Artificial intelligence models are increasingly used for stroke risk evaluation and clinical decision-making. However, the scarcity of expert masks, low contrast, and noise in ultrasound images affect segmentation performance. Our study innovatively integrated an adversarial PatchGAN discriminator into a batch-normalized semi-supervised U-Net generator to enhance carotid plaque segmentation in ultrasound images with limited expert annotation. The discriminator provides patch-level localized feedback to enhance boundary delineation and structural consistency of the predicted masks. The model was trained using a hybrid loss function and self-training strategy. The performance of our framework was compared with a fine-tuned semi-supervised U-Net model. Both frameworks were trained under identical experimental settings using 30% of the labeled data; the remaining unlabeled images were utilized to generate pseudo-labels for data augmentation. A confidence threshold of 0.7 was applied to filter unreliable predictions. We conducted experiments on a dataset of 970 internal carotid artery ultrasound images from the Imperial College of London, UK. The experimental results showed that the proposed framework achieved a Dice coefficient of 86.12 and a Jaccard index of 75.66, outperforming the baseline U-Net (84.22 and 73.12, respectively). The integration of patch-level adversarial feedback into a semi-supervised framework enhanced the segmentation accuracy and improved the reliability of pseudo-labels under limited supervision. The results were comparable to those of state-of-the-art deep learning models, which require a large pool of labeled data.
A backdoor attack poisons a victim model during training, causing it to predict attacker-specified target label during inference. This has emerged as a critical threat to AI security. To counter this threat, backdoor purification seeks to eliminate or mitigate backdoor effect without compromising clean accuracy. Pruning-based methods alter the model architecture and degrade clean accuracy, while fine-tuning methods rarely explore backdoor-related cues, leaving residual backdoor effects. In this paper, we propose a dual-pathway mask ranking guided selective fine-tuning method for backdoor purification, synthesizing the perspectives of pruning and fine-tuning. Since clean-poisoned separation forms the data prerequisite for backdoor purification, we first develop an erasure-based intervention strategy grounded in our finding that backdoor-related triggers are prioritized for spatial reconstruction. This enables direct trigger erasure, making poisoned data more susceptible to erasure-based intervention than clean data. Building on this, we establish a forget-then-recover mechanism that characterizes the degree of backdoor contamination for each neuron using soft-valued masks. We design a training-free dual-pathway mask ranking module to categorize neurons into distinct types based on the rankings of soft-valued masks. This type information guides the selection of fine-tuning policies (i.e., relearning, unlearning or nolearning) for each neuron. Experimental results across four benchmark datasets (MNIST-M, SVHN, CIFAR-10, and CIFAR-100) demonstrate that our proposed method outperforms seven baseline competitors under various backdoor attacks. Our method attains the lowest average ASR (Attack Success Rate) ranging from 0.23% to 1.29%, and incurs the smallest average CA (Clean Accuracy) degradation ranging from 0.15% to 1.53%. Ablation studies further validate the effectiveness of our clean-poisoned separation strategy and selective fine-tuning policy.
Human casualties are among the most severe consequences of human-tiger conflict (HTC), and recent evidence from Nepal indicates an increasing trend in tiger-related incidents. This decade-long study (2014/15-2023/24 A.D.) examined the status, trends, associated factors, conflict patterns, and mitigation preferences related to HTC in and around Chitwan and Parsa National Parks, including surrounding forests of Parsa, Bara, Makawanpur, Chitwan, and Nawalpur districts. Primary data were collected through semi-structured interviews with all reported victims or household representatives. Analyses included descriptive statistics, Chi-squared tests, correlation analysis, Friedman ANOVA, and hotspot mapping using MS Excel, SPSS, ArcGIS, and Python. A total of 80 tiger attacks were recorded, with fatalities and injuries occurring in nearly equal proportions. Most cases occurred in the Chitwan National Park buffer zone, and males comprised more than three-fourths of victims, largely reflecting their involvement in high-risk forest-related activities. Recorded casualties increased during the study period and showed a positive association with tiger population estimates; however, this relationship should be interpreted as associative rather than causal because multiple ecological and human-use factors may also contribute. Hotspot analysis identified Budhirapti BZUC, Madi Valley, and adjoining areas as major conflict zones. Incidents were more frequent during the monsoon and spring seasons, and solitary individuals were more vulnerable than those in groups. Most attacks involved older, injured, or sub-adult tigers. Awareness and training were ranked as the most preferred mitigation measures, and local communities generally maintained positive attitudes toward tiger conservation despite increasing conflict risks. The findings highlight the need for targeted, context-specific mitigation strategies, including awareness and training, strengthened physical barriers, management of problem tigers, and further testing of rear-face masks to support safer human-tiger coexistence.
