BackgroundCoronary artery disease (CAD) remains a leading cause of mortality globally. Accurate identification and characterization of significant coronary artery lesions via CT is important for proper diagnosis and patient management. However, current supervised techniques for lesion identification require detailed manual annotations, which are both labor-intensive and prone to human error. While semi-supervised or weakly supervised approaches can partially reduce manual annotation burden, they still depend on annotated datasets, which remain limited and inconsistent in CAD research. Therefore, there is a strong clinical motivation for developing robust unsupervised methods that eliminate annotation dependency.AimThis study aims to develop and evaluate a novel unsupervised method utilizing clustering algorithms to automatically classify and characterize significant lesions in coronary arteries from CT images, thereby addressing limitations of supervised and semi-supervised methods.MethodsWe analyzed 45 anonymized coronary artery CT scans from patients hospitalized between 2018 and 2022, selected based on the presence of plaques causing at least 30% stenosis. Although small, the dataset represented a clinically relevant population with a diverse range of lesion types (calcified, mixed, and soft plaques). Vessel segmentation was performed using nnU-Net, followed by skeletonization and extraction of statistical and Haralick texture features. Dimensionality reduction was executed using principal component analysis, and lesion clustering was conducted using both k-means and a hybrid clustering algorithm. Supervised methods are defined as algorithms that require labeled data for training, whereas unsupervised methods, as applied in this study, do not require labeled data and instead rely solely on inherent patterns within the imaging features. The effectiveness of lesion classification, including calcified, mixed, and soft plaques, was assessed. Additionally, hemodynamic significance was verified by comparison with fractional flow reserve (FFR) measurements.ResultsVessel segmentation yielded a mean Dice coefficient of 0.93, indicating high segmentation accuracy. The hybrid clustering algorithm demonstrated superior lesion classification performance, achieving sensitivity rates of 95.6% for calcified plaques, 88.3% for mixed plaques, and 74.1% for soft plaques. These performance indicators compare favorably to previously reported supervised and unsupervised approaches. Furthermore, the method reliably identified hemodynamically significant lesions as confirmed by FFR (n = 15 lesions).ConclusionsOur proposed unsupervised clustering-based method effectively classifies and characterizes coronary artery lesions without the need for manual annotations. However, the small sample size and limited number of lesions validated by FFR (n = 15) restrict broad generalizations and clinical translation. External validation on larger, multicenter datasets is essential to confirm these promising findings. This method offers a practical, accurate, and efficient diagnostic approach, potentially streamlining clinical workflow and improving patient outcomes.

ABSTRACTThis study examines the effects of varying concentrations of zinc oxide nanoparticles (ZnO NPs) on the biochemical profile of mung plants, utilizing Raman and ultraviolet–visible (UV–Vis) spectroscopy combined with machine learning algorithms for data analysis. Mung plants, developed in the lab under optimized growth conditions, were exposed to different ZnO NPs concentrations (0.2, 0.4, 0.6, 0.8, 1.0, 1.2, and 1.4 mM, particle size < 30 nm). UV–Vis measurements show that concentrations of photosynthetic pigments declined with higher ZnO NPs treatment, implying a decrease in photosynthetic efficiency due to oxidative stress. The analysis of acquired Raman spectra shows disruptions in key biochemicals, including carotenoids, pectin, lignin, cellulose, protein, and aliphatics on exposure to nanoparticles, indicating effects on the metabolic processes of plants. To assess these changes, various machine learning algorithms were employed, including unsupervised methods (k‐means clustering, density‐based spatial clustering of applications with noise, agglomerative clustering, and principal component analysis) and supervised methods (support vector machine, random forest, k‐nearest neighbor, decision tree, logistic regression, gradient boosting, and linear discriminant analysis). The support vector machine and random forest models achieved the highest classification accuracy, precision, recall, and f1 scores, effectively differentiating between NPs‐induced biochemical changes. Additionally, unsupervised algorithms revealed distinct clustering patterns, aiding in the identification of NPs treatment effects on plants. These findings demonstrate the potential of integrating confocal micro‐Raman and UV–Vis spectroscopy with machine learning as a rapid, early, nondestructive, and robust tool for providing valuable insights for sustainable agricultural practices.

