Deep learning-based segmentation of aneurysmal subarachnoid hemorrhage: toward accurate and scalable prognostic imaging.

Quantification of rice root anatomical traits such as cortical aerenchyma lacunae is key to understanding rice adaptation to diverse water regimes and to support climate-smart breeding. Aerenchyma lacunae contributes to rice internal gas transport and influences methane emissions from flooded systems and can also limit rice water conductivity. It could be an interesting anatomical trait for breeding, however, large-scale anatomical phenotyping remains limited because manual analysis of root cross-sections is labor-intensive, subjective, and difficult to scale across heterogeneous imaging conditions. Existing pipelines often require parameter tuning and do not generalize well across environments. We developed a deep learning pipeline based on a vision transformer architecture to automatically segment rice root cross-sections and quantify cortical aerenchyma lacunae. The model was trained on 1,760 annotated images collected across multiple countries, growth stages, cultivation systems, and experimental contexts, using a collaboratively defined annotation protocol. The final model achieved high segmentation accuracy, with mean intersection over union values exceeding 0.92 for cortical tissues and lacunae. Quantification of the lacuna-to-cortex ratio showed strong agreement with manual annotations, with a coefficient of determination of 0.98 on an independent test set. An independent expert review indicated that model predictions were at least as consistent as manual annotations and reduced large annotation inconsistencies. The pipeline is released as open-source software and includes an interactive online demonstrator, and is accompanied by an online test dataset to support testing and reproducibility. Application across six experimental use cases revealed reproducible differences in aerenchyma lacunae across genotypes, water regimes, environments, and developmental stages. This work provides a robust, scalable, and transferable tool for automated root anatomical phenotyping under heterogeneous experimental conditions. Transformer-based segmentation enables consistent and high-throughput quantification of lacunae, facilitating integration of these anatomical traits into breeding, physiological studies, and climate-smart crop improvement programs.

The Red Queen hypothesis for the maintenance of sexual reproduction proposes that sex should be favored during parasite-host interactions. The presence of sexual reproduction in Escovopsis sensu lato (s.l.) (Hypocreaceae, Ascomycota), obligate specialized parasites on the fungal gardens of fungus-growing ants, has been debated; previous analyses have concluded that sex is likely absent based on Escovopsis s.l. appearing to lack a complete mating-type (MAT) locus, which controls sexual compatibility in fungi. Using 39 previously sequenced genomes, we found that computational annotation of these loci was inconsistent. Through manual annotation, we show that all sequenced Escovopsis s.l. have a complete MAT1 locus. The MAT1 locus is found in the typical genomic context for Hypocreaceae fungi, contains the expected genes (one in the MAT1-2 idiomorph and three in the MAT1-1 idiomorph), are highly conserved at a structural and functional level, and are under strong purifying selection. Using a Phi test, we also find evidence for recombination in one closely related group of samples. Past phylogenomic analyses of Escovopsis s.l. have generated two distinct topologies, and we find that the MAT1 genes also have differing topologies. Further, phylogenetic network analyses show large-scale gene tree discordance between early diverging Escovopsis s.l. taxa and some outgroups, supporting a potential past hybridization event. Taken together, these data suggest that Escovopsis s.l. undergoes cryptic sex, changing our understanding of the ecology and evolution of the model fungus-growing ant symbiosis and opening many exciting research avenues on the dynamics of sex and infection in this system.

Region-Specific Evaluation of Plaque Segmentation in Cross-sectional Projections of Carotid Ultrasound Images Using Deep Learning Models in a Sub-clinical Atherosclerosis Cohort.

Semi-supervised YOLO-DEP for high-resolution X-ray component localization and counting.

Accurate extraction of the left ventricular (LV)-centred region of interest (ROI) in stress perfusion cardiovascular magnetic resonance (CMR) remains challenging due to low signal-to-noise ratio (SNR), motion artefacts, high-dimensional image data, and limited annotated datasets. Efficient ROI localisation is an important preprocessing step for reducing irrelevant anatomical content and improving downstream AI-based analysis. We propose a scalable and annotation-efficient framework for LV-centred ROI localisation and preprocessing in low SNR stress perfusion CMR. A U-Net pretrained on the Multi-Centre, Multi-Vendor, and Multi-Disease (M&Ms) cine CMR dataset was used to generate initial LV localisation proposals on previously unannotated perfusion CMR data acquired at a UK tertiary cardiac centre. Rather than targeting pixel-accurate segmentation, the segmentation outputs were used to identify the LV-centred region, and fit circular ROIs encompassing the LV cavity and surrounding myocardium. A lightweight graphical user interface (GUI) enabled rapid visual assessment, manual annotation and refinement, and temporal propagation across 42-frame sequences. The framework was applied to 798 stress perfusion videos (33,520 frames). A subset of 2882 frames from 69 videos was manually annotated and reviewed using the proposed GUI-assisted framework with confirmation from clinical experts. GUI-assisted annotation required 1-5min per video (mean ≈ 2.5min), corresponding to an estimated 8-12 × reduction in annotation time compared with conventional manual annotation workflows. Fine-tuning on the reviewed annotations improved Dice score from 0.87 to 0.90 compared with training from scratch. The proposed framework enables scalable LV-centred ROI annotation and preprocessing in low-quality stress perfusion CMR.