Ultrasound imaging through sonolucent cranial implants is an emerging modality for post-neurosurgical monitoring of the adult brain, but quantitative interpretation remains challenging due to speckle, attenuation, shadowing, and the difficulty of consistently delineating thin anatomical landmarks. We present a deep learning system developed at Longeviti Neuro Solutions for segmenting key intracranial structures the ipsilateral and contralateral lateral ventricles and the cranial midline-in coronal-plane adult cranial ultrasound images from patients with Longeviti ClearFit® Acoustic Brain Interface (ABI)TM implants. The dataset comprises 457 proprietary, de-identified ultrasound frames with known pixel spacing, annotated in CVAT with ventricle and midline labels. We benchmark multiple encoder-decoder segmentation architectures and address severe class imbalance using class-weighted optimization with Dice and midline-focused focal-Tversky terms, followed by horizontal-flip test-time averaging. The best-performing configuration achieved a foreground macro Dice of 0.856 on a held-out test set, with Dice values of 0.926, 0.921, and 0.720 for the contralateral ventricle, ipsilateral ventricle, and midline, respectively. Finally, predicted masks are converted into geometry-based metrology overlays by estimating maximal perpendicular ventricle spans and ventricle-to-midline distances. These outputs provide standardized, millimeter-calibrated measurement visualizations for downstream review and future clinical validation.
Medical image segmentation using deep learning (DL) has enabled the development of automated analysis pipelines for large-scale population studies. However, state-of-the-art DL methods are prone to hallucinations, which can result in anatomically implausible segmentations. With manual correction impractical at scale, automated quality control (QC) techniques have to address the challenge. While promising, existing QC methods are organ-specific, limiting their generalizability and usability beyond their original intended task. To overcome this limitation, we propose no-new Quality Control (nnQC), a robust QC framework based on a diffusion-generative paradigm that self-adapts to any input organ dataset. Central to nnQC is a novel Team of Experts (ToE) architecture, where two specialized experts independently encode 3D spatial awareness, represented by the relative spatial position of an axial slice, and anatomical information derived from visual features from the original image. A weighted conditional module dynamically combines the pair of independent embeddings, or opinions to condition the sampling mechanism within a diffusion process, enabling the generation of a spatially aware pseudo-ground truth for predicting QC scores. Within its framework, nnQC integrates fingerprint adaptation to ensure adaptability across organs, datasets, and imaging modalities. We evaluated nnQC on seven organs using fifteen publicly available datasets. Our results demonstrate that nnQC consistently outperforms state-of-the-art methods across all experiments, including cases where segmentation masks are highly degraded or completely missing, confirming its versatility and effectiveness across different organs.
To develop and validate a multimodal deep learning model based on a CT perfusion (CTP)-defined infarct core for predicting parenchymal hematoma type 2 (PH2) after endovascular therapy (EVT) in patients with acute ischemic stroke (AIS). In this dual-center retrospective study, 487 patients with anterior-circulation large-vessel-occlusion AIS who underwent EVT were included, with 219 assigned to the training cohort and 268 to an independent external validation cohort. The infarct core was defined using CTP and transferred to non-contrast CT (NCCT) to generate three-dimensional image patches. A score-based baseline model (Model.Score) was constructed using logistic regression with ASPECTS as the predictor. A pure imaging model based on a 3D DenseNet architecture (Model.Core) was developed using NCCT images and infarct-core masks as dual-channel inputs. In addition, a conventional clinical-imaging model (Model.Clinical-Imaging) was constructed using logistic regression based on ASPECTS and baseline clinical variables. A multimodal fusion model (Model.Fusion) was subsequently developed by integrating infarct-core imaging features with baseline clinical variables and dual-phase CT angiography (CTA)-derived scores. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration analysis, decision curve analysis, and statistical comparison with DeLong, net reclassification improvement, and integrated discrimination improvement tests. Model interpretability was evaluated using gradient-weighted class activation mapping and permutation importance analysis. Model.Fusion achieved the best predictive performance, with AUCs of 0.906 in the training cohort and 0.845 in the external validation cohort, exceeding those of the score-based model, the conventional clinical-imaging model, and Model.Core. The fusion model also showed better calibration and greater clinical net benefit. Interpretability analyses demonstrated predominant model attention to low-density regions within the infarct core, while dual-phase CTA-derived variables and baseline NIHSS score were identified as major predictors. CTP-defined infarct core-based deep learning enables effective prediction of PH2 after EVT in AIS. Integration of imaging, clinical, and dual-phase CTA features further improves predictive performance and may facilitate individualized risk stratification before treatment.
