Breast cancer exhibits significant spatiotemporal heterogeneity. Traditional radiomics approaches usually rely on low-temporal-resolution imaging and discrete image phases, failing to capture the rapid and continuous kinetic evolution within the tumor microenvironment. Although ultrafast dynamic contrast-enhanced magnetic resonance imaging captures precise contrast agent permeation, current methodologies lack systematic, dedicated computational frameworks to extract these high-dimensional dynamic features. To bridge this gap, this study developed and validated a robust two-stage spatiotemporal radiomics framework for high-dimensional feature selection and predictive modeling. First, a novel gradient-dynamics-based feature selection algorithm was designed to robustly identify highly discriminative kinetic trajectory patterns and overcome the dimensionality curse of time-series data. Second, a lightweight Transformer-based predictive network incorporating a self-attention knowledge distillation mechanism was deployed to capture deep representational knowledge and enhance the differentiation between benign and malignant lesions. Experimental results demonstrated that the proposed framework outperformed traditional baseline models using delta features or pharmacokinetic parameters, achieving an area under the curve of 0.959 ± 0.022 and an accuracy of 92
Novel view synthesis, which is essential for three-dimensional (3D) reconstruction, traditionally requires dense image sets and pre-calibrated camera parameters. However, in industrial digital twin applications spatial constraints often limit data acquisition to sparse views that fail to provide the feature correspondences necessary for accurate modeling. To overcome this bottleneck, a semantic-guided 3D Gaussian splatting (3DGS) framework tailored to sparse-view industrial reconstruction was introduced. Rather than relying on precise initialization and dense inputs, the proposed method couples explicit 3D Gaussian representations with the dense and unconstrained stereo 3D reconstruction (DUSt3R) end-to-end pose estimation model. The severe occlusions, typical of factory environments, are handled by incorporating Segment Anything Model 2 (SAM2) to hierarchically decompose the scene, yielding a structured representation that improves optimization stability. A probability density field-driven algorithm is subsequently applied to extract lightweight meshes directly from the optimized Gaussian point clouds. Evaluations on the MipNeRF360 benchmark and a custom industrial dataset demonstrated substantial improvements in the structural similarity index measure. By enabling robust reconstruction from limited viewpoints, this pipeline offers a practical geometric foundation for automated inspection and remote equipment monitoring.
Abstract Large language models (LLMs) have been rapidly adopted in healthcare since 2022; however, field-level trends and specialty differences remain poorly characterized. This study aims to map the research landscape of LLMs in healthcare and generate comparative specialty profiles through a bibliometric analysis using the Web of Science Core Collection (2015–2025), including English-language articles and reviews analyzed for publication and citation trends, leading countries, institutions, authors, journals, co-citation networks, and keyword structures with VOSviewer and CiteSpace, with sub-analyses for medicine, general and internal; surgery; and radiology, nuclear medicine and medical imaging. Of the 2226 articles included, the annual output increased from four publications in 2022 to 1327 in 2025, yielding a compound annual growth rate (CAGR) of 592.26%; the United States of America contributed 43.5% of the publications, followed by China (14.3%), Germany (9.2%), Turkey (9.0%), and England (6.8%); the publications for each specialty were surgery (31.7%), medicine, general and internal (24.2%), and radiology, nuclear medicine and medical Iimaging (13.0%), with the last demonstrating the fastest 2023–2024 growth (CAGR 265%); and the collaboration networks were United States of America-centered, with dense trans-Atlantic ties. LLM research in healthcare is expanding rapidly with distinct specialty-specific trajectories, and interpreting these trajectories using the DECIDE-AI (developmental and exploratory clinical Iinvestigations of decision support systems driven by artificial intelligence) framework clarifies low-risk near-term applications and monitoring priorities, providing a specialty-aware baseline to support the safe, equitable, and regulation-aligned adoption of LLMs in clinical practice.