Advances in cytometry have led to increases in the number of cellular markers that are routinely measured. The resulting complexity of the data has prompted a shift from manual to automated analysis methods. Currently, numerous unsupervised methods are available to cluster cells based on marker expression values. However, phenotyping the resulting clusters is typically not part of the automated process. Manually identifying both marker definitions (e.g. CD4+, CCR7+, CD45RA+, CD19-) and descriptive cell type names (e.g. naïve CD4+ T cells) based on marker expression values can be time-consuming, subjective, and error-prone. In this work we propose an algorithm that addresses these problems through the creation of an automated tool, CytoPheno, that assigns marker definitions and cell type names to unidentified clusters. First, post-clustered expression data undergoes per-marker calculations to assign markers as positive or negative. Next, marker names undergo a standardization process to match to Protein Ontology identifier terms. Finally, marker descriptions are matched to cell type names within the Cell Ontology. Each part of the tool was tested with benchmark data to demonstrate performance. Additionally, the tool is encompassed in a graphical user interface (R Shiny) to increase user accessibility and interpretability. Overall, CytoPheno can aid researchers in timely and unbiased phenotyping of post-clustered cytometry data.

Radiotherapy is a crucial treatment for brain tumor malignancies. To address the limitations of CT-based treatment planning, recent research has explored MR-only radiotherapy, requiring precise MR-to-CT synthesis. This study compares two deep learning approaches, supervised (Pix2Pix) and unsupervised (CycleGAN), for generating pseudo-CT (pCT) images from T1- and T2-weighted MR sequences. 3270 paired T1- and T2-weighted MRI images were collected and registered with corresponding CT images. After preprocessing, a supervised pCT generative model was trained using the Pix2Pix framework, and an unsupervised generative network (CycleGAN) was also trained to enable a comparative assessment of pCT quality relative to the Pix2Pix model. To assess differences between pCT and reference CT images, three key metrics (SSIM, PSNR, and MAE) were used. Additionally, a dosimetric evaluation was performed on selected cases to assess clinical relevance. The average SSIM, PSNR, and MAE for Pix2Pix on T1 images were 0.964 ± 0.03, 32.812 ± 5.21, and 79.681 ± 9.52 HU, respectively. Statistical analysis revealed that Pix2Pix significantly outperformed CycleGAN in generating high-fidelity pCT images (p < 0.05). There was no notable difference in the effectiveness of T1-weighted versus T2-weighted MR images for generating pCT (p > 0.05). Dosimetric evaluation confirmed comparable dose distributions between pCT and reference CT, supporting clinical feasibility. Both supervised and unsupervised methods demonstrated the capability to generate accurate pCT images from conventional T1- and T2-weighted MR sequences. While supervised methods like Pix2Pix achieve higher accuracy, unsupervised approaches such as CycleGAN offer greater flexibility by eliminating the need for paired training data, making them suitable for applications where paired data is unavailable.

Achieving efficiency, robustness and interpretability in classification continues to pose significant challenges in data analysis. To address these issues, the granular-ball-based k-nearest neighbors (GBkNN) classifier has recently been proposed and demonstrated promising results. However, the performance of GBkNN heavily relies on the quality of GBs, and existing GB generation methods often use only purity as the evaluation criterion and apply a fixed threshold-based stopping rule. These limitations restrict their effectiveness in practical applications. To overcome this, we extend the advanced unsupervised GB generation method based on the Principle of Justifiable Granularity (POJG) to the supervised classification setting, aiming to enhance the overall performance of GBkNN. Furthermore, to mitigate the instability caused by the inherent randomness of

This paper provides a comparative study of unsupervised Deep Learning (DL) methods for anomaly detection in Flight Data Monitoring (FDM). The paper applies Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), Convolutional Neural Network (CNN), classic Transformer architecture, and LSTM combined with a self-attention mechanism to real-world flight data and compares the results to the current state-of-the-art flight data analysis techniques applied in the industry. The paper finds that LSTM, when integrated with a self-attention mechanism, offers notable benefits over other deep learning methods as it effectively handles lengthy time series like those present in flight data, establishes a generalized model applicable across various airports and facilitates the detection of trends across the entire fleet. The results were validated by industrial experts. The paper additionally investigates a range of methods for feeding flight data (lengthy time series) to a neural network. The innovation of this paper involves utilizing Transformer architecture and LSTM with self-attention mechanism for the first time in the realm of aviation data, exploring the optimal method for inputting flight data into a model and evaluating all deep learning techniques for anomaly detection against the ground truth determined by human experts. The paper puts forth a compelling case for shifting from the existing method, which relies on examining events through threshold exceedances, to a deep learning-based approach that offers a more proactive style of data analysis. This not only enhances the generalization of the FDM process but also has the potential to improve air transport safety and optimize aviation operations.

