BackgroundObesity is a disease with high heterogeneity. Both overall obesity and central obesity are associated with increased risks of having cardio-metabolic co-morbidities. This study is aimed to examine the cardio-metabolic characteristics and comorbidity profile of the middle-aged and elderly Chinese with general and central obesity by clustering them into different subgroups, which would lead to a deepened understanding of their distinct medical needs.MethodsAdopting an unsupervised machine learning approach, we conducted a clustering analysis of the adiposity and cardio-metabolic profiles of the middle-aged and elderly Chinese with general obesity and central obesity. The data was obtained from the China Health and Retirement Longitudinal Study (CHARLS). The subgroup features were examined. The risks of having obesity-related co-morbidities (i.e. hypertension, dyslipidemia, diabetes, heart problem, stroke) in each cluster were then compared.ResultsAmong the 7,970 subjects selected from the baseline cohort, 41.88% (n = 3,338) had general obesity, while 71.29% (n = 5,682) had central obesity. These individuals with either general obesity or central obesity were clustered into four groups, respectively: (1) obesity with relatively healthier metabolites; (2) hyperuricemia subtype; (3) hyperglycemia-insulin resistance subtype; and (4) the average subtype. The results indicated among people with either general obesity or central obesity, those with high levels in HbA1c level and TyG index concurrently demonstrated more severe adiposity issues and unhealthier cardio-metabolic profile.ConclusionsThis data-driven study identified a novel classification strategy to identify subtypes of the middle-aged and elderly Chinese with general obesity and central obesity and classify their adiposity and cardio-metabolic profiles. With clinically accessible metrics, this approach could inform precise risk stratification by revealing subtype-specific heterogeneity during initial assessments.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12872-025-05157-x.

ABSTRACT Cloud contamination is a common degradation in optical remote sensing images, adversely affecting the application of such images. Deep-learning-based cloud removal algorithms with auxiliary information have received increasing attention in recent years. Most of these methods rely on georeferenced, cloud-free optical images from other periods as references. However, the inherent gap between the reference and the target images often leads to inaccurate reconstruction. Unsupervised methods have also been proposed, mitigating the gap issue by eliminating the need for reference images. Yet, they typically and solely rely on reconstruction loss during training, often resulting in unnatural outcomes. To tackle these limitations, we propose ALM-CR (Adversarial Learning – based Multimodal Cloud Removal), an unsupervised two-stage framework that leverages synthetic aperture radar (SAR) as auxiliary input. The first stage performs SAR-to-optical translation for structural and approximate spectral recovery, followed by SAR-optical fusion to restore fine-grained spectral details. The proposed adversarial learning strategy removes the need for temporal reference images, enabling precise reconstruction of cloud-covered images while preventing overfitting. Experimental results demonstrate that our method surpasses existing unsupervised methods on both reference and no-reference metrics, and reconstructs spectral information more consistently than supervised methods.

ABSTRACT Hyperspectral images (HSI) are renowned for their high spectral resolution and extensive wavelength coverage, but suffer from limited spatial and temporal resolution due to imaging sensor constraints. However, hyperspectral images with low spatial resolution and low temporal resolution are difficult to be applied to subsequent more advanced tasks such as object detection, classification, and anomaly detection. Physical constraints make it impossible for a single satellite sensor to acquire images that simultaneously have high resolution in time, space, and spectrum, so image fusion is the most efficient choice to achieve this goal. Spatial-temporal-spectral fusion (STSF) has the purpose of synthesizing different information which has the advantage in temporal, spatial, and spectral aspect respectively from multisource satellite data to reconstruct HSI with high spatial resolution and high temporal resolution. In order to address the problems that the linear relationship in current spectral reconstruction is difficult to accurately map the complex relationships among space, time and spectrum, and that deep convolutional neural networks are prone to overfitting, as well as to enhance the model’s ability to extract spectral features, this paper designs an unsupervised STSF method. The proposed method has three stages: Stage 1 (Spatial-Spectral Downsampling) analyzes the spatial-spectral degradation of time 1 observed images; Stage 2 (Spectral Upsampling) develops a spectral upsampling network (with the shared spatial-spectral downsampling network) to upsample multispectral data to hyperspectral data; Stage 3 uses the trained network to upsample multispectral images of time 2 for high-spatial-resolution HSI. In order to verify the proposed method, it is compared with other state-of-the-art methods on simulated and real datasets. This proves the method’s advantages in richer spatial-spectral details and more accurate reconstruction.