Abstract This study focuses on tertiary lymphoid structure (TLS) semantic segmentation in whole slide images (WSIs). Unlike TLS binary segmentation, TLS semantic segmentation identifies boundaries and maturity and requires the integration of contextual information to discover discriminative features. Owing to the extensive scale of WSI (e.g., 100,000 $$\times$$ 100,000 pixels), TLS segmentation is typically performed using a patch-based strategy. However, this prevents the model from accessing information outside the patches, thereby limiting its performance. To address this issue, GCUNet, a graph neural network-based contextual learning network for TLS semantic segmentation, is proposed. Given an image patch (target) to be segmented, GCUNet first progressively aggregates the long-range and fine-grained contexts outside the target. Subsequently, a detail and context fusion block (DCFusion) was designed to integrate the context and details of the target to predict the segmentation mask. This study builds four TLS semantic segmentation datasets: TCGA-COAD, TCGA-LUSC, TCGA-BLCA, and PUMCH-PAAD. The first three, comprising 826 WSIs and 15,276 TLSs, will be made publicly available to promote TLS semantic segmentation. Experiments on these datasets demonstrate that GCUNet consistently improves the mean F1-score (mF1) performance compared with existing state-of-the-art methods, with an observed mF1 improvement of at least 7.41% over the best-performing baseline. These results highlight the potential of GCUNet for accurate TLS assessment and facilitate the development of computational pathology-based immune microenvironment analysis.
Text-to-motion generation aims to synthesize semantically consistent and naturally coherent motion sequences from natural language descriptions. Given the continuous nature of human motion, diffusion models operating in a continuous latent space offer inherent advantages over vector quantization-based methods, particularly in avoiding quantization errors and in modeling quality. However, existing diffusion models primarily rely on mean squared error loss. This stepwise regression paradigm often leads to ‘over-smoothed’ motion sequences and struggles to capture the subtle semantic nuances embedded in textual descriptions. To realize the potential for continuous diffusion generation, an enhanced latent-space diffusion framework designed to elevate generation capabilities across two dimensions, namely, distribution approximation and semantic alignment, is proposed. Specifically, a latent-space adversarial discriminator is incorporated. By applying decoupled adversarial supervision, this component mitigates the detail loss caused by mean regression, significantly enhancing the physical realism and dynamic sharpness. Concurrently, a latent-space contrastive alignment strategy is introduced during the denoising process that reinforces the correspondence of the generated motion sequences with the given textual inputs via explicit cross-modal constraints. Extensive experiments on standard benchmarks demonstrate that the proposed method effectively addresses the limitations of conventional diffusion models, thus validating the potential of continuous diffusion frameworks within the domain of text-driven motion synthesis.
The intersection between mental maps and visualization has attracted increasing research attention in recent years, as they exhibit a natural functional complementarity. Mental maps research explores how humans construct in-mind representations of the external world through the view of cognitive science, while visualization can transform implicit mental maps into concrete visual forms, thereby facilitating the analysis of their construction mechanisms. Visualization, a branch of computer science, investigates how to design human-friendly visual representations of digital data. Mental maps can provide human-perspective insights to guide visualization designs and improve the effectiveness of data understanding and analysis. This paper presents a literature survey of the recent research at this intersection. Regarding the visualizations of mental maps, the present study classifies relevant research into visualizations to depict structural topographies, attribute information in mental maps, and compare multiple mental maps. In terms of mental maps for visualization, this study categorizes the relevant research into the element structuring, structural mapping, and consistency maintenance mechanisms of mental maps to guide visualization designs. This paper introduces and discusses in detail representative works in these categories, providing future research opportunities at this intersection.
Deformable image registration is a critical task in medical image analysis. However, existing learning-based methods encounter difficulties in achieving high registration accuracy under large anatomical deformations while maintaining model interpretability. To overcome these limitations, a correlation reconstruction and refinement network (CRR-Net) was introduced, representing the first framework to incorporate super-resolution reconstruction at the feature level for deformable registration. The core innovation lies in the correlation reconstruction and refinement module (CRRM), which enables precise modeling of spatial correspondences by leveraging high-resolution feature spaces. This design captures richer structural details and contextual cues while expanding the receptive field during resolution recovery, thereby improving performance in large-deformation scenarios. Integrated within a pyramid registration framework, the CRRM supports a multi-scale coarse-to-fine strategy based on local correlation modeling, ensuring consistent deformation-field prediction across scales. Model interpretability was further enhanced through a hierarchical visualization of the deformation fields, providing an intuitive quality assessment. Extensive experiments on brain and cardiac datasets demonstrated that CRR-Net outperforms state-of-the-art deformable registration approaches. For example, it achieved comparable performance on the LPBA40 dataset while using 32% fewer parameters and running 31% faster than CorrMLP, a representative high-performance method. This code is publicly available at https://github.com/miracledrumstick/CRR-Net .
