biomarker extraction

In low-resource clinical settings, bedside monitors often lack network connectivity, creating an interoperability bottleneck that hinders integration of physiological data into electronic health record (EHR) systems. To address this, we propose a hardware-modification-free, computer vision–based digitization framework. Our method employs a hierarchical detection architecture: YOLOv11 for robust detection of both the monitor device and vital-sign display regions; a geometric perspective correction module to handle variable viewing angles and illumination conditions; and PaddleOCR for high-accuracy optical character recognition. Evaluated on 6,498 real-world clinical images, our approach achieves a monitor detection mAP@50–95 of 99.5%, a vital-sign region localization accuracy of 91.5%, and an end-to-end core parameter extraction accuracy exceeding 98.9%. To the best of our knowledge, this is the first lightweight, scalable, and hardware-agnostic solution for automated screen-based data extraction from bedside monitors—effectively bridging clinical information silos in resource-constrained environments.

This study addresses the need for automated analysis and report generation from medical imaging modalities (CT, MRI, X-ray, ultrasound). We propose a zero-shot vision-language multimodal framework leveraging Google Gemini 2.5 Flash’s integrated visual encoding and language understanding capabilities. To enhance robustness in anomaly localization, we introduce Gaussian probabilistic modeling and coordinate validation. Furthermore, multi-level visual explanations and structured text extraction enable interpretable, clinically meaningful report generation; an interactive platform is implemented using Gradio. Our key contributions are: (1) zero-shot cross-modal tumor detection, drastically reducing reliance on annotated data; (2) precise abnormality localization with only 80-pixel error, achieving both high accuracy and strong clinical interpretability; and (3) end-to-end support for heterogeneous imaging modalities and standardized radiology report generation, significantly advancing automation in radiological workflows.

Existing cerebral artery feature extraction methods—whether manual or automatic—suffer from high user dependency, insufficient validation, and limited quantification capability for subtle vascular changes. To address these limitations, we propose ArteryX, a semi-supervised framework integrating a vessel-tracking network with an in silico quantitative validation mechanism. The vessel-tracking component, implemented in MATLAB, performs keypoint localization and graph-structured modeling to mitigate vessel discontinuities and erroneous connections. The validation module employs a synthetic vascular phantom with predefined ground truth, enabling standardized, objective performance benchmarking. Evaluated on TOF-MRA data from patients with cerebral small vessel disease, ArteryX processes each case in 10–15 minutes with minimal manual intervention. It demonstrates significantly higher sensitivity to subtle vascular alterations compared to state-of-the-art semi-automatic approaches, underscoring its potential for clinical translation.

High-quality, diverse annotations for arteriole–venule (A/V) segmentation in fundus images are scarce, limiting model generalizability and clinical reliability. Method: We introduce the first publicly available, high-resolution (1024×1024) color fundus image dataset, capturing real-world clinical variability across imaging devices, age groups, and acquisition conditions. We propose an explicit connectivity-validated hierarchical A/V annotation paradigm—enabling, for the first time in a public fundus dataset, topology-aware vascular consistency verification. Low-quality images discarded by conventional quality control—but retaining clinical relevance—are intentionally retained to enhance robustness. We further develop a customized multi-level annotation interface, a connected-component–based visualization tool for manual correction, and a multimodal preprocessing pipeline (including raw images, contrast-enhanced images, and RGB-encoded connectivity-validation masks). Contribution/Results: This dataset significantly improves the robustness and clinical applicability of vascular analysis models in real-world deployment scenarios.

This study addresses the challenge of automating and scaling quantitative morphological analysis of cerebral vasculature to enable population-level investigations of vascular health and aging. We propose the first fully automated, multi-scale cerebral vascular analysis framework: it employs atlas-guided centerline extraction and skeletonization to generate topologically consistent vascular graph representations; subsequently, it systematically computes 15 quantitative features—spanning morphological, topological, fractal, and geometric domains—across hierarchical scales from whole-brain networks to arterial subregions. Validated on 570 high-quality cerebral vascular datasets, the features demonstrate excellent test–retest reliability (ICC > 0.9) and robustly capture age-, sex-, and education-associated gradients in vascular complexity. The framework establishes a reproducible, standardized computational foundation for large-scale cerebrovascular phenomics studies.

