speed, we propose a novel zero-shot approach that integrates initial data screening with diffusion-based data augmentation. Our framework begins with a similarity-based screening algorithm to curate a set of normal samples directly from the overall data distribution. A latent diffusion model then learns the patch-level distribution of these samples to generate a robustly augmented dataset, which is used to construct a comprehensive memory bank for anomaly detection. Evaluated on the MVTec AD dataset, our method achieves state-of-the-art performance, with average AUROC scores of 96.87
A cure of cancer remains challenging. With the advancements in gene editing techniques, immune cell-based treatments (e.g., natural killer (NK) cell therapy) have emerged as potent therapeutic modalities, aiming to address clinical bottlenecks associated with traditional approaches. This review recapitulates the inherent characteristics of NK cells as a safe allogeneic tool against cancers, as well as available gene-editing systems for enhancing NK-cell activity, improving persistence, and increasing safety. Furthermore, it summarizes the preclinical and clinical practices of genetically engineered NK cell therapies and highlights their potential, challenges, and future perspectives in cancer treatment. Overall, we provide a comprehensive insight into NK cell-based immunotherapies as a promising approach for cancer treatment.
Continuum robots (CRs) have shown significant clinical potential in minimally invasive surgery (MIS) due to their inherent compliance and high dexterity. However, achieving independently decoupled control and long-distance, stable transmission within strictly confined anatomical spaces remains a key challenge for proximal advancing mechanisms. This study presents a systematic review of advancing mechanisms for CRs and their representative applications. Advancing mechanisms are categorized into coupled-motion and decoupled-feed configurations, where the former translates the entire drive module to provide high kinematic rigidity and motion accuracy, and the latter employs friction-wheel or spooling mechanisms with decoupled transmission to enable compact architectures and extended stroke capabilities. Application-driven design strategies are analyzed across diverse surgical scenarios, highlighting how configuration selection and topology optimization can be tailored to address anatomical and environmental constraints. Key challenges, including transmission inaccuracies from nonlinear friction, sterility and modularity constraints, and embodied integration, are discussed. Future directions focus on high-fidelity force feedback, modular quick-release architectures, and integrated embodied sensing-actuation units to advance the next generation of continuum robotic systems.
To address the degradation of centroid extraction accuracy in star trackers under asymmetric point spread function (PSF) conditions, this study proposes an analytical centroid localization algorithm based on the energy distribution of a skewed Gaussian model. The method employs a skew-normal Gaussian function to characterize the asymmetric PSF and introduces an aspect ratio parameter to describe the variation in energy distribution. On this basis, a closed-form analytical expression for the centroid offset is derived, enabling fast, efficient high-precision centroid determination. Simulation and real-image experiments are conducted under various representative scenarios, including different PSF morphologies, noise levels, and sub-pixel centroid positions. Experimental results demonstrate that, in typical asymmetric PSF conditions, the proposed algorithm exhibits superior overall performance in terms of accuracy, computational efficiency, and robustness.
Deep learning has become indispensable for medical image analysis and computer-aided diagnosis. In real-world scenarios, the limited sample size available within a single institution often necessitates integrating multi-site datasets for model training. However, poor performance in multi-site learning attributed to non-identically distributed data stems primarily from two key factors: cross-site data heterogeneity and site-level volume imbalance. To address these challenges, a novel multi-site learning framework, AugNormNet, is proposed. A site-specific batch channel normalization module is designed to perform site-adaptive normalization across the batch and channel dimensions, aligning cross-site feature distributions while preserving site-specific characteristics. In addition, a feature-level contrastive enhancement loss is introduced, which conducts semantic-direction feature enhancement and contrastive learning to explicitly increase diversity of samples from small-scale sites and mitigate the model’s bias toward data-rich centers. Evaluation on four publicly available COVID-19 CT datasets from different clinical centers demonstrates that the proposed method achieves absolute accuracy improvements of 2.42, 1.88, 0.85, and 1.41 percentage points over state-of-the-art multi-site learning methods. All paired t-tests yield p-values less than 0.05, confirming statistical significance. Experimental results demonstrate that the proposed method consistently achieves higher classification accuracy across all sites, while comprehensive ablation studies and in-depth analyses further verify the effectiveness of each component, highlighting its practical applicability to real-world multi-site medical image classification.