Coronary heart disease (CAD) is the leading cause of death worldwide, and coronary angiography (CAG) serves as the gold standard for its assessment. Valvular heart diseases, such as severe aortic stenosis (AS) and severe mitral regurgitation (MR), frequently coexist with CAD yet are often underdiagnosed. Opportunistic screening for these conditions at the time of CAG could influence therapeutic strategies and improve prognosis. This study developed and validated a foundation model for the automated screening of severe AS and severe MR from CAG videos. The study presents CAGFound, a video-based foundation model that was self-supervised pre-trained on CAG sequences from seven medical centers and subsequently adapted to two downstream tasks: screening for severe AS and severe MR. Two internal and external validation datasets were retrospectively enrolled from the First Medical Center and the Sixth Medical Center of Chinese PLA General Hospital, respectively. A total of 117,383 unlabeled CAG sequences were used to build CAGFound. For the detection of severe AS, CAGFound achieved an area under the receiver operating characteristic curve (AUROC) of 0.932 (sensitivity 0.767, specificity 0.921) on the internal test dataset and maintained robust performance on the external validation dataset, with an AUROC of 0.879 (sensitivity 0.800, specificity 0.955). For the detection of severe MR, the model demonstrated an AUROC of 0.933 (sensitivity 0.738, specificity 0.938) on the internal dataset and an AUROC of 0.896 (sensitivity 0.754, specificity 0.855) on the external cohort. The performance of CAGFound was also compared with other video-based foundation models, VideoMAEv2 and Video Swin. CAGFound achieved the highest AUROC and demonstrated the best calibration performance (Brier score 0.122, R2 0.478) compared with VideoMAEv2 (Brier score 0.159, R2 0.306) and Video Swin (Brier score 0.162, R2 0.306). CAGFound enables accurate, automated screening for severe AS and severe MR during CAG. It has the potential to increase detection rates, facilitate timely clinical referral, and improve prognosis without requiring additional contrast administration or procedures.
Few-shot medical image classification is a highly challenging problem in computer-aided diagnosis, with the central difficulty being enabling deep models to learn discriminative features conducive to classification from limited labeled samples. Vision transformers (ViTs) have recently demonstrated outstanding performance across various visual tasks. However, owing to their large parameter counts and dependence on massive pretraining data, ViTs are prone to overfitting in sample-scarce scenarios typical of few-shot learning. Parameter-efficient fine-tuning (PEFT) techniques, such as low-rank adaptation (LoRA), have alleviated some of these issues. However, conventional PEFT approaches still encounter difficulties in complex medical image classification tasks. To address this, this study proposes a general fine-tuning framework called a hierarchical probing and fusion network (HPF-Net), which integrates three core innovations to allow smarter and more efficient adaptation for few-shot medical image classification. First, a Fisher information-driven layer selection strategy strengthens the layer-selection robustness in few-shot settings. Subsequently, the attention-guided multiscale fusion module aligns and improves the features drawn from the selected critical layers. Subsequently, LoRA is incorporated into this efficient fine-tuning pipeline to reduce the parameter overhead while improving the accuracy. Extensive experiments on the public few-shot medical image benchmark, the medical imaging meta-dataset, demonstrated that HPF-Net significantly outperformed baseline methods, and ablation studies validated the necessity of each proposed component. The source code will be released upon acceptance.