This study addresses the challenges of data scarcity and insufficient model trustworthiness in automated pneumothorax diagnosis in dogs by proposing a novel approach that integrates generative segmentation with first-principles statistical analysis. The method employs a vision-language model to guide iterative flow-matching optimization for pixel-level lesion segmentation, followed by spectral anomaly detection based on random matrix theory to identify eigenvalues significantly deviating from normal distributions as indicators of pathology. The work introduces the first publicly available canine pneumothorax dataset with pixel-level annotations, achieving high boundary-precision segmentation and high-sensitivity anomaly detection. This framework not only ensures diagnostic accuracy but also provides interpretable evidence grounded in statistical principles.

This study addresses the challenge of effectively detecting pre-stroke neurovascular instability (NVI) using conventional unimodal wearable devices. To this end, we propose Melaguard, a lightweight multimodal Transformer model that establishes the first edge-computing-oriented physiological signal fusion framework. Melaguard integrates heart rate variability, peripheral perfusion index, blood oxygen saturation, and bilateral phase coherence to generate a composite NVI score. Built upon a 1.2M-parameter Transformer-lite architecture with four-head self-attention, the model enables real-time inference on resource-constrained edge devices such as Cortex-M4 microcontrollers, achieving a worst-case execution time of ≤4 ms. Evaluated on the PhysioNet CVES cohort, Melaguard attains an AUC of 0.755, significantly outperforming LSTM, SVM, and random forest baselines. Notably, PPG-derived metrics exhibit strong correlation with ECG gold standards (r > 0.69), and their morphological features achieve an AUC of 0.923 for cerebrovascular disease classification on an independent dataset, offering an efficient and viable solution for community-level stroke risk screening.

To address the critical bottleneck that millions of scanned paper-based electrocardiogram (ECG) recordings remain inaccessible to AI-based diagnosis, this work proposes the first robust, fully automated digitization framework for real-world clinical paper ECGs. Methodologically, it integrates four sequential stages: image preprocessing, adaptive grid detection, noise-resilient waveform tracing, and geometric correction-based signal reconstruction—effectively suppressing artifacts from noise, paper creases, stains, and perspective distortion. Our key contributions include an open-source, modular system designed for both clinical deployment and research reproducibility. Evaluated on 37,191 real-world ECG images, the framework achieves a mean signal-to-noise ratio of 19.65 dB on the Akershus dataset and consistently outperforms state-of-the-art methods across all metrics on the Emory dataset. This work establishes a high-fidelity digital signal foundation essential for scalable, AI-driven ECG interpretation.

This study addresses the limitations of conventional electrocardiogram (ECG) screening, which relies on manual interpretation and often fails to detect subtle or progressive morphological changes, leading to missed early abnormalities. The authors propose an interpretable, motif-based framework that extracts representative cardiac cycles within a fixed window using dynamic time warping (DTW) and aligns them with fiducial points for morphological visualization. They introduce three novel metrics of morphological drift: deviation from normal sinus rhythm, divergence from an individual’s personalized baseline, and motif instability index, enabling longitudinal monitoring and early warning. Evaluated on the MIT-BIH and PTB-XL datasets, the method significantly discriminates between normal and abnormal ECGs (p<0.01 and p<1e-4, respectively), achieving a Cliff’s delta as high as 0.93, thereby demonstrating strong clinical potential and validity.

This study addresses the challenges in automatic report generation from multi-view coronary X-ray angiography, where lesion localization and stenosis grading suffer from projection-dependent inconsistencies and a lack of supervisable cross-view anatomical alignment. To overcome these limitations, the authors propose a geometry-driven multi-view fusion framework that leverages synthetically generated angiograms with controllable geometry to provide plaque-level correspondence supervision. A dedicated anatomical correspondence module explicitly aligns features from auxiliary views into the primary view’s coordinate space before feature aggregation. This approach achieves, for the first time, explicit cross-view anatomical alignment, circumventing the unsupervised nature of real clinical data. Experiments demonstrate significantly improved alignment consistency on synthetic data and successful zero-shot transfer to real angiograms, with structured stenosis reporting performance substantially outperforming both single-view and conventional multi-view methods.