Recognizing the crossing behavior of non-motorized traffic at intersections is crucial for urban transportation safety and efficiency. However, traditional methods often overlook environmental context, while deep learning approaches require extensive labeled data. To address these limitations, we propose an unsupervised learning framework that infers waiting and passing behaviors by combining motion features with spatio-temporal semantic context. Our model employs a convolutional autoencoder with an integrated clustering layer, trained through a joint optimization of reconstruction, clustering, and K-neighborhood contrastive losses. Experimental results show our framework outperforms other unsupervised methods. Notably, incorporating semantic features improves clustering accuracy from 87.3% to 91.3%, validating the effectiveness of our approach in capturing complex behavior.

Abstract There is a growing need for labeled data, yet manual annotation is costly, error-prone, and often infeasible in privacy-sensitive, highly imbalanced domains such as fraud detection. We introduce a fully unsupervised framework that combines unsupervised SHapley Additive exPlanations (SHAP) feature selection with our novel unsupervised labeling method. We apply unsupervised SHAP to the Kaggle Credit Card Fraud Detection and Medicare Part D datasets to produce high-impact feature subsets, and then label the datasets with our unsupervised labeling approach. To effectively evaluate the labels generated by our novel methodology, we apply a baseline unsupervised learner, Isolation Forest (IF), to both the original datasets and their subsets. We calculate Matthew’s Correlation Coefficient (MCC), Jaccard Index (JI), Precision, Recall, and F1-score by comparing our generated labels against the ground truth labels. It is important to note, the ground truth labels were used solely for evaluation. Our empirical results surpass the results obtained with the full feature dataset and baseline. By improving label quality while reducing computational complexity and preserving privacy, our approach offers a practical solution for learning from unlabeled, severely imbalanced data.

Cryo-electron tomography (cryo-ET) datasets are rich sources of information capable of describing the localizations, structures, and interactions of macromolecules. However, most current methods for localizing particles in cryo-electron tomograms are limited to macromolecules with known structures, require extensive manual annotations, and/or are computationally expensive. Here, we present PickET, a method for localizing macromolecules in tomograms that does not rely on expert annotations and prior structures. Its performance is demonstrated on a diverse dataset comprising over a hundred tomograms from publicly available datasets, varying in sample types, sample preparation conditions, microscope hardware, and image processing workflows. We demonstrate that PickET can simultaneously localize macromolecules of various shapes, sizes, and abundance. The predicted particle localizations can be used for 3D classification and de novo structural characterization. Our fully unsupervised approach is efficient and scalable, and enables high-throughput analysis of cryo-ET data.

Deformable image registration is essential in medical image analysis. The state-of-the-art approaches are unsupervised methods based on convolutional neural networks (CNN) and vision transformers (ViT). While CNNs perform well in extracting local features, ViTs perform better in extracting global features. This study aimed to compare the performance of CNN and ViT in unsupervised deformable image registration. We have proposed a unified registration framework and evaluated both architectures. Experiments have been conducted using 4D-CT. The results have shown ViT-based registration to achieve superior performance compared to CNN-based methods. The findings have indicated vision transformer architectures to be more effective than convolutional networks for unsupervised deformable registration on 4D-CT data. Foundation Item: This work is supported by the National Natural Science Foundation of China (No.61801413).