Slide-based lectures remain the primary means by which undergraduate students learn about the mathematical, physical, and systems-level foundations of medical imaging. However, despite their central educational role, no openly available dataset pairs imaging lecture slides with clean, well-aligned explanatory narration suitable for scientific and educational research. The authors introduced MEDI-SLATE: medical imaging slide-lecture aligned teaching ensemble, constructed from a complete undergraduate biomedical engineering medical imaging course. The dataset contains 1117 high-resolution slides paired with refined narration derived from classroom audio through automatic speech recognition, followed by careful manual cleanup. MEDI-SLATE encompasses linear systems, Fourier analysis, signal processing, X-ray physics, computed tomography, positron emission tomography/single photon emission computed tomography, magnetic resonance imaging , ultrasound, and optical imaging. In addition to the slide-text pairs, the dataset includes lecture-level difficulty tags, key ideas, common student misunderstandings, and practice questions sourced directly from the instructor’s materials. A fully reproducible preprocessing pipeline covering slide extraction, narration refinement, alignment, and corpus-level analyses is provided. MEDI-SLATE offers a high-fidelity, openly available resource for medical imaging education, curriculum development, multimodal learning research, and creation of artificial intelligence-assisted instructional tools, with all data and codes released for transparent use and future extension.
The construction of complex volumetric parametric models has long been a bottleneck in achieving integrated design and simulation modeling. To enhance model quality and simplify the modeling process, this paper proposes an innovative method for improving the continuity of complex volumetric parametric models. First, depending on whether the input consists of design parameters or surface models, a G^0/C^0 -continuous volumetric parametric model is generated using either a creation or recreation approach. A new data structure is also developed to store the control-point indices and their topological relationships. Next, based on the actual connections among different patches in the volumetric models generated using the two aforementioned methods, the continuity conditions for three different scenarios are formulated. For each scenario, the corresponding systems of equations are established to determine the control points. Subsequently, an algorithm is developed to automatically sort, store, and adjust the relevant control points so that the volumetric parametric model satisfies the G1 continuity condition. The generated examples demonstrate that the volumetric parametric modeling method proposed in this study is effective for constructing complex models, significantly improving the model quality and rendering them more suitable for subsequent analysis and processing.
In the evolving nexus of neuroscience and robotics, the symbiotic fusion of electroencephalography (EEG) and electromyography (EMG) is emerging as a paradigm-shifting avenue for enhancing human-machine interfaces. While EEG, which captures the subtle electrical nuances of the brain, offers a potent channel for nuanced brain-machine communication, EMG serves as a bridge, converting neuromuscular intentions into actionable directives for robotic apparatuses. This review highlights the current methodologies in which EEG and EMG not only function in silos but also converge harmoniously to dictate robotic control. By delving deeper into this, the intricate synergy between cognitive processes, muscular responses, and machine actions can be unraveled. Subsequently, the discourse also navigates through the myriad challenges encountered in realizing real-time, seamless integration of these bio-signals with robotics and the innovative solutions poised to address them. The aim is to provide a comprehensive understanding of the interplay between neuroscience and robotics. This insight will help drive breakthroughs in adaptive human-machine collaboration.
Medical imaging technology has advanced rapidly in recent years; however, abnormalities in medical images are often rare and complex, making sample labels difficult to obtain for supervised learning of detection models. Existing unsupervised anomaly detection methods, which are the mainstream approaches, often struggle with issues such as blurred edges and varying scales of abnormal regions. To address these issues, a novel unsupervised method for medical image anomaly detection is proposed: unsupervised medical image anomaly detection based on edge guidance and multi-scale flow fusion (UMIAD-EGMF). This method excavates rich edge information with scale adaptation and progressively identifies discriminative information for anomaly detection. UMIAD-EGMF captures contextual information around anomaly boundaries via low-level feature fusion (enhancing boundary details with the edge guidance module; EGM), integrates EGM-extracted edge information into deeper features using the edge aggregation module, and merges multi-scale feature maps to capture common anomaly features (subtle and significant) through multi-scale flow fusion. Experiments on breast ultrasound images (BUSI), brain magnetic resonance imaging (brain MRI), and head computed tomography (head CT) datasets demonstrate that UMIAD-EGMF outperforms the state-of-the-art methods. Specifically, on the BUSI dataset, the segmentation area under the precision-recall curve for object localization (AUPRO) of UMIAD-EGMF reaches 63.36%, surpassing that of the multi-scale low-level feature enhancement U-Net (MLFEU-net) by 0.01%; on the brain MRI dataset, its segmentation AUPRO is 90.83%, outperforming that of MLFEU-net by 0.33%; and on the head CT dataset, its segmentation AUPRO is 62.24%, exceeding that of MedMAE by 2.37%.