General practice is the cornerstone of primary healthcare, but it faces increasing challenges from workforce shortages and the rising burden of chronic diseases with complex comorbidities. The integration of medicine and engineering offers new opportunities to transform service delivery and enhance the sustainability of primary care systems. This narrative review aims to synthesize recent practices and emerging research on medicine–engineering integration in general practice and proposes strategic directions for building a people-centered intelligent health ecosystem. A comprehensive search was conducted across PubMed, Web of Science, IEEE Xplore, Scopus, and the Cochrane Library using relevant keywords. Sixty-four studies were included in the final review. It was summarized key applications of engineering technologies in general practice, including disease screening, clinical decision support, chronic disease management, healthcare accessibility, and professional education. Major barriers to integration are also analyzed across data infrastructure, clinical workflows, workforce capacity, and ethical and regulatory domains. Technologies such as wearable devices, artificial intelligence, telemedicine platforms, and immersive educational tools have demonstrated substantial potential to enhance diagnostic accuracy, improve chronic disease outcomes, expand access to care, and strengthen workforce training. However, persistent challenges (including data silos, algorithmic bias, limited real-world usability, interdisciplinary talent shortages, and ethical concerns) continue to hinder large-scale implementation. Moving beyond fragmented pilot initiatives towards a coordinated, ecosystem-level approach is essential for the effective integration of intelligent health technologies into general practice. A people-centered intelligent health ecosystem, supported by interdisciplinary collaboration and robust governance, can strengthen primary healthcare delivery and contribute to the achievement of universal health coverage.
Surgical navigation technology provides surgeons with precise guidance through real-time tracking of the spatial pose of surgical instruments and serves as a crucial technical foundation for modern precision medicine. As a visual reference marker system, AprilTag demonstrates the advantages of low cost and flexible deployment for surgical instrument localization. However, traditional geometric constraint-based pose estimation methods face challenges such as cumulative detection errors and environmental interference sensitivity in complex surgical environments, making it difficult to meet the requirements of high-precision surgical navigation. This study proposes a novel AprilTag pose tracking method that integrates geometric and deep learning approaches. By constructing a deep residual learning network, the method learns the influence patterns of image features on pose errors to achieve intelligent correction of traditional geometric methods. The approach employs ResNet18 to extract deep image features, encodes the AprilTag’s pose information, and utilizes a dual-branch fusion network to learn the residual error between geometric predictions and actual instrument pose. A weighted loss function is designed to address rotation and translation errors. Experimental results on real surgical instrument datasets show that compared to traditional geometric methods, this approach reduces the average position error by 48.44
This study presents a computational and experimental framework for assessing magnetic resonance imaging (MRI)-induced radiofrequency (RF) heating risks for a transcranial puncture needle at 3 T MRI. Due to the one-dimensional structure of the puncture needle, as well as its clinical interventional use, the transfer function method was used to evaluate the needle’s electromagnetic model. Multi-variable computational simulations encompassing over 100 000 tests were performed using five anatomical virtual human models, including various coil configurations and clinical implantation trajectories. The in vivo RF heating was predicted by convolving the simulated tangential electric field along each path with the measured transfer function, scaled by an experimentally validated factor. The maximum temperature rise of the puncture needle in the human body does not exceed 6 °C after a 15-min scan under a given power limit exposure under MRI. This integrated approach provides a robust safety assessment for transcranial MRI-guided intervention.
To investigate the prosthesis prediction accuracy of an artificial intelligence (AI)-assisted preoperative planning system, this study compared standard anteroposterior (AP) pelvic radiographs with long-leg radiographs (LLRs) acquired by a robotic advanced X-ray (RAX) system (Multitom Rax). A total of 29 patients (29 hips) undergoing primary total hip arthroplasty (THA) were retrospectively analyzed, excluding those with severe structural anomalies such as Crowe type IV developmental dysplasia of the hip. Preoperative planning was performed using the “Cuantian” AI system across four groups: AP pelvic radiographs with and without calibration markers, and RAX-acquired LLRs with and without calibration markers. The actual implanted component sizes recorded in intraoperative logs served as the standard. The results demonstrated that for AP pelvic radiographs, the inclusion of calibration markers was associated with improved prediction accuracy, achieving a 100
Cyclopeptide compounds have emerged as a significant source of drug candidates in pharmaceutical development due to their unique cyclic structures, enhanced metabolic stability, and exceptional biological activities. Ulm16, a penicillin-binding protein-type thioesterase, promiscuously catalyzes the macrocyclization of linear peptides of different sizes between the N- and C-terminal residues with L- and D-configurations, respectively However, the mechanism governing its selectivity towards substrates with various lengths remains unclear. This study comparatively investigated the structural differences in the interaction modes between Ulm16 and its linear substrates — hexapeptide (WLA-B1) and octapeptide (SGM) — by integrating molecular docking, molecular dynamics simulations, and rational design approaches. Findings revealed that the cooperative interactions with the oxygen anion cavity (Ser71/Thr299) and Arg431 in the active pocket stabilize WLA-B1 at both the C- and N-termini, which is crucial for its efficient macrocyclization. Conversely, SGM exhibits impaired pre-reaction state proportion due to the absence of key hydrogen-bonding networks. Based on structural analysis, we designed a series of mutants (A225S, G226E, D147V, S144F, and L300G) to optimize substrate binding by enhancing the negative charge density within the active pocket. Among them, A225S variant demonstrated a slight improvement in the yield of cyclized WLA-B1 (1.3-fold higher than the wild type). This study established a computational methodology spanning from substrate loading to catalytic macrocyclization, elucidated the molecular mechanism of substrate selectivity with Ulm16, and provided crucial theoretical foundations for rationally designing highly efficient cyclic peptide biosynthetic enzymes.