Human cancer is highly heterogeneous, resulting in variable drug resistance and clinical outcomes. This complexity hinders accurate prognosis prediction and the development of targeted therapies. Molecular subtyping addresses these challenges by grouping cancers into more homogeneous subsets based on molecular characteristics, enabling subtype-specific treatment strategies. Subtyping is crucial for early diagnosis, personalized therapy, and improved survival by capturing differential therapeutic responses. Existing approaches to cancer subtyping fall into supervised and unsupervised categories. Supervised methods, often trained on The Cancer Genome Atlas (TCGA), rely on predefined subtype annotations but face limitations in generalizability and novel subtype discovery. Unsupervised methods, while capable of identifying new subtypes, may overlook widely recognized ones, hindering consistency with established classifications. Multi-omics approaches improve accuracy but are constrained by costs and data collection. We propose CancerSubminer, a hybrid subtyping framework that integrates supervised and unsupervised learning. A subtype classifier is first trained on labeled data, after which clustering is applied to extracted features, with low-confidence samples reassigned to refine subtype boundaries. Model is retrained with the refined subtypes, and adversarial training corrects batch effects and learns domain-invariant features across labeled TCGA and unlabeled external datasets. A subsequent semi-supervised fine-tuning phase aligns subtypes between datasets and designates low-confidence samples as potential novel candidates. CancerSubminer was evaluated on five cancer types, including breast, bladder, brain, kidney, and thyroid cancers, using TCGA methylation data with annotated subtypes and unlabeled datasets from the Gene Expression Omnibus. The framework outperformed state-of-the-art subtyping models (iClusterPlus, iClusterBayes, NEMO) and clustering methods (Spectral, K-means). Kaplan-Meier survival analysis demonstrated significant prognostic separation (p < 0.05) for all cancers, including thyroid cancer where predefined subtypes showed no significance but CancerSubminer-derived subtypes did. These findings highlight CancerSubminer's ability to identify distinct prognostic subtypes, mitigate batch effects, and improve prognostic stratification across heterogeneous datasets. CancerSubminer is publicly available at https://github.com/joungmin-choi/CancerSubminer .

We present a computational pipeline that links nuclear morphology to mRNA expression–based cell phenotypes under diverse biological conditions, including aging, disease progression, and drug response, using RNAscope imaging. The pipeline consists of three components: nuclear segmentation from RNAscope images, nuclear morphology identification, and downstream statistical analysis. Central to our approach is a novel unsupervised method, based on deep disentangled representation learning, which effectively captures diverse nuclear morphologies in large-scale datasets, as validated on synthetic benchmarks. We applied the full pipeline to RNAscope data targeting dopaminergic and glutamatergic neuron populations in the midbrains of mice and humans. Our analyses uncovered distinct nuclear morphology differences between dopaminergic and non-dopaminergic, as well as glutamatergic and non-glutamatergic neurons, in both species. Moreover, we identified a significant interaction between neurotransmitter identity and healthy aging in mice, reflected in systematic changes in nuclear morphology. These findings position nuclear morphology as a scalable and informative imaging-based readout of cell identity and physiological state.

CADRE: A novel unsupervised reconstruction algorithm for limited-angle CT of ancient wooden structures.

Reverse logistics (RL) plays a crucial role in promoting circularity and sustainability in supply chains, particularly in the face of increasing waste generation and growing environmental demands. In recent years, machine learning (ML) has emerged as a strategic tool to enhance processes, decision-making, and outcomes in RL. This article presents a systematic review of ML applications in reverse logistics, highlighting trends, challenges, and research opportunities. The analysis covers 52 articles retrieved from the Scopus and Web of Science databases, following the PRISMA protocol. The results show that the most frequently employed techniques are supervised models, followed by unsupervised methods and, to a lesser extent, reinforcement learning. The main ML applications in RL focus on return and waste generation forecasting, process optimization, classification, pricing, reliability assessments, and consumer behavior analysis. The studies examined predominantly use traditional evaluation metrics, such as MAPE and F1-score, while few consider multidimensional indicators encompassing long-term social or environmental impacts. Key challenges identified include data scarcity and quality, inherent uncertainties in reverse supply chains, and the high computational cost of models. This article also points to research gaps concerning metadata standardization, the absence of public benchmarks, model explainability, and the integration of ML with simulations and digital twins, indicating pathways toward more robust, transparent, and sustainable RL.

