ABSTRACT Academic freedom is considered a fundamental principle of higher education, indicating that scholars may engage in research without limitations, irrespective of geographical location, personal identity or institutional ties. However, trends in funding distribution, publication assessment, citation methodologies and promotion frameworks indicate a more limited reality. This article posits that the selection of research topics in global higher education is influenced not only by intellectual curiosity but also by incentive structures that favour specific subjects, methodologies and publication outlets, while making others professionally precarious. This paper utilises the sociology of knowledge, higher education studies and postcolonial and decolonial thought to introduce the concept of uneven universality, elucidating the mechanisms by which certain traditions achieve “universal” status while others are relegated to the category of local. The analysis of funding dependencies, journal hierarchies, accreditation systems, language dominance and evaluation metrics reveals that while academic autonomy is formally recognised, it remains structurally constrained. The article employs music scholarship as a case study to reconceptualise academic freedom as a conditional practice influenced by global governance. It concludes by proposing pathways towards more diverse forms of scholarly recognition within global higher education frameworks.

Traditional non-negative matrix factorization (NMF) faces challenges when dealing with complex data, primarily characterized by sensitivity to noise, neglect of data geometric structure, and inability to effectively utilize supervised information. To address these limitations, this paper proposes a class-driven robust non-negative matrix factorization with dual-hypergraph regularization (CRNMFDH) framework. The core contributions of this framework include the following: Firstly, the design of a novel dual-hypergraph regularization term that symmetrically captures and preserves the higher-order geometric structures of both the sample space and feature space, establishing a mutually reinforcing topological relationship between them. Secondly, an introduction of a class-driven mechanism to effectively integrate label information into the decomposition process, significantly enhancing the discriminative capability of the low-dimensional representations. Finally, the employment of a loss function based on correntropy to replace the traditional Euclidean distance, thereby enhancing the model’s robustness against noise and outliers. Extensive experiments across nine datasets demonstrate that CRNMFDH significantly outperforms existing state-of-the-art algorithms in multiple clustering evaluation metrics and noise robustness, providing an effective new solution for complex data clustering tasks.

Semantic segmentation of metallographic micrographs is a key task for quantitative microstructural analysis in additive manufacturing, yet it remains challenging due to phase heterogeneity, complex morphologies, and the scarcity of annotated data. The MetalDAM dataset, composed of 42 labeled scanning electron microscopy images of steel microstructures, has been widely adopted as a benchmark, with U-Net commonly reported as the strongest supervised baseline. Nevertheless, the encoder–decoder structure of U-Net imposes architectural constraints that hinder the precise delineation of heterogeneous and irregular phase boundaries under severe data limitations. To address this limitation, this paper investigates a Fully Convolutional Network (FCN)-based architecture as an alternative approach for semantic segmentation on the MetalDAM dataset. The FCN is trained and evaluated under the same experimental protocol as the U-Net baseline, enabling a direct and fair comparison. Performance is assessed using multiple evaluation metrics, including Intersection over Union (IoU), precision, recall, and mean Average Precision at an IoU threshold of 0.5. The results show that the FCN achieves comparable overall IoU values (0.75) while delivering substantial improvements at the class level, particularly for minority and morphologically complex phases, with gains of up to 25–30% in class-specific IoU. Additional metrics confirm enhanced robustness, with consistently higher precision, recall, and mAP@0.5 values. These findings demonstrate that FCN-based architectures constitute a competitive and robust alternative to U-Net for metallographic segmentation in additive manufacturing scenarios characterized by limited annotated data.

Abstract In recent years, Large Language Models (LLMs) have emerged as transformative tools across numerous domains, impacting how professionals approach complex analytical tasks. This systematic mapping study comprehensively examines the application of LLMs throughout the data science lifecycle. By analyzing relevant papers from Scopus and IEEE databases, we identify and categorize the types of LLMs being applied, the specific stages and tasks of the data science process they address, and the methodological approaches used for their evaluation. Our analysis includes a detailed examination of evaluation metrics employed across studies and systematically documents both positive contributions and limitations of LLMs when applied to data science workflows. This mapping provides researchers and practitioners with a structured understanding of the current landscape, highlighting trends, gaps, and opportunities for future research in this rapidly evolving intersection of LLMs and data science.