To enhance the reasoning stability of small-parameter medical large language models in internal medicine question-answering tasks, this paper proposes a training methodology based on explicit chain-of-thought (CoT) modeling and hybrid supervised fine-tuning (SFT). First, a hierarchical dataset comprising general internal medicine instructions and explicit CoT data was constructed. On this basis, a two-stage hybrid SFT process was implemented, incorporating direct preference optimization to align the model with clinical preferences. Experimental results demonstrate that the proposed method improves the accuracy on Chinese medical benchmarks while reducing the proportion of redundant reasoning, effectively enhancing the logical rigor of complex clinical inquiries. Furthermore, these findings validate the potential of this approach for deploying low-cost, highly reliable, and localized auxiliary diagnostic systems in privacy-sensitive and compute-constrained clinical scenarios.
Large segmental bone defects have long posed a significant clinical challenge in orthopedics, primarily due to limited osteogenic potential and extended healing durations. Metallic biomaterials, such as tantalum, have been extensively employed in bone regeneration due to their superior mechanical strength and favorable early-stage osseointegration. However, their inherent bioinertness frequently results in complications, including implant subsidence and aseptic loosening over prolonged periods, ultimately compromising the repair outcome. To address these challenges, we developed a porous tantalum–biodegradable polymer composite scaffold with a triply periodic minimal surface (TPMS) architecture, aiming to systematically elucidate how structural design modulates polymer degradation kinetics and regulates the biological behavior of bone marrow mesenchymal stem cells (BMSCs). Through evaluating composite hydrogels composed of lithium magnesium silicate (LAP), gelatin (GE), and sodium alginate (SA) at various concentrations, we identified the 20 g/L LAP formulation as injectable and capable of forming a well-organized porous architecture upon freeze-drying. Subsequently, three distinct TPMS-based porous tantalum scaffolds, i.e., Diamond (D)-type, Gyroid (G)-type, and fused D–G (FDG) hybrid, were fabricated, and the optimized GE–SA–LAP polymer hydrogel was infused into each structure. The results demonstrated that the pore architecture played a decisive role in regulating polymer degradation dynamics: the D-type scaffold exhibited slower degradation (35.2
When capturing images through glass surfaces, it is inevitable to include reflections in acquired images. Such reflections not only degrade the aesthetic quality of the captured image but also adversely impact the performance of subsequent computer vision tasks. Residual reflections and artifacts still remain in the results, even though existing reflection removal algorithms are able to remove part of the reflection interference. Motivated by the observation that reflected objects present in input image are absent in corresponding ground-truth (reflection-free) image, this study proposes a reflection location awareness (RLA) approach for single image reflection removal. The proposed RLA method first detects reflection regions in the input image, generating an explicit representation of reflection locations. The representation is then leveraged as spatial prior information and fed into subsequent network modules, along with the input image, to recover texture details of background transmission layer. Extensive qualitative and quantitative experiments conducted on multiple benchmark datasets validate the effectiveness of the proposed method.