The integration and analysis of multi-modal data are increasingly essential across various domains including bioinformatics. As the volume and complexity of such data grow, there is a pressing need for computational models that not only integrate diverse modalities but also leverage their complementary information to improve clustering accuracy and insights, especially when dealing with partial observations with missing data. We propose Generalized Probabilistic Canonical Correlation Analysis (GPCCA), an unsupervised method for the integration and joint dimensionality reduction of multi-modal data. GPCCA addresses key challenges in multi-modal data analysis by handling missing values within the model, enabling the integration of more than two modalities, and identifying informative features while accounting for correlations within individual modalities. GPCCA demonstrates robustness to various missing data patterns and provides low-dimensional embeddings that facilitate downstream clustering and analysis. In a range of simulation settings, GPCCA outperforms existing methods in capturing essential patterns across modalities. Additionally, we demonstrate its applicability to multi-omics data from TCGA cancer datasets and a multi-view image dataset. GPCCA offers a useful framework for multi-modal data integration, effectively handling missing data and providing informative low-dimensional embeddings. Its performance across cancer genomics and multi-view image data highlights its robustness and potential for broad application. To make the method accessible to the wider research community, we have released an R package, GPCCA, which is available at https://github.com/Kaversoniano/GPCCA.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12859-025-06227-9.

Abstract Earth system models have deficiencies and biases that raise large uncertainties in the simulated ocean temperature and salinity. Methods of assessing vertical thermohaline structures traditionally rely on computing the mean profile in somewhat ad hoc regions. However, the global ocean has complex spatial variability, complicating regional classification. Here, to assess the global thermohaline profiles simulated by CMIP6 models, we apply an unsupervised clustering method to classify the Argo-observed profiles and compare them with the historical simulations of eight CMIP6 models. The results show that in each season, the Argo-observed thermohaline profiles in the global ocean can be divided into around 60 classes, which have distinct geographical distributions. CESM2 (MRI-ESM2-0) performs best in simulating the main climatological spatial patterns and vertical structures of temperature (salinity) in 0–1000 m, with an overall bias of 1.19°C (0.31). Temperature biases in 8-CMIP6 multimodel mean highlight several hotspots: 1) the tropical Atlantic Ocean and the Gulf of Mexico; 2) the Kuroshio Extension; 3) south of the equator in the western Pacific Ocean; and 4) the Agulhas retroflection region. Vertical structures of temperature and salinity biases vary in these four regions. Composite analysis of vertical bias patterns and dynamic characteristics suggests that region 1 may reflect remote advective influences, regions 2 and 4 may be associated with the inaccuracy of mode water simulations, and region 3 appears attributable to wind stress and mixed layer depth biases. The proposed evaluation framework will provide important insights for optimizing global ocean simulations in future modeling efforts.

The distribution of seagrass ecosystems along the coast of Pemenang District, North Lombok Regency, plays an important ecological and economic role. However, this ecosystem is vulnerable to damage due to human activities. The purpose of this study was to analyze the spatial distribution and extent of seagrass cover based on its cover density level. This study used remote sensing methods through the interpretation of Sentinel-2A images. The analysis process included image merging, image cropping, water column correction using the Lyzenga algorithm, image classification using supervised and unsupervised methods, and field validation at 60 observation points. The results showed that seagrass meadows were widespread along the coast and Gili islands (Gili Trawangan, Gili Meno, and Gili Air), with a total area of ​​685.26 ha. Seagrass cover density was classified into three categories, namely high (33.69%), medium (45.02%), and low (21.29%), with a mapping accuracy level of 73.33%. Based on the Decree of the Minister of Environment No. 200 of 2004, most seagrass meadows were categorized as damaged. Factors influencing this condition include domestic waste, tourism activities, ship anchors, human movement (madak), and turtle grazing activity. Seven seagrass species were identified, with Cymodocea rotundata being the most dominant. These findings emphasize the importance of regular monitoring and ongoing management to maintain the sustainability of seagrass ecosystems in the Pemenang coastal area.