Spatial domain identification is an essential task for revealing spatial heterogeneity within tissues, providing insights into disease mechanisms, tissue development, and the cellular microenvironment. In recent years, spatial multi-omics has emerged as the new frontier in spatial domain identification that offers deeper insights into the complex interplay and functional dynamics of heterogeneous cell communities within their native tissue context. Most existing methods rely on static graph structures that treat all neighboring cells uniformly, failing to capture the nuanced cellular interactions within the microenvironment and thus blurring functional boundaries. Furthermore, cross-modal reconstruction performance is often degraded by overfitting to modality-specific noise, which may impair the precise delineation of spatial domains. Therefore, we present GATCL, a novel deep learning framework that integrates a graph attention network with contrastive learning (CL) for robust spatial domain identification. First, GATCL leverages the graph attention mechanism to dynamically assign weights to neighboring spots, adaptively modeling the complex cellular architecture. Second, it implements a cross-modal CL strategy that forces representations from the same spatial location to be similar while pushing those from different locations apart, thereby achieving robust alignment between modalities. Comprehensive experiments across six distinct datasets (spanning transcriptome, proteome, and chromatin) reveal that GATCL is superior to seven representative methods across six key evaluation metrics.

Breast cancer is the most common malignancy among women and a leading cause of cancer-related mortality, making early and accurate detection essential. This review summarises advances in breast imaging and computational diagnostics across mammography, ultrasound, and magnetic resonance imaging (MRI), highlighting challenges in differentiating benign from malignant lesions and identifying rarer tumour types. Key preprocessing steps—denoising, deblurring, and contrast enhancement—are reviewed as they improve image quality prior to analysis. Classical methods (e.g., thresholding, edge detection, and region growing) are compared with deep learning approaches for segmentation and classification. CNNs, RNNs, and emerging transformer-based models consistently outperform handcrafted pipelines, with representative studies reporting 5–15% gains in AUC/accuracy and deep models achieving AUC > 0.85–0.95 on several benchmarks. The review also discusses dataset constraints, common evaluation metrics (AUC, Dice, sensitivity, specificity), and clinical translation barriers such as interpretability and domain shift. Overall, AI-driven methods show strong potential to enhance early detection and support improved breast cancer outcomes.

This paper reviews the use of machine learning techniques (ML) in diagnosing and predicting Chronic Kidney Disease (CKD) in articles published between 2020 and 2025. Across twenty selected studies, we systematically analyzed data preprocessing techniques, feature selection methods, classification algorithms, and performance evaluation metrics. We conclude that there is a significant trend towards ensemble and hybrid models, which outperform traditional single algorithms. At the same time, deep learning still shows remarkable predictive results with limited interpretability challenges. Feature selection remains one of the few constants for improving accuracy, efficiency, and transparency. Major challenges include limited datasets, class imbalance, and the gap between the proposed models and clinical implementation. Our review highlights the need for standardized evaluation methods, diverse multicenter datasets, and increased concentration on explainable AI to ease and enhance clinical integration of machine learning based CKD diagnostic tools.

Existing Infrared and Visible Image Fusion (IVIF) methods typically assume high-quality inputs. However, when handing degraded images, these methods heavily rely on manually switching between different pre-processing techniques. This decoupling of degradation handling and image fusion leads to significant performance degradation. In this paper, we propose a novel VLM-Guided Degradation-Coupled Fusion network (VGDCFusion), which tightly couples degradation modeling with the fusion process and leverages vision-language models (VLMs) for degradation-aware perception and guided suppression. Specifically, the proposed Specific-Prompt Degradation-Coupled Extractor (SPDCE) enables modality-specific degradation awareness and establishes a joint modeling of degradation suppression and intra-modal feature extraction. In parallel, the Joint-Prompt Degradation-Coupled Fusion (JPDCF) facilitates cross-modal degradation perception and couples residual degradation filtering with complementary cross-modal feature fusion. Extensive experimental results indicate that the proposed VGDCFusion demonstrates marked superiority in degraded image fusion tasks, surpassing existing state-of-the-art methods in both qualitative visual quality and quantitative evaluation metrics (e.g., the AG and SF measures achieve average improvements of approximately 15% and 14.75%, respectively). Our code is available at https://github.com/Lmmh058/VGDCFusion.