SummaryHow protein move and deform determines their interactions with the environment and is thus of utmost importance for cellular functioning. Following the revolution in single protein 3D structure prediction, researchers have focused on repurposing or developing deep learning models for sampling alternative protein conformations. In this work, we explored whether continuous compact representations of protein motions could be predicted directly from sequences, without exploiting 3D structures. SeaMoon leverages protein Language Model (pLM) embeddings as input to a lightweight convolutional neural network. We assessed SeaMoon against ~ 1 000 collections of experimental conformations exhibiting diverse motions. It predicts at least one ground-truth motion with reasonable accuracy for 40% of the test proteins. SeaMoon captures motions inaccessible to normal mode analysis, an unsupervised physics-based method relying solely on 3D geometry, and generalises to proteins without detectable sequence similarity to the training set. SeaMoon is easily retrainable with novel or updated pLMs.

We present Gene-Embedded Multi-modal Networks (GEM-Net), a semi-supervised framework for constructing multi-modal networks centered on genes. GEM-Net uses gene-level modules and selectively incorporates heterogeneous omics profiles using a correlated meta-analysis strategy that accounts for scale imbalance, missingness, and intra-modular correlation. Prior to network inference, we developed a harmonized data processing protocol that adjusts each omic layer independently through a shared mathematical workflow involving transformation, dimensionality reduction, and regression-based covariate adjustment. GEM-Net modules were inferred and benchmarked against unsupervised methods using transcriptomic, metabolomic, and lipidomic data from the Long Life Family Study (LLFS), a unique cohort enriched for exceptional familial longevity and health. GEM-Net modules were more diverse and biologically interpretable, with stronger support from protein–protein interactions, transcriptional regulation, and metabolic annotations. Applying GEM-Net to metabolic health in LLFS revealed an axis between the microbiome-derived metabolite N-acetylglycine and immune genes (FCER1A, HDC, CPA3, MS4A2) associated with improved insulin sensitivity and reduced inflammation in healthy older individuals. GEM-Nets offer a reusable reference from a long-lived population and a generalizable framework for multi-omics discovery. https://doi.org/10.5281/zenodo.15003731.

During the operation of industrial equipment, substantial volumes of both normal and faulty unlabeled data are frequently generated. Traditional data-driven models are often susceptible to significant performance degradation in the presence of large amounts of unlabeled data, and their diagnostic decision-making processes typically lack robust physical interpretability. To address these challenges, this paper proposes an unsupervised Physical Information Neural Network (PINN) that utilizes the dynamic updating of pseudo-labels for the intelligent diagnosis of industrial bearings. The PINN is designed to compute the gradient of each network layer parameter through backpropagation, effectively combining data-driven learning with physical constraints and incorporating a regularization mechanism to enhance generalization capabilities. In the experiments, wavelet time-frequency diagrams are constructed based on three distinct fault states alongside the normal state of rolling bearings, elucidating the time-frequency characteristics under various operational conditions. Additionally, initial pseudo-labels are generated using the fault octave amplitude ratio. A dynamic pseudo-label update mechanism is implemented during the training phase, incorporating adaptive correction based on physical constraints. This mechanism ensures that the pseudo-labels are refined adaptively to accurately represent the current physical signal state within an unsupervised context, thereby improving the model’s self-learning capability and fault diagnosis accuracy. Ultimately, the proposed model is validated on a publicly available bearing dataset, demonstrating robust fault recognition performance and strong physical interpretability, even in scenarios characterized by limited labeled data and substantial background noise. The model exhibits promising potential for practical applications in the field of industrial diagnostics.