Accurate weed detection from unmanned aerial vehicle (UAV) images remains difficult because weed targets often blend with surrounding vegetation, vary substantially in scale, and exhibit irregular boundaries under cluttered grassland backgrounds. This study presents a UAV grassland weed detection framework that combines intelligent annotation with task-oriented detector adaptation. For label generation, SAM produces instance-level vegetation masks, SigLIP performs few-shot semantic matching, and a vision-language model audits ambiguous candidates using few-shot weed references, local candidate patches, and full-image context. Candidate outputs are converted into YOLO-format labels for detector training. For detection, a YOLOv11-based model is adapted with four complementary components: EIEStem for shallow boundary preservation, C3k2-EMA for multi-scale feature aggregation, SPPF-LSKA for contextual modeling, and LDConv for adaptive downsampling of irregular weed targets. On 200 reference images, the SAM-SigLIP-VLM workflow achieved 92.4% label precision, a mean IoU of 0.813, and an F1-score of 0.905, while reducing the annotation time per image from 96.8 s to 31.6 s. Under five-fold image-level cross-validation on a UAV grassland weed dataset, the improved detector achieved 0.762 ± 0.004 mAP@0.5 and 0.545 ± 0.005 mAP@0.5:0.95, improving the YOLOv11 baseline by 4.0 and 4.7 percentage points, respectively. Additional evaluations cover annotation quality, module ablation, module placement, downsampling mechanisms, detector comparison, and validation on an external public crop-weed dataset. These results show that the proposed framework improves both annotation efficiency and UAV grassland weed detection performance in complex vegetation backgrounds.
Entomopathogenic fungi represent promising eco-friendly bioinsecticides, but are hindered by low virulence and host immunity. Here, we engineer a novel immune-evasion mechanism by exploiting the host's own pathogen recognition system. The insect β-1,3-glucan recognition protein 1 (βGRP1), containing a β-glucan-binding CBM39 domain and a glycoside hydrolase (GH16) domain, was identified as a key immune activator in insects. Silencing GmβGRP1 in Galleria mellonella significantly increased susceptibility to Beauveria bassiana. We engineered B. bassiana to secrete a catalytically inactive GmβGRP1 variant (GH16-deficient GmβGRP1cbm) that preemptively binds fungal β-glucans but lacks immune-activating capacity. This engineered strain exhibited significantly accelerated killing across diverse insects and transferred enhanced virulence to Metarhizium robertsii. Mechanistic studies confirmed that GmβGRP1cbm masks fungal β-glucan epitopes, preventing recognition and downstream immune activation. This "stealth" strategy─hijacking host immunity through preemptive occupation with decoy proteins─provides a paradigm shift toward next-generation bioinsecticides combining efficacy with environmental sustainability.