Synthetic positron emission tomography (PET) imaging, enabled by deep learning, represents a promising approach to minimize radiation exposure while preserving diagnostic accuracy. However, variability in methodologies, performance metrics, and clinical applications needs to be assessed. This systematic mapping review examines the current state of research in synthetic PET generation, analyzing their methodological frameworks and evaluating the clinical relevance. A systematic search in Scopus, PubMed, and Google Scholar (2019-2024) identified peer-reviewed studies on deep learning-based synthetic PET. Review articles, conference abstracts, and inaccessible full texts were excluded. Data extraction covered study characteristics, imaging modalities, architectures, and evaluation metrics. Due to study heterogeneity, the risk of bias was not formally assessed. Results were synthesized through descriptive and quantitative analysis. Of the initial 116 studies retrieved, 34 were included, 25 of them (73.5%) on brain/neuro using magnetic resonance imaging, computed tomography, or low-dose PET data to generate full-dose or tracer-specific PET. Common architectures included convolutional neural networks, generative adversarial networks, and U-Nets. Peak signal-to-noise ratio (PSNR) ranged 22.69-56.87 dB, structural similarity index measure (SSIM) 0.38-1.00 and mean absolute error (MAE) 1.37-72.00%. Whole-body applications were less frequent (9/34, 26.5%) but showed improvements in oncologic imaging, in particular for tumor detection and image quality. Despite promising advancements, challenges remain, including limited data availability, variability in tracer uptake, and the lack of standardized evaluation metrics. The absence of large/multicenter datasets limits the generalizability of findings. This review highlights promising advancements in synthetic PET imaging using deep learning, with several studies demonstrating the potential for high-quality image generation and substantially reduced radiation exposure. These developments are particularly significant in pediatric populations, where minimizing radiation dose is crucial to ensure patient safety and long-term health. Nonetheless, methodological variability and limited clinical validation continue to pose substantial challenges. Future research should prioritize the development of standardized evaluation protocols, the use of larger and more diverse datasets-including pediatric cohorts-and comprehensive real-world clinical validation to support the safe and effective translation of synthetic PET techniques into clinical practice. Deep learning-based synthetic PET imaging enhances diagnostics while reducing radiation, but requires methodological standardization and clinical validation for broader adoption. Deep learning can create full-dose PET images with less radiation exposure. Neurological applications dominate synthetic PET research, maintaining essential diagnostic detail. Challenges include limited datasets and variability in tracer uptake, necessitating further advancements.

Performance monitoring of roadside green infrastructure: The different perspectives of alternate water quality metrics.

Spatial clustering is a critical analytical task in spatial transcriptomics (ST) that aids in uncovering the spatial molecular mechanisms underlying biological phenotypes. Along with the numerous spatial clustering methods, there comes the imperative need for an effective metric to evaluate their performance. An ideal metric should consider three factors: label agreement, spatial organization, and error severity. However, existing evaluation metrics focus solely on either label agreement or spatial organization, leading to biased and misleading evaluations. To fill this gap, we propose CEMUSA, a novel graph-based metric that integrates these factors into a unified evaluation framework. Extensive testing on both simulated and real datasets demonstrate CEMUSA's superiority over conventional metrics in differentiating clustering results with subtle differences in topology and error severity, while maintaining computational efficiency. The source code and data is freely available at https://github.com/YihDu/CEMUSA. CEMUSA is implemented as an R package at https://yihdu.github.io/CEMUSA. Supplementary data are available at Bioinformatics online.

Benchmark datasets for assessing polyploid genome assembly are limited because real polyploid genomes often contain unknown structural variations, complex repeats, and heterogeneous divergence among homologous copies. In this study, a minimal, fully controlled virtual dataset is provided for reproducible benchmarking of triploid de novo assembly using short reads. A 1-Mbp haploid reference sequence is generated and iteratively mutated to produce three genome copies (A-C) across 100 mutation steps, creating a divergence gradient that transitions from nearly identical to moderately diverged triploid genomes. For each divergence level, paired-end Illumina reads are simulated at uniform coverage and processed through error correction followed by de Bruijn graph assembly across multiple k-mer sizes. The dataset provides the full set of reference genomes, read sets, assemblies, and evaluation metrics, allowing direct reproduction of trends such as overcollapsed contigs at low divergence and improved genome separability at higher divergence. This compact resource offers method developers and users a transparent, reproducible standard for evaluating k-mer strategies and assembly behavior in triploid genomes.