Visual monitoring of vital parameters in premature infants has become an intensively studied area in recent years. Among these parameters, respiration rate (RR) is one of the most critical vital signs, making non-contact measurement of respiration a key research focus. Many published algorithms achieve improved performance when an appropriate region of interest (ROI) is detected prior to RR estimation. Typically, such ROIs are generated using data-driven segmentation methods. However, modern deep learning-based ROI detection algorithms require thousands of annotated samples for training, and manual data collection and annotation are time-consuming and labor-intensive. In this work, we propose a motion-periodicity-based method to automatically generate respiration-related region masks that capture the abdominal or chest area of neonates. The predicted masks were validated against independent expert annotations, achieving high localization consistency on the torso ([Formula: see text]) and significant overlap with clinical ground truth (mean IoU=0.580 and Dice=0.735). We further show that these automatically produced labels can be directly used to train common segmentation architectures, eliminating the need for manual annotation and enabling the creation of large, high-quality datasets for neonatal respiration analysis. Our findings demonstrate that automatic dataset generation is both feasible and effective for training deep learning-based ROI detectors in this domain.

Accurate tumor segmentation is essential for early diagnosis, treatment planning, and prognostic evaluation. Although manual annotation can achieve high accuracy, it is time-consuming and requires substantial expert involvement. While deep learning has significantly advanced medical image analysis, fully automated methods often fail to segment atypical lesions within complex abdominal anatomy, leading to missed lesions and misclassification of normal tissues, which may compromise clinical decision-making. To address these challenges, we incorporated guidance masks into a convolutional neural network (CNN)-based deep learning framework. Using our Star-Rain software, users place interactive clicks on lesion locations, and the system adaptively generates task-specific guidance masks. This approach directs the model's attention to relevant regions, particularly in atypical or anatomically complex cases. Our method is validated on four independent cohorts comprising 1,217 CT scans from 726 patients, encompassing hepatic, renal, and pancreatic tumors. Across these datasets, our approach outperforms state-of-the-art baseline models on independent test sets, achieving Dice scores consistently above 0.7 and reducing the false negative rate (FNR) by 0.006 to 0.346 compared to the best fully automated approaches. In addition, the model's segmentation outputs effectively support downstream prognosis tasks, highlighting its clinical value. These findings underscore the promise of semi-automatic deep learning frameworks that integrate minimal user input for reliable tumor segmentation. The proposed approach offers a practical and robust solution for clinical applications, enhancing segmentation accuracy and decision support while reducing the annotation burden.

BackgroundAcne is a chronic skin condition that primarily affects adolescents and young adults but can persist into adulthood. It can have repercussions on physical and mental health, self-esteem, and body image. The increasing use of social media for health information and peer support offers an opportunity to explore real-life experiences with acne.ObjectiveThis study aims to analyze social media messages from users in the United States and the United Kingdom using artificial intelligence to assess the impact of acne on quality of life (QoL), identify discussion topics, and explore unmet needs.MethodsThe data were extracted from public platforms using a query containing the word “acne” between January 1 and December 31, 2024. Data cleaning and filtering were performed using natural language processing, machine learning methods, and algorithms. Biterm topic modeling was used to identify the main discussion topics, and QoL impact was assessed using a deep learning algorithm adapted from the EuroQol 5-Dimension Questionnaire or the 36-Item Short Form Health Survey. Unmet needs were identified through manual annotation using the saturation method.ResultsA total of 646,809 messages posted by 432,234 users were identified. The main topics included skincare routines and product recommendations (n=154,907, 23.9%), acne scars (n=135,643, 21%), and general treatment information (n=97,177, 15%). Engagement varied across topics and platforms. On Instagram, dietary and nutritional strategies (0.16%, SD 6.36%) showed the highest mean engagement, followed by skincare routines and product recommendations (0.11%, SD 4.81%). In general, engagement scores were higher in the United Kingdom compared to the United States across all topics. On TikTok, content about makeup and acne had the highest mean engagement score (3.03%, SD 92.65%). Overall, 52.9% (228,613/432,234) of the users expressed at least 1 QoL impact, most frequently related to signs and symptoms (175,604/228,613, 76.8%), social functioning (n=149,234, 65.3%), mental health (n=107,155, 46.9%), and cost (n=62,008, 27.1%). Of 3200 annotated messages, 582 contained unmet needs, including effective solutions for hormonal acne (111/582, 19.1%), clarity in identifying acne triggers (n=84, 14.4%), treatment guidance (n=68, 11.7%), and psychological support (n=68, 11.7%).ConclusionsThis study revealed the significant physical, psychological, social, and financial impact of acne on QoL and identified several unmet needs. Given the growing role of social media, these findings highlight opportunities for dermatologists and health professionals to educate and engage with the acne community through digital platforms.