Timely and accurate detection of colorectal polyps plays an important role in the prevention and early diagnosis of colorectal cancer. Despite the advancement of deep learning-based methods, automatic polyp detection remains a challenging problem due to factors such as the small size of polyps, apparent similarity of polyps to surrounding tissue, variable quality of colonoscopy images, and the presence of noisy samples in the training data. In this study, a lightweight, fast, and robust framework for colorectal polyp detection was proposed that combines Local Outlier Factor (LOF)-based preprocessing with the YOLOv11n object detection model. In this study, five public datasets were used, including CVC-ColonDB, CVC-ClinicDB, Kvasir-SEG, ETIS, and EndoScene. Since these datasets originally contained segmentation labels, the binary masks of polyps were converted into bounding boxes to be used for training the object detection model. To improve the quality of the training data, LOF with 30 neighbors and a contamination rate of 5% was applied only to the training set to remove outliers and potentially noisy samples, while the validation and test data were left unchanged. Then, the YOLOv11n model was trained using the cleaned data and its performance was evaluated using five-fold cross-validation. The results showed that the proposed LOF-YOLOv11n framework achieved Precision of 94.73%, Recall of 91.46%, F1-score of 93.28%, mAP@0.5 of 96.54%, and mAP@0.5:0.95 of 78.01%. Also, the model showed an average speed of 56.1 frames per second, indicating its potential for image-level real-time inference under the evaluated experimental conditions. Supplementary analyses and ablation studies showed that LOF-based preprocessing can provide descriptive and limited improvements in the stability of the learning process and model performance, without biasing the model evaluation by removing difficult examples from the test set. Overall, the results suggest that combining training-data quality improvement with a lightweight YOLOv11n detector may provide a promising image-level framework for computer-aided colorectal polyp detection, although further validation on independent clinical and video-based datasets is required before clinical deployment.
Whole Slide Images are central to modern pathology and computational histopathology. However, tissue processing and slide scanning can introduce artifacts that degrade image quality and hinder computer-aided diagnosis (CAD) systems. Therefore, automated artifact processing has drawn much attention in recent years. Nevertheless, current methods exhibit limitations, including reliance on annotated data and inadequate pixel-level localization capabilities, resulting in fragmented workflows. To address these challenges, we propose an unsupervised automatic pipeline for artifact detection, localization, and restoration. First, based on the features extracted by the histopathology foundation model, anomaly heatmaps are generated using Normalizing Flow to enable precise artifact detection. Subsequently, optimal artifact masks are identified via anomaly heatmaps and unsupervised semantic segmentation for accurate localization. Finally, a mask-guided artifact restoration is performed using a diffusion model. Experimental results demonstrate that our method effectively handles synthetic and real-world artifacts using only normal training data and improves downstream performance on both BCSS segmentation and TCGA-BLCA tumor staging classification, confirming our method's efficacy in bolstering the robustness of CAD systems, reducing manual verification, and providing an automated solution for artifacts in histopathology images.
Schizophrenia (SCZ) and bipolar disorder (BIP) share substantial common-variant liability but differ in cognition, medical comorbidity, and treatment response. Here we decomposed this overlap into schizophrenia-predominant, bipolar-predominant, and shared psychosis dimensions to test whether these components show distinct pleiotropic and biological profiles. Using the largest available SCZ and BIP GWAS, we applied bidirectional mtCOJO and Genomic SEM to derive SCZcondBIP, BIPcondSCZ, and PSY-shared and validated them using inter-component genetic correlations, FinnGen psychiatric endpoints, and Genomic SEM latent factors. We then characterized each component across cognitive, cardiometabolic, and immune traits, followed by genomic risk-locus discovery, pathway analysis, developmental expression profiling, and drug-target enrichment. The three components showed marked divergence. SCZcondBIP was negatively genetically correlated with cognition, education, metabolic syndrome, C-reactive protein, and neutrophil percentage, whereas BIPcondSCZ showed the opposite cognitive profile and shifted toward positive cardiometabolic and immune correlations. PSY-shared retained the mixed cognitive pattern seen at the disorder level and intermediate peripheral correlations, indicating that shared psychosis liability masks stronger disorder-specific differences. Between-component contrasts were approximately twice the magnitude of the corresponding SCZ-versus-BIP contrasts. We identified 248 consensus genomic risk loci, including 81 not detected in the input disorder GWAS. Biologically, PSY-shared was enriched for synaptic signalling, ion-channel, and neurodevelopmental pathways; SCZcondBIP primarily implicated synaptic-signalling and cellular-homeostasis pathways; and BIPcondSCZ showed weaker but distinct enrichment for synaptic-vesicular biology. Drug-target enrichment further separated the components, with strong antipsychotic enrichment for PSY-shared and distinct non-antipsychotic signals for the conditional factors. These findings show that SCZ and BIP genetic risk is best understood as biologically distinguishable shared and disorder-predominant dimensions that differentially map onto cognitive, cardiometabolic, immune, and molecular architecture. These findings provide a framework for evaluating whether component-specific polygenic scores improve stratification of cognitive, cardiometabolic, and inflammatory heterogeneity across severe psychiatric illness.
