Spontaneous Echo Contrast (SEC), characterized by smoke-like dynamic echogenicity in trans esophageal echocardiography (TEE), serves as a critical imaging biomarker for assessing thromboembolic risk. However, current diagnostic approaches depend on manual interpretation or regression-based algorithms, which are labor-intensive, prone to inconsistency, and limited in generalizability. To address these challenges, we propose SECNet, an end-to-end deep learning framework for auto mated SEC detection in TEE videos. SECNet incorporates two key components: (1) a region-of-interest segmentation module that accurately localizes SEC-prone cardiac structures while suppressing surrounding noise; and (2) a spa tiotemporal feature extraction module that captures both temporal dynamics and spatial distribution of SEC. These modules are linked through a novel differentiable mask guided bridging mechanism that enables joint optimization of localization and representation learning. Trained on a clinical dataset of 1,106 annotated TEE videos, SECNet achieves 93.7% accuracy and a 90.0% F1-score, outper forming the compared methods by 3.1–23.3 percentage points in accuracy. An auxiliary experiment on EchoNet Dynamic further demonstrates the applicability of the architecture to a related echocardiographic video-analysis task. This study supports standardized and reproducible SEC assessment to assist thromboembolic risk evaluation in clinical workflows.
The 12-lead electrocardiogram (ECG) is the standard for diagnosing myocardial infarction (MI), but its use in pre-hospital and resource-limited settings is limited by hardware complexity and electrode placement requirements. This study addresses a biomedical and health informatics challenge by investigating MI screening from minimal-lead ECG suitable for mobile and telemedicine applications. We propose LeadMix-MST, a lead-mixing multi scale spatio-temporal architecture for reduced-lead ECG analysis. Using three 12-lead ECG datasets, we conducted patient-level cross-dataset evaluations across devices and populations. The predefined lead set was an important design factor, with I–aVL–V5 and I–II–V3 performing most consistently. On these leads, the architecture remained competitive with larger baseline models at under 1Mparameters. A model trained on the Shandong Provincial Hospital dataset and evaluated on the Physikalisch-Technische Bun desanstalt dataset achieved an F1-score of 0.933 ± 0.003 and an AUROC of 0.937 ± 0.004 using a 3-lead configuration (I–aVL–V5), exceeding the full 12-lead setting on this dataset. The proposed framework maintains a favor able balance between diagnostic performance, parameter efficiency, and practical deployability, primarily by reducing electrode placement burden.
With the advent of Industry 5.0, increasing attention has been directed toward workers' mental well-being, particularly work-related stress. Given the complexity of dynamic work environments, achieving reliable monitoring while minimizing the impact on workers remains challenging. To address this issue, we propose a multimodal stress monitoring system that integrates a wearable polymer optical fiber sensor with a thermal camera to acquire respiratory, cardiac, and regional facial-temperature information. A graph neural network (GNN) was developed to model modality-specific features and their cross-modal relationships. Multimodal recordings from 20 participants were evaluated under blocked five-fold within-subject and strict leave-one-subject-out (LOSO) protocols. The proposed GNN achieved accuracies of 0.7233, 0.9475, and 0.9225 for within-subject three-class classification, relaxation versus low stress, and relaxation versus high stress, respectively. Under strict LOSO evaluation, the corresponding accuracies were 0.5550, 0.7925, and 0.7375. Compared with conventional machine-learning and temporal deep-learning baselines, the proposed GNN achieved the best performance in five of the six evaluation settings. The experimental and ablation results demonstrate that the proposed framework effectively exploits complementary cardiorespiratory and facial thermal information and supports continuous stress monitoring in complex industrial and healthcare settings.
Lung cancer is the leading cause of cancer death, with CT central to screening, prognosis, and treatment. Manual segmentation is variable and time consuming, while deep learning (DL) offers automation but faces adoption barriers. Guided by the Knowledge-to-Action framework, this study develops a clinician-guided and clinician-evaluated pipeline to improve reproducibility, prediction, and clinical trust. Multi-center lung-cancer CT from 999 patients across 12 public datasets were analyzed. Five DL models (3D Attention U-Net, ResUNet, VNet, ReconNet, SAM-Med3D) were benchmarked against expert contours on whole and cropped images. Segmentation reproducibility was tested with 497 PySERA-extracted radiomics features using Spearman correlation, ICC, Wilcoxon tests, and MANOVA. Prognostic modeling compared supervised (SL) and semi-supervised learning (SSL) with 38 dimensionality reduction strategies and 24 classifiers. Six physicians qualitatively evaluated masks across 21 questions spanning seven domains, including clinical meaningfulness, boundary quality, survival prediction, trust, and workflow integration. VNet-based segmentations achieved the highest performance (Dice 0.83±0.07, IoU 0.71±0.09), strong radiomics reproducibility (average Spearman correlation coefficient 0.76, ICC 0.65), and strong predictive SSL performance (accuracy 0.88±0.003, F1 = 0.83±0.004). SSL generally outperformed SL across models. Descriptive physician ratings indicated favorable boundary quality and peritumoral representation for VNet, although Friedman tests showed no statistically significant differences in selected physician-rated domains. Importantly, experts preferred AI to provide initial masks for refinement rather than replace physicians. This study shows that VNet, integrated with SSL, can deliver accurate, reproducible, and potentially clinically useful tools for CT based lung cancer prognosis. Embedding AI as a supportive assistant within physician-centered workflows offers a feasible path to clinical translation.
Stain normalization is a fundamental preprocessing technique in digital and computational pathology, and it is important for improving the consistency of clinical diagnosis by reducing color heterogeneity. However, existing stain normalization methods are often limited by content-stain entanglement and an ambiguously modeled content-stain relationship, resulting in unstable structural preservation, cross-center adaptation burden, and sensitivity to template selection. To address these limitations, we propose a Vector-Quantized Content-Stain Decoupling Network (VQ-CSDN) that decouples structure-defining content from stain appearance and learns a center-specific content-to-stain mapping, enabling structure-preserving, template-free stain transfer from unseen source centers without source-specific retraining. VQ-CSDN employs two independent encoders to extract content and stain features respectively and a generator for reconstruction. To better characterize cross-center content feature spaces and capture staining styles, we introduce vector quantization with a global content codebook shared across centers and center-specific stain codebooks. Moreover, a content-to-stain module directly infers stain features from the content feature and a target staining style, enabling end-to-end stain transfer without template selection. To further promote structural fidelity and enhance staining style separability, we propose center-specific gradient-driven discriminators that jointly assess center-level stain conformity and the morphological consistency of generated images. Extensive experiments on multiple public and private pathology datasets show that VQ-CSDN consistently enhances normalization quality, reflected by increased SSIM/PSNR and decreased FID, and further improves downstream benign-malignant classification performance (AUC 0.9738) over state-of-the-art methods.