One of the key objectives in spatial transcriptomics (ST) studies is to map the complex organization and functions of tissues. We introduce GraphScrDom, a reference-informed and weakly supervised contrastive learning model that uniquely integrates expert-provided manual annotations (i.e., scribbles) on spatial grids or histology images with cell type-specific gene expression profiles derived from reference single-cell RNA-seq data to perform tissue segmentation. With only limited scribble annotations, GraphScrDom consistently outperforms existing methods across various ST platforms and at both spot-level and single-cell resolution, as evaluated by six widely used metrics, demonstrating strong generalizability and robustness. Additionally, we have developed an integrative software toolkit that includes an interactive annotation interface and a model training module for spatial domain detection, providing a unified and user-friendly framework to facilitate spatial domain analysis.

Graft-versus-host disease (GVHD) remains a critical complication affecting patients undergoing hematopoietic stem cell transplantation (HSCT), with multifaceted symptoms requiring systematic management. Current clinical practices lack structured tools to guide nurses in addressing GVHD-related symptoms across diverse organ systems. This review aims to synthesize evidence on GVHD symptom management to construct a knowledge graph, enhancing nursing interventions and improving patient outcomes. Relevant authoritative books, guidelines, and expert consensuses such as the "Clinical Nursing Manual for Hematopoietic Stem Cell Transplantation" and the "Chinese Expert Consensus on Diagnosis and Treatment of Acute GVHD After Allogeneic Hematopoietic Stem Cell Transplantation (2024 edition)" were referenced. Literature related to GVHD in HSCT was searched, with a time limit from 2014 to 2024. The knowledge graph was constructed using a combination of manual annotation and computer processing. A GVHD knowledge graph centered on symptom management for HSCT patients was constructed, encompassing 7 aspects of symptom management, including skin, oral cavity, liver, eyes, gastrointestinal tract, lungs, and bones and joints, as well as health knowledge related to nutrition and psychology. This knowledge graph provides a structured, evidence-based framework for GVHD symptom management, addressing gaps in standardized nursing protocols. By integrating multidisciplinary recommendations, it supports clinical decision-making, reduces practice variability, and prioritizes patient-centered care. Future studies should validate its impact on complication rates and quality of life in real-world settings. The tool's adaptability allows updates as new evidence emerges, ensuring alignment with evolving guidelines. For nurses, this resource enhances competency in managing complex GVHD presentations, ultimately improving care delivery for HSCT recipients.

This study proposes a low-resource pattern generation framework for cultural heritage digital innovation, which effectively solves the problems of resource consumption and data scarcity in traditional pattern generation by integrating the LORA (Low-Rank Adaptation) technology and multi-dimensional evaluation system. The study selects the ancient Egyptian traditional tattoos as the target, and by freezing the parameters of the Stable-Diffusion backbone network and optimizing the iteration of the low-rank matrix only, we achieve the reduction of the training memory and the iteration time, which verifies the efficiency of the low-rank adaptation in the image generation task. Aiming at the scarcity and heterogeneity of pattern data, we propose a preprocessing process based on the two dimensions of content-form, and construct a pattern hierarchical dataset, which effectively eliminates the problems of image background interference and style heterogeneity. A complete workflow including LoRA fine-tuning, parameter optimization and multi-dimensional evaluation (shape and color similarity, aesthetics, innovation and application value) is established, and a label optimization framework assisted by manual annotation is developed. The cross-cultural validation experiments show that the method maintains stable performance in the tasks of generating Chinese Qin-Han patterns and cloud-shouldered dress patterns, providing a technical path that combines professionalism and engineering feasibility for the digital inheritance of intangible cultural heritage. In the revised experimental design, five quantitative indicators are additionally adopted, including FID, CLIPScore, pairwise LPIPS diversity, Inception Score, and a CLIP-based aesthetic score. The proposed framework is further quantitatively compared with DreamBooth, Textual Inversion, HyperNetworks, full fine-tuning, ControlNet-based adaptation, and Adapter tuning, complemented by ablation studies and overfitting/diversity analysis.The pipeline reduces trainable parameters by over 99% (~ 860M → ~ 3.2M) and lowers GPU memory from 14.6GB to 5.8GB.