BackgroundEffective prevention of cardiac malformations is constrained by limited understanding of etiology. We used 2011-2021 MarketScan US insurance claims data to identify and characterize associations between maternal and paternal characteristics and non-chromosomal cardiac malformations.MethodsAmong 693,483 singleton live-birth pregnancies of women linked to infants (of which 488,146 linked to fathers), odds ratios were estimated between 2000 clinical diagnostic and medication codes (500 clinical and 500 medication codes each for mothers and fathers) and cardiac malformations (n = 7522 affected pregnancies) using logistic regression. Associations were selected using procedures to control the false discovery rate (FDR). Selected codes were grouped using latent semantic analysis alongside hierarchical clustering.ResultsAt the 5% FDR, 67 codes are selected of which 63 are maternal and four paternal. Elevated risk with maternal diabetes, obesity, and chronic hypertension, highlights the importance of maternal cardiometabolic health for cardiac malformations. Additional potential signals included maternal fingolimod or azathioprine use. The relative lack of paternal associations is consistent with prior findings of few replicated associations with paternal non-genetic exposures.ConclusionsScreening associations, with interpretation aided by unsupervised machine learning methods, identifies, in this study, both known risk factors and potential signals. Signals might be explained by confounding, other systematic errors, or chance, and warrant further investigation.

Segmentation and detection of biological objects in fluorescence microscopy is of paramount importance in cell imaging. Deep learning approaches have recently shown promise to advance, automatize and accelerate analysis. However, most of the interest has been given to the segmentation of static objects of 2D/3D images whereas the segmentation of dynamic processes obtained from time-lapse acquisitions has been less explored. Here we adapted DeepFinder, a U-Net originally designed for 3D noisy cryo-electron tomography (cryo-ET) data, for the detection of rare dynamic exocytosis events (termed ExoDeepFinder) observed in temporal series of 2D Total Internal Reflection Fluorescence Microscopy (TIRFM) images. ExoDeepFinder achieved good absolute performances with a relatively small training dataset of 12000 events in 60 cells. We rigorously compared deep learning performances with unsupervised conventional methods from the literature. ExoDeepFinder outcompeted the tested methods, but also exhibited a greater plasticity to the experimental conditions when tested under drug treatments and after changes in cell line or imaged reporter. This robustness to unseen experimental conditions did not require re-training demonstrating generalization capability of our deep learning model. ExoDeepFinder, as well as the annotated training datasets, were made transparent and available through an open-source software as well as a Napari plugin and can directly be applied to custom user data. The apparent plasticity and performances of ExoDeepFinder to detect dynamic events open new opportunities for future deep learning guided analysis of dynamic processes in live-cell imaging.

Removing contaminants is a delicate, yet crucial step in neutral hydrogen ( intensity mapping and often considered the technique's greatest challenge. Here, we address this challenge by analysing intensity maps of about $100$ deg^2 at redshift z≈0.4 collected by the MeerKAT radio telescope, an SKA Observatory (SKAO) precursor, with a combined 10.5-hour observation. Using unsupervised statistical methods, we removed the contaminating foreground emission and systematically tested, step-by-step, some common pre-processing choices to facilitate the cleaning process. We also introduced and tested a novel multiscale approach: the data were redundantly decomposed into subsets referring to different spatial scales (large and small), where the cleaning procedure was performed independently. We confirm the detection of the cosmological signal in cross-correlation with an ancillary galactic data set, without the need to correct for signal loss. In the best set-up we achieved, we were able to constrain the distribution through the combination of its cosmic abundance (Ω_ and linear clustering bias (b_ up to a cross-correlation coefficient (r). We measured Ω_ hi b_ hi r = 0.93 ± 0.17 with a ≈6σ confidence, which is independent of scale cuts at both edges of the probed scale range ($0.04 łesssim k łesssim 0.3 ,h$ Mpc^-1), corroborating its robustness. Our new pipeline has successfully found an optimal compromise in separating contaminants without incurring a catastrophic signal loss. This development instills an added degree of confidence in the outstanding science we can deliver with MeerKAT on the path towards intensity mapping surveys with the full SKAO.