Speech enhancement aims to improve speech quality and intelligibility in noisy environments and is important in applications such as hearing aids, mobile communications and automatic speech recognition (ASR). This paper shows a structured review of speech enhancement techniques, classified depending on the channel configuration and signal processing framework. Both traditional and modern approaches are discussed, including classical signal processing methods, machine learning techniques, and recent deep learning-based models. Furthermore, common noise types, widely used speech datasets, and standard evaluation metrics for evaluating speech quality and intelligibility are reviewed. Key challenges such as non-stationary noise, data limitations, reverberation, and generalization to unseen noise conditions are highlighted. This review presents the advancements in speech enhancement and discusses the challenges and trends of this field. Valuable insights are provided for researchers, engineers, and practitioners in the area. The findings aid in the selection of suitable techniques for improved speech quality and intelligibility, and we concluded that the trend in speech enhancement has shifted from standard algorithms to deep learning methods that can efficiently learn information regarding speech signals.

With the integration of high-proportion renewable energy, the operation modes of the power system are becoming increasingly complex and diverse. The typical operation modes selected with manual experience cannot comprehensively represent system operating characteristics. To more accurately analyze system operating characteristics, an analysis method for power system operation modes based on autoencoder clustering is proposed. Compared to other clustering methods, the autoencoder clustering method can adapt to data of different types and structures, extract features and perform clustering in a reduced-dimensional space, and suppress noise in the data to a certain extent. First, multi-dimensional analysis metrics for power system operation modes are proposed. The metrics are used to evaluate system characteristics such as cleanliness, security, flexibility, and adequacy. The evaluation metrics for clustering are designed based on the metrics. Second, an operation mode analysis framework is constructed. The framework uses an autoencoder to extract implicit coupling relationships between system operation variables. The encoded feature vectors are used for clustering, which helps to find the internal similarities of the operation modes. Regulation resources such as pumped hydro storage are also considered in the framework. Finally, the proposed method is tested on the IEEE 39-node system. In the test, the comparison of clustering evaluation metrics and operation mode analysis errors shows that the proposed method has the best clustering performance and operation mode analysis effect compared to other clustering methods. The results prove that the proposed method can effectively extract the inner correlations and coupling relations of high-dimensional operating vectors, form consistent operation mode clusters, select typical operation modes, and accurately assess the characteristics and risks of the power system with high-proportion renewable energy integration. This paper helps to build a stronger power system that can integrate a higher proportion of renewable energy, replace fossil fuel generation, and contribute to a higher level of sustainable development.

Recent advances in deep learning (DL) have enabled the integration of diverse biomedical data for disease prediction and risk stratification. Building on this progress, the overall objective of this study was to develop and evaluate a multimodal DL framework for robust multi-label classification (MLC) of major comorbidities in patients with obstructive sleep apnea (OSA) using physiological time series signals and clinical data. This study proposes a robust framework for multi-label classification (MLC) of comorbidities in patients with OSA using diverse physiological and clinical data sources. We conducted a retrospective observational study including a convenience sample of 144 patients referred for overnight polysomnography at the Sleep Medicine Center (SleepLab Split), University Hospital Centre Split (KBC Split), Split, Croatia. Patients were selected based on predefined inclusion criteria and data availability. A one-dimensional Convolutional Neural Network (1D-CNN) was developed to process and fuse time series signals, oxygen saturation (), derived features, and nasal airflow (FP0), with demographic and physiological parameters, enabling the identification of key comorbidities such as arterial hypertension, diabetes mellitus, and asthma/COPD. The instruments included polysomnography-derived signals (SpO2 and FP0 airflow) and structured demographic/physiological parameters. Signals were preprocessed and used as inputs to the proposed fusion model. The proposed model was trained and fine-tuned using the Optuna hyperparameter optimization framework, addressing class imbalance through weighted loss adjustments. Its performance was comprehensively assessed using multi-label evaluation metrics, including macro/micro F1-score, AUC-ROC, AUC-PR, subset and partial accuracy, Hamming loss, and multi-label confusion matrix (MLCM). The study protocol was approved by the Ethics Committee of the School of Medicine, University of Split (Approval No. 003-08/23-03/0015, Date: 17 October 2023). The 1D-CNN achieved superior predictive performance compared to traditional machine learning (ML) classifiers with macro AUC-ROC = 0.731 and AUC-PR = 0.750. The model demonstrated consistent behavior across age, gender, and BMI groups, indicating strong generalization and minimal demographic bias. In conclusion, the results confirm that and airflow signals inherently encode comorbidity-specific physiological patterns, enabling efficient and scalable screening of OSA-related comorbidities without the need for full polysomnography. Although the study is limited by data set size, it provides a methodological basis for the application of multi-label DL models in clinical decision support systems. Future research should focus on the expansion of multi-center datasets, thereby improving model interpretability and potential clinical adoption.