Computed tomography (CT) is a powerful imaging modality widely used in medicine, research, and industry for noninvasive visualization of internal structures. However, conventional CT systems rely on X-rays, which involve radiation exposure, high equipment costs, and complex regulatory requirements, making them unsuitable for educational or low-resource settings. To address these limitations, we developed a compact, low-cost, optically emulated CT scanner that uses visible light to image semi-transparent specimens. The system consists of a rotating stage enclosed within a light-isolated box, backlight illumination, and a fixed digital single-lens reflex camera. A Teensy 2.0 microcontroller regulates the rotation of the stage, while MATLAB is used to process the captured images using the inverse Radon transform and visualize the reconstructed volume using the Volumetric 3D MATLAB toolbox. Experimental results using a lemon slice demonstrate that the scanner can resolve internal features such as the peel, pulp, and seeds in both 2D and 3D renderings. This system offers a safe and affordable platform for demonstrating CT principles, with potential applications in education, industrial inspection, and visual computing.
Efficient and accurate assignment of journal submissions to suitable associate editors (AEs) is critical in maintaining review quality and timeliness, particularly in high-volume, rapidly evolving fields such as medical imaging. This study investigates the feasibility of leveraging large language models for AE-paper matching in IEEE Transactions on Medical Imaging. An AE database was curated from historical AE assignments and AE-authored publications, and extracted six key textual components from each paper title, four categories of structured keywords, and abstracts. ModernBERT was employed locally to generate high-dimensional semantic embeddings, which were then reduced using principal component analysis (PCA) for efficient similarity computation. Keyword similarity, derived from structured domain-specific metadata, and textual similarity from ModernBERT embeddings were combined to rank the candidate AEs. Experiments on internal (historical assignments) and external (AE Publications) test sets showed that keyword similarity is the dominant contributor to matching performance. Contrarily, textual similarity offers complementary gains, particularly when PCA is applied. Ablation studies confirmed that structured keywords alone provide strong matching accuracy, with titles offering additional benefits and abstracts offering minimal improvements. The proposed approach offers a practical, interpretable, and scalable tool for editorial workflows, reduces manual workload, and supports high-quality peer reviews.
The limited availability of annotated data presents a major challenge in applying deep learning methods to medical image analysis. Few-shot learning methods aim to recognize new classes from only a few labeled examples. These methods are typically investigated within a standard few-shot learning paradigm, in which all classes in a task are new. However, medical applications, such as pathology classification from chest X-rays, often require learning new classes while simultaneously leveraging the knowledge of previously known ones, a scenario more closely aligned with generalized few-shot classification. Despite its practical relevance, few-shot learning has rarely been investigated in this context. This study presents MetaChest, a large-scale dataset of 479,215 chest X-rays collected from four public databases. It includes a meta-set partition specifically designed for standard few-shot classification, as well as an algorithm for generating multi-label episodes. Extensive experiments were conducted to evaluate both the standard transfer learning (TL) approach and an extension of ProtoNet across a wide range of few-shot multi-label classification tasks. The results indicate that increasing the number of classes per episode and the number of training examples per class improves the classification performance. Notably, the TL approach consistently outperformed the ProtoNet extension, even though it was not specifically tailored for few-shot learning. Furthermore, higher-resolution images improved the accuracy at the cost of additional computation, whereas efficient model architectures achieved performances comparable to larger models with significantly reduced resource requirements.
Photoacoustic imaging (PAI), a modality that combines the high contrast of optical imaging with the deep penetration of ultrasound, is rapidly transitioning from preclinical research to clinical practice. However, its widespread clinical adoption faces challenges such as the inherent trade-off between penetration depth and spatial resolution, along with the demand for faster imaging speeds. This review comprehensively examines the fundamental principles of PAI, focusing on three primary implementations: photoacoustic computed tomography, photoacoustic microscopy, and photoacoustic endoscopy. It critically analyzes their respective advantages and limitations to provide insights into practical applications. The discussion then extends to recent advancements in image reconstruction and artifact suppression, where both conventional and deep learning (DL)-based approaches have been highlighted for their role in enhancing image quality and streamlining workflows. Furthermore, this work explores progress in quantitative PAI, particularly its ability to precisely measure hemoglobin concentration, oxygen saturation, and other physiological biomarkers. Finally, this review outlines emerging trends and future directions, underscoring the transformative potential of DL in shaping the clinical evolution of PAI.