To address the problems of feature extraction, cost of obtaining labeled samples, and large differences in domain distribution in bearing fault diagnosis on variable operating conditions, an unsupervised domain-adaptive bearing fault diagnosis method based on migration learning using MSACNN-IJMMD-DANN (multi-scale and attention-based convolutional neural network, MSACNN, improved joint maximum mean discrepancy, IJMMD, domain adversarial neural network, DANN) is proposed. Firstly, in order to extract fault-type features from the source domain and target domain, this paper establishes a MSACNN based on multi-scale and attention mechanisms. Secondly, to reduce the feature distribution difference between the source and target domains and address the issue of domain distribution differences, the joint maximum mean discrepancy and correlation alignment approaches are used to create the metric criterion. Then, the adversarial loss mechanism in DANN is introduced to reduce the interference of weakly correlated domain features for better fault diagnosis and identification. Finally, the method is validated using bearing datasets from Case Western Reserve University, Jiangnan University, and our laboratory. The experimental results demonstrated that the method achieved higher accuracy across different migration tasks, providing an effective solution for bearing fault diagnosis in industrial environments with varying operating conditions.

Unsupervised SAR Image Change Detection Method Based on Fused Difference Images

With the rapid development of single-cell RNA sequencing (scRNA-seq) technology, researchers can now investigate gene expression at the individual cell level. Identifying cell types via unsupervised clustering is a fundamental challenge in analyzing single-cell data. However, due to the high dimensionality of expression profiles, traditional clustering methods often fail to produce satisfactory results. To address this problem, we developed NetworkSSC, a network-guided sparse subspace clustering (SSC) approach. NetworkSSC operates on the same assumption as SSC that cells of the same type have gene expressions lying within the same subspace. In addition, it integrates a regularization term incorporating the gene network's Laplacian matrix, which captures functional associations between genes. Comparative analysis on nine scRNA-seq datasets shows that NetworkSSC outperforms traditional SSC and other unsupervised methods in most cases.

Bolted flanged joints are essential for connecting piping and process equipment but are vulnerable to localized corrosion that leads to sudden, unpredictable leaks. Electrochemical noise (EN) measurements can detect such corrosion, yet processing EN data is time-consuming and requires expertise. This study applies recurrent neural networks (RNNs) to automate corrosion type identification on flange surfaces using raw EN signals from spontaneous electrochemical reactions. In this work, supervised, hybrid, and unsupervised ML approaches are evaluated using experimentally obtained EN data. Among supervised models, the long short-term memory (LSTM) model achieves 93.62% accuracy. A hybrid method combining LSTM autoencoder features with a random forest classifier improves accuracy to 97.85%. An unsupervised method using LSTM autoencoder, principal component analysis, and k-means clustering also shows strong potential for real-time corrosion monitoring. Automated identification of corrosion types on flanged joints supports more effective material protection strategies, reducing the risk of failure in critical infrastructure.

Abstract Smart mobility has ushered in advanced sensing technologies. These, together with high‑level data analytics, are revolutionizing how we analyze driving behavior. Excellent performance in dealing with real-world, high-technology complexities for machine learning has made wide enthusiasm to utilize them to study driver behavior. This article gives a thorough overview of the important machine learning methods—especially clustering and classification techniques—that help analyze complex driving behaviors, predict fuel and energy use, and improve vehicle safety systems. The review specifically explains unsupervised methods like fuzzy c-means, k-means, and density-based spatial clustering of applications with noise, as well as supervised techniques such as artificial neural networks, k-nearest neighbors, and support vector machines. Also, this review discusses the integration of clustering and classification techniques with hybrid deep learning models, and examines their applications in eco-driving, energy forecasting, and intelligent transport systems while offering novel findings that contribute to more sustainable mobility. Emphasis is placed on how these methods transform vast, heterogeneous driving data into actionable insights that support real-time monitoring and personalized feedback for eco-driving and smart transportation applications. Finally, current benefits and barriers, and future research opportunities and challenges in integrating machine learning into intelligent transportation systems are reviewed. The potential to advance to safer, better, and more sustainable forms of mobility is emphasized.