This review summarizes current evidence on the biological, diagnostic, prognostic, and therapeutic significance of microRNAs (miRNAs/miRs) in prostate cancer (PCa). Dysregulated miRNA networks contribute to PCa initiation, progression, metastasis, and therapy resistance by regulating androgen receptor signaling, proliferation, apoptosis, epithelial-mesenchymal transition, angiogenesis, hypoxia signaling, bone tropism, and castration-resistant evolution. Oncogenic miRNAs, including miR-21, miR-93, miR-9, miR-181a, and miR-182, promote malignant phenotypes through survival, TGF-β, PI3K/AKT, MAPK, HIF-1α, and EMT-associated pathways, whereas tumor-suppressor miRNAs, including the miR-34 family, miR-145, miR-122, and miR-382, restrict proliferation, stem-like traits, invasion, metastasis, and treatment resistance. Circulating, urinary, and exosomal miRNAs have potential as minimally invasive biomarkers for PCa detection, risk stratification, recurrence monitoring, metastatic risk prediction, and assessment of castration-resistant disease. Multi-miRNA panels, such as miR-375-3p/miR-182-5p, miR-34b-3p/miR-361-5p/miR-200c-3p, and miR-200c/miR-605/miR-135a/miR-433/miR-106a, may outperform individual markers by capturing multiple biological pathways and reducing single-marker variability; however, most remain at the discovery or validation stage rather than routine clinical implementation. Translation requires standardized sample handling, hemolysis control, reproducible RNA extraction, validated normalization strategies, assay harmonization, locked thresholds, and multicentre prospective validation against PSA, imaging, histopathology, and established risk models. High-throughput platforms, including qRT-PCR, microarray, next-generation sequencing, and nCounter digital counting, support miRNA discovery and validation but differ in sensitivity, specificity, cost, throughput, bioinformatic complexity, and clinical deployability. Therapeutically, anti-miRs, miRNA mimics, sponges, masks, CRISPR-based approaches, and nanocarrier-assisted delivery systems provide experimental strategies for inhibiting oncogenic miRNAs or restoring tumor-suppressor miRNAs. Preclinical evidence supports miR-21 inhibition and replacement of miR-34a, miR-145, miR-15a/miR-16-1, miR-124, and miR-205. Still, clinical translation is limited by off-target effects, immune activation, delivery efficiency, endosomal escape, tumor heterogeneity, pharmacokinetics, toxicity, scalability, and manufacturing reproducibility. Overall, miRNAs provide mechanistic insight into PCa heterogeneity and offer promising opportunities for precision diagnosis, prognosis, and therapy, with the most realistic near-term application being integration of validated miRNA panels with PSA, multiparametric MRI, pathology, and multi-omic or AI-assisted risk models.
Residual respiratory events (RREs) are common in continuous positive airway pressure (CPAP) users. They can lead to poor CPAP adherence and early termination. The management of RREs has not been standardized yet. The purpose of this study was to test a structured management algorithm for RREs detected by CPAP devices in OSA patients. This was a prospective observational clinical study that aimed to test a specific algorithm in patients exhibiting RREs. This study also included a case-control comparison. Cases were OSA patients (AHI ≥ 15) exhibiting RREs (AHIflow >10) during the first 3 months of CPAP use, reported by telemonitoring. The control group were similar CPAP patients with no RREs. All underwent in-lab polysomnography (PSG) with manual titration to set the fixed CPAP pressure on the Resmed® CPAP device. In case of RRE, a specific algorithm was followed to manage and resolve these events, applied only after leak correction. A total of 19/273 patients exhibited RREs. Compared to matched controls, cases had higher RREs on titration PSG and used more often naso-buccal masks (p < 0.05). Among the 19 RRE patients, predominant obstructive AHIflow was observed in 15 patients, which generally resolved rapidly with an increase in CPAP pressure, allowing an AHIflow ≤10 to be obtained. Predominant central AHIflow and Cheyne-Stokes respiration (CSR) was observed in four patients. Individualized management was successful and one was moved to auto-servo ventilation. Once leaks are controlled, occurrence of RREs in compliant CPAP users is uncommon. A structured algorithm helped clarify the likely mechanism of device-detected RREs and guided individualized management during the first months of CPAP therapy. The most common cause of RRE was insufficient CPAP pressure, but cases of COMISA and of "idiopathic" central sleep apnea (CSA)/CSR also occurred that required specific management.