Breast cancer is a heterogeneous disease whose molecular subtypes differ in biological behavior, prognosis, and therapeutic response. This study investigated whether whole-breast radiomic features extracted from digital mammograms and showing statistically significant differences between HER2 + tumors and other molecular subtypes or healthy controls could also provide discriminatory information for exploratory HER2 + characterization. An automated whole-breast segmentation and feature-extraction workflow, without manual lesion-centered delineation, was applied to the breast region to capture broader parenchymal and microenvironmental texture patterns while reducing dependence on manual lesion annotation. Intensity-based, first-order, and second-order texture features were extracted from DICOM mammograms, followed by pairwise statistical testing, false discovery rate correction, effect-size assessment, Gaussian distribution analysis, normalized feature visualization, univariate AUC analysis, and classifier evaluation using logistic regression and linear support vector machines. First-order and intensity-based descriptors showed limited subtype-specific value, whereas second-order texture features provided more informative discriminatory patterns. Among the evaluated feature families, GLCM and NGLDM descriptors showed the most coherent evidence across statistical, visual, and classifier-based analyses, with NGLDM yielding the broadest set of statistically significant features. Classification performance was strongest and most balanced for HER2 + versus healthy controls, while discrimination between HER2 + and other malignant molecular subtypes was modest, context-dependent, and affected by sensitivity-specificity imbalance in several models. Therefore, the present findings more strongly support sensitivity to malignancy-related whole-breast texture alterations than reliable HER2-specific classification among malignant subtypes. Whole-breast mammographic radiomics should be interpreted as an exploratory and complementary source of candidate imaging biomarkers for future validation.

In this study, we describe a deep learning framework for automated seizure annotation in stereo electroencephalography (SEEG) data of patients with focal epilepsy. We use a one-dimensional Variational Autoencoder (VAE) for feature extraction of single-channel temporal series and a linear classifier for segment classification. We trained the network using data from 37 patients containing manual annotations of ictal and Low-Voltage Fast Activity (LVFA) by clinicians. The 1D VAE encodes two-second SEEG segments into a low-dimensional representation in latent space and then classifies them as interictal, ictal, or LVFA segments. We used 5-fold cross-validation for training and validation. Our system classified ictal vs. interictal 2-second segments with an average recall of 0.88. For whole-channel seizure annotation, we compute the Area Under Curve (AUC) of the exponentially smoothed probability signal, marking the onset of both ictal and LVFA with a high average recall of 0.86, and 0.91 on channels identified as the Seizure Onset Zone (SOZ). Markers were temporally accurate, with a mean time lag of 9.8 seconds for ictal onset and 2.0 seconds for LVFA. Latent space analysis suggests dimensions correlate with class-relevant features like amplitude and spectral power. As a secondary objective, we obtain a seizure detection recall of 99% with a specificity of 95%. Our findings suggest that a VAE-based approach can produce a meaningful latent space from SEEG data to detect seizures and fast onset patterns, potentially helping clinicians reduce the workload of SEEG review.

SPELL: A scalable NLP method using regular expressions and large language models for clinical information extraction.

Explainable artificial intelligence (XAI) is increasingly important for computational pathology, where reliable and interpretable predictions are required for clinical use. Pre-trained Vision-Language Models (VLMs) offer a natural pathway to connect visual evidence with textual concepts, but adapting them to the domain of cancer pathology remains challenging due to fine-grained and heterogeneous morphological patterns. Visual prompt learning enables efficient task adaptation with minimal fine-tuning; however, existing prompting techniques face critical limitations: soft prompts often lack clinical specificity, while manually designed hard prompts reduce adaptability. To address these issues, we propose RAPT, a retrieval-augmented and text-guided visual prompting framework for explainable pathology classification. Given an input image, RAPT retrieves semantically related exemplars and leverages class-specific textual descriptions to construct disease-aware prompt tokens, injecting diagnostic cues without manual annotation. An adaptive weighting mechanism attenuates unreliable retrieval evidence, and bridge prompt tokens facilitate the integration of retrieved cues with image representations. Extensive experiments on three public cancer pathology datasets (PatchGastric, BACH, and LC25000) demonstrate that RAPT consistently outperforms prompting baselines across diverse backbone settings. Additional analyses and qualitative case studies demonstrate clinically actionable cues, robustness to imperfect retrieval, and clear failure-mode boundaries for trustworthy pathology decision support, highlighting the potential of robust and explainable prompting for computational pathology.