To enhance the generalization of networks and avoid redundant training efforts, cross-image hyperspectral anomaly detection (HAD) based on deep learning has been gradually studied in recent years. Cross-image HAD aims to perform anomaly detection on unknown hyperspectral images after a single training process on the network, thereby improving detection efficiency in practical applications. However, the existing approaches may require additional supervised information or stacking of networks to improve model performance, which may impose high demands on data or hardware in practical applications. In this paper, a simple and lightweight unsupervised cross-image HAD method called Variational Joint Discrimination Network (VJDNet) is proposed. We leverage the reconstruction and distribution representation ability of the variational autoencoder (VAE), learning the global and local discriminability of anomalies jointly. To integrate these representations from the VAE, a probability distribution joint discrimination (PDJD) module is proposed. Through the PDJD module, the VJDNet can directly output the anomaly score mask of pixels. To further facilitate the unsupervised paradigm, a sample pair generation module is proposed, which is able to generate anomaly samples and background representation samples tailored for the cross-image HAD task. The experimental results show that the proposed method is able to maintain the detection accuracy with only a small number of parameters.

Anomaly detection methods play a crucial role in automated quality control within modern manufacturing systems. In this context, unsupervised methods are increasingly favored due to their independence from large-scale labeled datasets. However, existing methods present limited multi-scale feature extraction ability and may fail to effectively capture subtle anomalies. To address these challenges, we propose UAMS, a pyramid-structured normalization flow framework that leverages the symmetry in feature recombination to harmonize multi-scale interactions. The proposed framework integrates a Multi-Scale Attention Aggregation (MSAA) module for cross-scale dynamic fusion, as well as a Self-Supervised Predictive Convolutional Attention Block (SSPCAB) for spatial channel attention and masked prediction learning. Experiments on the MVTecAD dataset show that UAMS largely outperforms state-of-the-art unsupervised methods, in terms of detection and localization accuracy, while maintaining high inference efficiency. For example, when comparing UAMS against the baseline model on the carpet category, the AUROC is improved from 90.8% to 94.5%, and AUPRO is improved from 91.0% to 92.9%. These findings validate the potential of the proposed method for use in real industrial inspection scenarios.

With the continuous advancement of industrial automation, intelligent defect detection has become a crucial means of ensuring product quality. However, the cost and labor required to obtain high-quality labeled data limit the widespread practical application of traditional supervised learn-ing methods. Therefore, semi-supervised learning (SSL) and unsupervised learning (UL) methods have received extensive attention from researchers due to their superior performance in low-labeled or unlabeled scenarios. This paper provides a systematic survey of typical semi-supervised and unsupervised learning methods in the field of industrial defect detection, analyzes the core concepts and key technologies of unsupervised learning and semi-supervised learning, as well as compara-tive analysis on relevant datasets, and finally proposes future development direc-tions.