This study addresses current gaps in the 3D printing literature for filament composition (polymer type, color, and brand), aerosol emissions, and human health risks. Six metals (aluminum, magnesium, manganese, chromium, iron, and copper) were found in nonmetal filaments depending on polymer type, color, and brand using inductively coupled plasma-mass spectrometry. For copper- and steel-filled filaments, scanning electron microscopy and elemental analysis indicated a higher metal concentration within the filaments than on the surface. More VOCs were emitted from acrylonitrile butadiene styrene (ABS) filaments compared to polylactic acid (PLA) filaments, according to thermal desorption unit samples analyzed by gas chromatography. Styrene, emitted from ABS filaments, appears to pose the greatest potential health risk among known VOCs, given its concentration and reported effects. Toluene and benzaldehyde were emitted in lower concentrations from both ABS and PLA filaments and may pose a potential risk to human health. Based on three inflammatory markers across two exposure times using BEAS-2B epithelial lung cells, steel showed the highest proinflammatory response, possibly due to chromium and overall particulate matter. 3D printer users should minimize aerosolized emissions by taking proper precautions, such as ensuring adequate ventilation and wearing face masks, particularly during extended exposure.
Marine oil spills pose serious threats to marine ecosystems and coastal socio-economic systems, while the scarcity of annotated SAR oil-spill samples limits the training and generalization of deep learning-based detection models. To address this problem, this study proposes an improved Pix2PixGAN method for SAR oil-spill sample generation. The proposed method uses real SAR images as background priors and combines oil-spill masks with multi-channel random noise to jointly constrain background realism, target morphology, and texture diversity. A spatially adaptive normalization module, PatchGAN discriminator, least-squares adversarial loss, Dropout regularization, and segmentation-consistency constraint are integrated to enhance local texture realism and structural consistency. Experiments on the SOS dataset show that the proposed method outperforms Pix2Pix, PGGAN, and BEGAN in FID, KID, Global BIHD, ENL difference, and GMHD, indicating better consistency with real SAR samples in both feature distribution and SAR-related image-level statistical characteristics. Ablation experiments further confirm the complementary effects of spatial semantic modulation and multi-channel noise injection. In downstream U-Net segmentation validation, the model trained only with generated samples achieved performance comparable to that trained with real samples only, while combining real and generated samples produced slight numerical improvements. These results indicate that the generated samples contain learnable oil-spill structural and textural information, but the improvement should be interpreted as evidence of sample usability rather than statistically significant superiority. Cross-dataset validation on the independently constructed A Symphony-Yellow Sea dataset and additional experiments using DeepLabV3Plus, PSPNet, and SegNet further suggest that the generated samples can provide useful supplementary training cues for independent SAR oil-spill scenes when real annotated samples are limited.
Chronic calcium imaging offers a window into how single neurons and ensemble activity change across days where identifying the same neurons from one session to the next is the prerequisite for answering questions regarding learning, drift, and plasticity over time. Yet only ∼2-3% of imaging laboratories publish longitudinal cross-session work, because existing registration tools depend on spatial-footprint or temporal correlations that degrade under repeated recording sessions. Here, we introduce Stars2Cells (S2C) , a tracking pipeline inspired by astrometric plate-solving that represents each neuron's local geometry as a four-dimensional quad descriptor invariant to rotation, translation, and uniform scaling. S2C operates purely on centroid coordinates and combines descriptor-space matching, Random Sample Consensus (RANSAC) verification, and Hungarian assignment. Across a synthetic benchmark of 1,262 paired runs spanning 100-1,000 neurons and 8 perturbation conditions plus 1 identity sanity-floor, S2C reached pooled F1 = 98.4% compared to the standard ROI-based matching of 36.0%. To show what this enables, we applied the pipeline to dorsomedial striatum (DMS) imaging during oral fentanyl behavioral-economics self-administration. Here, we show that a conserved population-rewarded lever press response in DMS masks near-complete single-neuron turnover. This representational-drift signature we demonstrated is invisible to the bulk photometry, and resolving it requires the same-cell tracking S2C provides. S2C is distributed as a GUI-driven standalone application for both macOS and Windows, requiring no Python, command line, or virtual environment setup.