To address the decision-making requirements for drainage gas recovery in horizontal gas wells within low-permeability tight reservoirs, this study proposes an intelligent classification approach that integrates supervised and unsupervised learning techniques. Initially, the static and dynamic performance characteristics of gas wells are characterized across multiple dimensions, including static performance, liquid production intensity, liquid drainage capacity, and liquid carrying efficiency. These features are then quantitatively categorized using Linear Discriminant Analysis (LDA). Subsequently, a hybrid classification framework is developed by integrating LDA with the K-means clustering algorithm. The effectiveness of this supervised–unsupervised fusion method is validated through comparative analysis against direct K-means clustering, demonstrating enhanced classification accuracy and interpretability. Key findings are summarized as follows: (1) Classification based on individual dynamic or static parameters exhibits low consistency, indicating that single-parameter approaches are insufficient to fully capture the complexity of actual production conditions. (2) By incorporating both dynamic and static parameters and applying a strategy combining LDA-based dimensionality reduction with K-means clustering, gas wells are precisely classified into five distinct categories. (3) Tailored optimization strategies are proposed for each well type, including production allocation optimization, continuous production (without the need for drainage gas production measures), mandatory drainage measures, foam-assisted drainage, and optimal tubing or plunger lift systems. The methodologies and findings of this study offer theoretical insights and technical guidance applicable to the classification and management of horizontal gas wells in other unconventional reservoirs, such as shale gas formations.

Full-field digital mammography (FFDM) is the most common imaging technique for breast cancer screening programs. Still, it is limited by noise from quantum effects, electronic issues, and X-ray scattering, affecting the image quality. Traditional denoising methods based on filters and transformations perform poorly due to the complex, tissue-dependent nature of noise, while supervised deep learning methods require extensive, often unavailable datasets with paired noisy and noiseless images. Consequently, unsupervised denoising methods, which do not require clean images as ground truth, are gaining attention. However, their application to FFDM is poorly explored. This study investigates the use of Noise2Void (N2V), an unsupervised denoising approach adapted to digital mammography images for the first time. N2V employs blind spot masking to remove noise without requiring noiseless images. The method was assessed using different metrics on real clinical images and artificially noised images: contrast-to-noise ratio (CNR), and structural similarity index (SSIM). A qualitative evaluation was also made based on a questionnaire provided to radiologists. The results show that evaluated metrics increase on N2V images; these results are comparable with traditional methods. Despite showing quantitative performance comparable to traditional methods, N2V retains potential for clinical application as a flexible, annotation-free approach for retrospective, low-dose mammography imaging.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-19499-1.

Accurate detection of mitosis is crucial in automated cell analysis, yet many existing methods depend heavily on deep learning models or complex detection techniques, which can be computationally intensive and error-prone, particularly when segmentation is incomplete. This study presents a novel unsupervised method for mitosis detection, leveraging the geometric properties of the Cassini oval to reduce computational costs and enhance robustness. Our approach integrates a newly developed deep learning model, MaxSigNet, for initial cell segmentation. We subsequently employ the Cassini oval in its single-ring mode to detect mother cells in the initial frame and switch to double-ring mode in subsequent frames to identify daughter cells and confirm mitosis events. The success of this method hinges on the presence of equal non-zero foci values in the mother cell and distinct non-zero foci values in the daughter cells, which indicate accurate mitosis detection. The method was evaluated across six datasets from four different cell lines, achieving perfect F1, Recall and Precision scores on four datasets, with scores of 96% and 85% on the remaining two. Comparative analysis demonstrated that our method outperformed similar approaches in F1 and Recall metrics. Additionally, the method showed substantial robustness to incomplete segmentation, with only a 20% average drop in F1 scores when tested with older segmentation methods like K-means, Felzenszwalb and Watershed. The proposed method offers a significant advancement in mitosis detection by leveraging the Cassini oval's properties, providing a reliable and efficient solution for automated cell analysis systems. This approach promises to enhance the accuracy and efficiency of cellular behaviour studies, with potential applications in various biomedical research fields.