Solid-state electrolytes (SSEs) constitute a compelling alternative to liquid organic electrolytes in lithium-ion batteries. Non-toxic and non-flammable, SSEs not only enhance safety but could potentially increase energy density by enabling the use of metal lithium anodes. Several potential candidates demonstrating fast ionic conductivity (IC) have been identified to date, and progress in rationally optimizing their performance without compromising their chemical or mechanical stability requires a fundamental understanding of the mechanisms underlying their high IC. Atomistic simulation techniques, such as ab initio molecular dynamics (AIMD) and classical molecular dynamics (CMD), have notably contributed to this task. In combination with structural and electrochemical spectroscopy methods, these simulation schemes have enabled the detailed description of ionic diffusion pathways along specific sublattices within the host material. However, external factors such as temperature, pressure, and aliovalent doping can significantly perturb the location, geometry, and connectivity of these sites. Moreover, exceedingly fast ionic transitions may render the precise identification and classification of crystallographic sites a challenging task from both an experimental and computational standpoint.A closely related problem, strongly dependent on the accurate identification of diffusion pathways, is that of unravelling the impact of collective motion on the IC. The traditional approach to studying the collective motion of ions is through van Hove correlation functions, which are limited to pair-correlations [1,2]. This limited perspective fails to capture the concentration effects and multi-ion interactions, which have been recognized as key in systems with high concentrations of mobile ions, such as SSE materials [3,4]. While the cooperative diffusion has been shown to potentially enhance the IC by lowering the activation energy [3], identifying and classifying these diffusion "chains" from long simulation trajectories efficiently and accurately remain a difficult task.In light of these challenges, the development of advanced computational tools capable of analysing atomistic trajectories with minimal supervision has emerged as a critical area of focus in materials science. To address the complexity of analysing ionic transport and its relation to material structure from CMD and AIMD trajectories, several schemes that require a priori knowledge of crystallographic sites location, geometry, and associated cutoff values have been reported [5,6]. However, these methods are sensitive to structural details and thermal noise, impacting their precision. Recently, unsupervised methods requiring fewer structural parameters have been developed [7], but generally require high-frequency sampling to achieve accurate results, which can be computationally expensive.In this work, we introduce a novel density-based unsupervised method, enabling the accurate and efficient identification and categorization of diffusion sites (Figure 1). Furthermore, this method offers a robust framework for analysing atomistic trajectories, capturing both independent and collective properties of diffusive ions without a priori knowledge of the crystallographic details. Our method bridges existing gaps in previous schemes, providing a comprehensive and adaptable tool for understanding the relationship between material structure and ionic diffusion in SSEs. The new approach, implemented in the open-source CrySF package, was successfully applied to CMD and AIMD simulation trajectories of several prospective SSEs with a wide range of ICs, such as Li7La3Zr2O12, Li10GeP2S12, Li6PS5Br, Li3PS4, Li3OCl, LiGaO2, LiMn(HCO2), and Li2B12H12. Our results align well with experimental and computational data, demonstrating the method’s accuracy and adaptability, as well as its potential to elucidate the interplay between structural properties, ionic mobility, and collective migration phenomena in SSEs. In particular, we demonstrate through CrySF that the mere presence of long strings of atoms diffusing concertedly does not necessarily imply a high IC. Traditionally regarded as the hallmark of superionic conductors, our analysis shows that the frequency of such long strings, the number of distinct sites visited, and the persistence of motion in a particular direction must all be considered.

Technological advancements and AI-based research have significantly influenced our daily lives. Human activity recognition (HAR) is a key area at the intersection of various AI technologies and application domains. In this study, we present our novel time series classification approach for monitoring the physical behaviors of the elderly and patients. This approach, which integrates supervised and unsupervised methods with generative models, has been validated for HAR, showing promising results. Our method was specifically adapted for healthcare and surveillance applications, enhancing the classification of physical behaviors in the elderly. The hybrid approach proved its effectiveness on the HAR70+ dataset, surpassing traditional recurrent convolutional network-based approaches. We further evaluated the surveillance system for the elderly (Surv-Sys-Elderly) model on the HARTH and HAR70+ datasets, achieving an accuracy of 94,3% on the HAR70+ dataset for recognizing elderly behaviors, highlighting its robustness and suitability for both clinical and domestic environments.

BackgroundThe spread of SARS-CoV-2 Omicron variant and its subvariants, highly transmissible but responsible of milder disease, has increased unreported infection cases. Identifying unaware infected individuals is crucial for estimating the true prevalence of infection and evaluating the breadth of hybrid immunity. In this study, this challenge was addressed by applying several Machine Learning approaches.MethodsA group of 116 participants, vaccinated against SARS-CoV-2, was enrolled in the IMMUNO_COV study at Siena University Hospital, Italy. Blood samples were collected before and six months after third vaccine dose. Machine Learning analysis, involving dimensionality reduction techniques, unsupervised clustering methods and classification models, were applied to serological data including antibody responses specific for wild type SARS-CoV-2 strain as well as Delta, Omicron BA.1 and Omicron BA.2 variants. Spike- and nucleocapsid-specific B cells were also assessed in each participant.ResultsUsing dimensionality reduction and unsupervised clustering, participants are grouped into high- and low-responders, with infected participants mainly distributed within the high-responders. Implementation of a consensus-based approach, including k-NN, RF, and SVM models, identifies 14 participants unaware of previous infection. Their immunological profile, characterized by strong spike- and nucleocapsid-specific humoral and B cell responses, significantly differs from that of non-infected participants.ConclusionsMachine Learning approaches are applied to identify participants unaware of prior infection and to dissect their hybrid immunity profiles. Based on serological data, this cost-effective method can be a valuable tool for estimating the true prevalence of infection, improving comprehension of immune responses elicited by vaccination alone or combined with infection, and tailoring public health interventions.