Image quality and diagnostic performance between a fast spin-echo sequence with standard parameters (FSE-SD) and a deep learning-reconstructed fast spin echo with low parameters (FSE-DL) are compared for temporomandibular joint (TMJ) magnetic resonance imaging (MRI) at 1.5-T and 3-T scanners. This study enrolled 183 patients (94 scanned at 1.5-T and 89 at 3-T) who underwent both FSE-SD and FSE-DL acquisitions for temporomandibular disorders. Scan time, subjective image quality, detection rates of pathological features, and the inter-protocol and inter-reader agreement were assessed by two readers. The Wilcoxon signed-rank test, McNemar test, and unweighted/linearly weighted Cohen’s κ statistics were used. Using the deep learning algorithm, total scan time at 1.5-T was reduced from 316 s (FSE-SD) to 211 s (FSE-DL), representing a 33.23
To investigate and analyze the application techniques and principles of 3D printing-based medical–engineering interaction for precision surgery of the foot and ankle, we summarized 216 patients treated at the Foot and Ankle Surgery Team of the Department of Orthopaedic Surgery, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, between February 2020 and March 2025, who underwent 3D printing-assisted interventions under medical–engineering interaction. Among them, 152 patients underwent pre-operative 3D reconstruction based solely on CT/MRI imaging, 32 patients underwent deformity localization and quantitative measurements performed in the regions of interest based on these reconstruction, and 26 patients received personalized 3D-printed surgical guides for osteotomy, lesion localization, or broken implant removal. Additionally, 3 patients received customized fixation and rehabilitation orthoses, and 3 patients received personalized 3D-printed prostheses. Using 3D reconstruction, deformity localization and quantitative measurements, personalized 3D-printed surgical guides, and personalized 3D-printed prostheses, individualized treatment strategies were formulated for selected foot and ankle patients, accounting for 17.2
Osteoporotic fractures pose a growing public health burden in aging populations with chronic comorbidities. Current fracture risk tools (e.g., FRAX and QFracture) remain limited by static assessment frameworks and inadequate racial adaptation. Artificial intelligence (AI), particularly multimodal deep learning and temporal modeling, demonstrates superior predictive accuracy in elderly patients with osteoporosis by integrating dynamic physiological, genetic, and clinical variables. This review critically evaluates the clinical applicability of existing prediction tools and synthesizes advances in AI-driven modeling, highlighting four transformative frontiers: multimodal data fusion (imaging, genomics, and real-world monitoring), temporal analysis, meta-learning for cross-population generalization, and interpretable AI for clinical transparency.
With the accelerating pace of population aging, the number of older stroke patients continues to rise. Lower-limb motor impairments severely compromise gait stability and quality of life. Conventional rigid exoskeletons are often bulky and provide insufficient comfort and compliance, making it difficult to meet the safety and lightweight requirements of older users. To address these limitations, this study establishes mechanical models of Bowden-cable friction and elasticity, and on this basis proposes a dual-loop assistance strategy that integrates admittance control with PID control to achieve compliant compensation of gait deviations at the knee and ankle joints. Wearable experiments involving two older post-stroke participants demonstrate that the proposed system significantly improves knee flexion and markedly reduces the peak inversion angle in the participant with severe foot inversion, while also decreasing joint angle fluctuations and enhancing postural stability. These findings support the feasibility and application potential of the proposed exoskeleton-assisted strategy for gait rehabilitation in older stroke populations.
Electroencephalography (EEG) foundation models are increasingly used as general-purpose backbones for brain-computer interfaces (BCIs) by leveraging large-scale pretraining and task-specific adaptation. This review summarizes recent progress in EEG foundation models from three perspectives: datasets and task coverage, with emphasis on how generalization goals are operationalized by split protocols and concrete evaluation procedures; model design choices, including input construction and tokenization, masked pretraining objectives, and Transformer backbones for spatiotemporal modeling across heterogeneous channel layouts; and downstream adaptation, comparing linear probing, full fine-tuning, and parameter-efficient tuning, while clarifying the conditions under which each setting is most informative. We emphasize that reported gains are often protocol-dependent, as differences in task scope, preprocessing, training budget, and baseline selection can substantially affect comparability and the extent to which conclusions generalize. Finally, we outline future directions for EEG foundation models in BCI, focusing on standardized evaluation infrastructure, EEG-tailored modeling choices, and deployment-aware adaptation under real-world constraints.
This study investigates passive multi-target indoor localization in reconfigurable intelligent surface (RIS)-assisted wireless local area network (WLAN) systems. Unlike conventional methods that rely on the receiver’s spatial resolution, we propose a novel formulation that directly exploits the spatial resolution of the RIS for improved localization accuracy. To tackle the challenge of high-dimensional joint parameter estimation, we develop a low-complexity off-grid sparse representation model by applying geometric approximations and reformulating the problem into a coarse, non-uniform framework. This transformation enables efficient and accurate joint localization by fully leveraging signal sparsity. Moreover, we design a variational Bayesian algorithm for joint sparse signal recovery, which simultaneously estimates all target positions and incorporates a grid refinement strategy to enhance accuracy. Numerical results demonstrate the effectiveness and robustness of the proposed approach.
Dynamic cardiac magnetic resonance imaging is a robust non-invasive imaging technique capable of quantitatively evaluating cardiovascular function. Nevertheless, its clinical application is hindered by the relatively slow data acquisition. Accelerated imaging strategies that rely on acquiring undersampled k-space data offer a practical approach to mitigate this drawback, yet producing high-quality reconstructions with limited measurements remains difficult. To address this, we propose a multi-scale sparse learning network (MSSL-Net) built upon an iterative unfolding framework. Each iterative reconstruction unit of the network integrates a multi-scale attention fusion module, a deep sparse module, and a gradient update module, enabling the full exploitation of multi-scale spatial features and sparse priors of the image. Extensive experimental results on two public cardiac MRI datasets demonstrate that MSSL-Net clearly outperforms the comparison methods under various undersampling patterns and acceleration factors.