Identifying patient diagnoses from hospital discharge letters is essential for large-scale cohort selection and epidemiological research, but traditional supervised approaches require extensive manual annotation, which is often impractical for large textual datasets. We present a weakly supervised Natural Language Processing (NLP) pipeline for classifying Italian discharge letters without document-level manual annotation. The method extracts diagnosis-related sentences, generates semantic embeddings using a transformer model further pre-trained on Italian medical documents, and applies a two-level clustering procedure to derive weak labels that are then used to train a document-level classifier. The approach was evaluated in a case study on bronchiolitis using 33,176 discharge letters of children admitted to 44 emergency rooms or hospitals in the Veneto Region, Italy, between 2017 and 2020. The best weakly supervised model achieved an AUROC of 77.68% ([Formula: see text]), an AUPRC of 73.13% ([Formula: see text]), and an F1-score of 78.14% ([Formula: see text]) against manually annotated data. Performance surpassed unsupervised baselines and approached fully supervised models, while reducing the need for manual annotation by more than 1,500 hours for a dataset of this size. Similar model rankings were observed in a secondary validation on a smaller bronchitis dataset (3,188 discharge letters, 2020-2025), where the best weakly supervised model achieved an AUPRC of 76.72% ([Formula: see text]). These results suggest the potential of weakly supervised NLP methods for scalable disease identification from clinical discharge letters.

We present Apollo 3, a new manual genome annotation tool that integrates with the JBrowse 2 genome browser. Its functionality is inspired by existing manual genome annotation tools such as Apollo, Artemis, and Otter, but uses an updated and more scalable architecture and technology stack. It allows the simultaneous editing of multiple genomes, including the visualization of synteny to inform those annotations. Apollo 3 can be used as a standalone annotation editor, or it can be installed on a server and used collaboratively. We describe the application’s design, features, and use cases.

In this study, we present a semi-supervised inspection framework that integrates second harmonic generation (SHG) microscopy with pixel-level anomaly detection (PLAD) for high-sensitivity defect mapping in AlGaN/GaN heterostructures grown on Si (111). SHG is intrinsically sensitive to local inversion symmetry breaking and modulation of the effective second-order nonlinear susceptibility. As a result, spatial variations in the nonlinear optical response serve as a physical contrast mechanism for detecting structural inhomogeneities. The SHG-derived defect features are further supported by correlation with cathodoluminescence (CL) based RDE/NBE defect metrics. Complementary to conventional supervised object-centric detectors such as YOLOv8 and Mask R-CNN, the PLAD approach statistically models nominal lattice patterns without manual annotations. This strategy enables improved sensitivity to weak and spatially diffuse crystalline irregularities. Quantitative evaluation demonstrates substantially improved spatial overlap and detection sensitivity compared with object-based models, resolving more than 5,000 micron-scale defect-affected regions within a single field of view that are largely inaccessible to bounding-box or mask-based architectures. By bridging nonlinear optical microscopy with pixel-level anomaly modeling, the SHG-PLAD framework provides a non-contact strategy for defect-sensitive inspection and offers a useful approach for crystalline defect characterization and quality assessment in AlGaN/GaN heterostructures.

Segmentation of maxillary sinus conditions (MSC) in cone-beam computed tomography (CBCT) images may support preoperative assessment in the posterior maxilla, including implant planning and sinus floor augmentation. Supervised deep learning methods for MSC segmentation typically rely on labor-intensive manual annotation of MSC for network training. This study aimed to develop and evaluate a condition-label-free deep learning framework (SinusNet+) for MSC segmentation in CBCT images, in which network training does not require manual MSC annotations and instead relies on synthetic conditions generated within the normal maxillary sinus (MS). To generate synthetic MSC in normal MS, a synthetic condition generator was introduced to simulate MSC within the normal MS by varying texture, shape, and noise, thereby approximating a range of radiographic appearances of MSC in CBCT images. SinusNet+ achieved an average Dice similarity coefficient of 0.820±0.110, precision of 0.878±0.061, and recall of 0.777±0.145, respectively. The proposed method outperformed unsupervised baselines and achieved segmentation performance comparable to that of the supervised approach. The proposed framework demonstrates the feasibility of condition-label-free segmentation of MSC in CBCT images, while still requiring anatomical annotation of the normal MS during dataset preparation.