Text-to-3D scene generation is pivotal for digital content creation; however, existing methods often struggle with global consistency across views. We present 3DS-Gen, a modular “generate-then-reconstruct” framework that first produces a temporally coherent multi-view video prior and then reconstructs consistent 3D scenes using sparse geometry estimation and Gaussian optimization. A cascaded variational autoencoder (2D for spatial compression and 3D for temporal compression) provides a compact and coherent latent sequence that facilitates robust reconstruction. An adaptive density threshold improves detailed allocation in the Gaussian stage under a fixed computational budget. While explicit meshes can be extracted from the optimized representation when needed, our claims emphasize multiview consistency and reconstructability; the mesh quality depends on the video prior and the chosen explicitification backend. 3DS-Gen runs on a single GPU and yields coherent scene reconstructions across diverse prompts, thereby providing a practical bridge between text and 3D content creation.
Effective survival analysis is essential for identifying optimal preventive treatments within smart healthcare systems and leveraging digital health advancements; however, existing prediction models face limitations, primarily relying on ensemble classification techniques with suboptimal performance in both target detection and predictive accuracy. To address these gaps, this paper proposes a multimodal framework that integrates enhanced facial feature detection and temporal predictive modeling. For facial feature extraction, this study developed a lightweight face-region convolutional neural network (FRegNet) specialized in detecting key facial components, such as eyes and lips in clinical patients that incorporates a residual backbone (Rstem) to enhance feature representation and a facial path aggregated feature pyramid network for multi-resolution feature fusion; comparative experiments reveal that FRegNet outperforms state-of-the-art target detection algorithms, achieving average precision (AP) of 0.922, average recall of 0.933, mean average precision (mAP) of 0.987, and precision of 0.98–significantly surpassing other mask region-based convolutional neural networks (RCNN) variants, such as mask RCNN-ResNeXt with AP of 0.789 and mAP of 0.957. Based on the extracted facial features and clinical physiological indicators, this study proposes an enhanced temporal encoding-decoding (ETED) model that integrates an adaptive attention mechanism and a gated weighting mechanism to improve predictive performance, with comparative results demonstrating that the ETED variant incorporating facial features (ETEncoding-Decoding-Face) outperforms traditional models, achieving an accuracy of 0.916, precision of 0.850, recall of 0.895, F1 of 0.884, and area under the curve (AUC) of 0.947–outperforming gradient boosting with an accuracy of 0.922, but AUC of 0.669, and other classifiers in comprehensive metrics. The results confirm that the multimodal dataset (facial features + physiological indicators) significantly enhances the prediction accuracy of the seven-day survival conditions of patients. Correlation analysis reveals that chronic health evaluation and mean arterial pressure are positively correlated with survival, while temperature, Glasgow Coma Scale, and fibrinogen are negatively correlated.
Epilepsy is a chronic neurological disorder characterized by recurrent seizures that can lead to death. Seizure treatment usually involves antiepileptic drugs and sometimes surgery, but patients with drug-resistant epilepsy often remain effectively untreated owing to the lack of targeted therapies. The development of a reliable technique for detecting and predicting epileptic seizures could significantly impact clinical treatment protocols and the care of patients with epilepsy. Over the years, researchers have developed various computational techniques using scalp electroencephalography (EEG), intracranial EEG, and other neuroimaging modalities, evolving from traditional signal processing methods (e.g., wavelet transforms and template matching) to advanced machine learning (ML, e.g., support vector machines and random forests) and deep learning (DL) algorithms (e.g., convolutional neural networks, recurrent neural networks, transformers, graph neural networks, and hybrid architectures). This review provides a detailed examination of epileptic seizure detection and prediction, covering the key aspects of signal processing, ML algorithms, and DL techniques applied to brainwave signals. We systematically categorized the techniques, analyzed key research trends, and identified critical challenges (e.g., data scarcity, model generalizability, and real-time processing). By highlighting the gaps in the literature, this review serves as a valuable resource for researchers and offers insights into future directions for improving the accuracy, interpretability, and clinical applicability of EEG-based seizure detection systems.