Corneal confocal microscopy (CCM) enables non-invasive imaging of the sub-basal nerve plexus for early diagnosis of diabetic neuropathy, but its utility is hindered by inherent noise and low contrast in raw images. We present NerveBoost, a weakly supervised framework for CCM denoising and enhancement guided by anatomical priors. Unlike supervised methods requiring paired clean data, NerveBoost uses binary nerve masks to construct a pseudo-target via region-specific gamma correction. A composite loss function integrates weighted reconstruction, gradient consistency, background smoothness, and foreground-background contrast constraints to jointly optimize noise suppression and structural enhancement within an encoder-decoder architecture. Results indicate that NerveBoost effectively enhances nerve visibility while maintaining structural fidelity, offering a robust and efficient preprocessing solution for clinical CCM analysis without requiring paired ground-truth data.
This study challenges a common misconception in current flood hazard research that floods will decrease under a warming climate in many rivers due to changes in land-surface processes, particularly decreases in snowmelt. This view persists because most studies on flood changes rely on daily resolution data, which masks flood changes due to subdaily rainfall intensification. We show that daily streamflow projections systematically underestimate flood changes compared with hourly projections in Alpine catchments. Our results highlight that the sign of change can even switch from negative to positive in strongly snow-influenced catchments when taking an hourly perspective because intensified subdaily precipitation can outweigh snowmelt decline. These results highlight that using daily instead of hourly projections may lead to wrong conclusions on both the magnitude and direction of flood changes. An hourly resolution perspective is, therefore, crucial to reliably guide adaptation strategies to extreme Alpine floods in a warming world.
Ensuring food safety in agri-food production requires reliable inspection data for detecting foreign objects and other contamination risks. Hyperspectral imaging (HSI) is a relevant non-destructive modality for this purpose because it combines spatial information with detailed spectral responses. This Data Descriptor presents HSI-AgriFoodAnomaly, an open hyperspectral image dataset acquired under conveyor-based industrial-like conditions for foreign object annotation in an oat, white-chocolate and dark-chocolate mixture. The dataset contains 147 calibrated hyperspectral cubes, associated red-green-blue (RGB) renderings, pixel-level binary masks, and polygon annotations. Each hyperspectral cube was acquired in the visible to near-infrared range, with 300 contiguous spectral bands covering approximately 381-1016 nm. The annotated foreign objects (FOs) include textile and fibre-based materials, plastics, paper-based materials, metals, wood and plant residues, minerals, glass, mixed-object scenes, and anomaly-free scenes. The dataset is organised into training, validation and test subsets at the cube level, in order to prevent data leakage between splits. HSI-AgriFoodAnomaly can be reused for hyperspectral image analysis, foreign object localisation, binary classification, object detection, semantic segmentation, and benchmarking of data processing pipelines. The dataset and the accompanying code are publicly available.
This study aimed to determine hearing benefits and challenges for children using cochlear implants (CIs) in one ear and acoustic hearing in the other (bimodal hearing) through hearing aids (HA-CI) or normal hearing (single-sided deafness/SSD-CI). Participants were 34 children with CIs [MAge (SD)=12.2 (3.2) years] and 6 peers with typical hearing/controls [MAge (SD)=13.8 (1.7) years]. Side of better hearing (aural preference) was measured by word recognition in quiet and noise, and spondee-word recognition thresholds in noise co-located/0° or separated (left/right 90°). Self-reported hearing was measured using the Speech, Spatial and Qualities of Hearing Scale (SSQ). Localization of stationary and moving sound (and unrestricted head movements) and sensitivity to interaural level and timing cues were measured (separate controls [n=5, MAge (SD)=14.0 (1.6) years]). Speech perception using the CI alone was similar in children with SSD-CI and HA-CI (p=0.84) but SSD-CI users had greater aural preference in their acoustic ear (p<0.01). CI users had high errors localizing stationary sound and poor moving sound detection (p<0.001) which was associated with earlier onset of hearing loss, and poor access to interaural cues (p<0.001) compared to controls. Head movements in children with CIs tended to favour their non-implanted ear (p<0.05). Results indicate that children with asymmetric hearing gain speech perception with their CI but have aural preference for the non-implanted ear, particularly in the presence of good residual hearing and later onset of hearing loss. Bimodal listeners had poor access to binaural cues, poor spatial hearing with slightly better results in children with later onset of hearing loss and ineffective gaze movements.