Abstract Accurate fault detection and diagnosis are critical components of any fault-tolerant control system, especially for Unmanned Aerial Vehicles (UAVs) where reliability is paramount. Traditionally, both model-based and data-driven approaches have been applied for fault diagnosis. However, the increasing complexity of high-dimensional UAV systems has shifted focus toward data-driven methods, which leverage advanced classification algorithms to enhance fault identification and isolation. This study builds on this evolution by developing a sophisticated condition-based monitoring (CBM) system specifically designed for multirotor UAVs. In contrast to earlier studies that primarily relied on raw data for classifier training, this work introduces advanced preprocessing techniques and multi-domain feature extraction, significantly improving the robustness and accuracy of fault detection. A comparative analysis is performed between feature selection methods, including Recursive Feature Elimination with Cross-Validation (RFECV) and Variational Autoencoder (VAE), to extract critical insights into UAV operational behavior. Through testing and evaluating various classification models on data from a hexarotor UAV under diverse actuator fault conditions, this research identifies optimal approaches for real-time fault detection and diagnosis. Results demonstrate notable improvements across all evaluation metrics, establishing this approach as a substantial advancement in UAV fault tolerance.

LEFF-ViT: A locally enhanced vision transformer framework for accurate Alzheimer's Disease classification from brain MRI.

Deep learning models often exhibit bias when trained on imbalanced datasets, tending to favor majority classes. While Deep Synthetic Minority Over-sampling Technique (DeepSMOTE) mitigates this challenge by generating synthetic minority class images within an autoencoder’s latent space, the proposed architecture achieves further performance enhancement through the optimization of the latent space structure. This strategy introduces a composite loss function that minimizes intra-class distances while maximizing inter-class distances by incorporating class centroid information into the autoencoder training process. The approach was evaluated on six benchmark image datasets (MNIST, Fashion-MNIST, EMNIST, SVHN, GTSRB, and CIFAR10) with severe class imbalance ratios of 750:1, using five-fold cross-validation. Results demonstrate that Deep Class-Latent Synthetic Minority Oversampling Technique (DeepCLSMOTE) consistently outperforms baseline methods, including Balancing Generative Adversarial Network (BAGAN) and DeepSMOTE across all evaluation metrics, achieving statistically significant improvements in macro-average accuracy, precision, recall, and F1-measure. The enhanced performance is attributed to improved discriminative feature extraction in the optimized latent space, resulting in superior classification performance on highly imbalanced image datasets, particularly for critical minority classes.

Abstract The evolving threat landscape demands intrusion detection systems that adapt quickly to novel attack patterns and operate across heterogeneous environments. Recent studies show that Generative Adversarial Networks (GANs) can improve intrusion detection performance by generating synthetic attack traffic, balancing imbalanced datasets, enhancing adversarial robustness, and serving as anomaly detectors. This survey provides a comprehensive and systematic review of GAN-based intrusion detection system (IDS) research, analyzing the architectures employed—including Wasserstein GANs, conditional GANs, self-attention GANs, and specialized multi-generator designs—together with their applications, datasets, and evaluation metrics. Unlike previous surveys, we extend the scope to resource-constrained Internet of Things (IoT) and federated scenarios, where lightweight and tabular GANs can process sensor data and operate on edge devices. We also examine deployments in software-defined networking environments. We propose a unified evaluation framework that reports class-wise precision, recall and macro-F1-scores, per-attack metrics, computational cost, and statistical similarity tests, and we emphasize the need for interpretable and multi-modal approaches that fuse network flows with logs or threat intelligence. Emerging paradigms including GANs combined with large language models, quantum GANs, diffusion models, and reinforcement learning are surveyed, and open challenges such as training instability, mode collapse, hyper-parameter tuning, and ethical dual-use concerns are discussed. By synthesizing recent advances and outlining future research directions, this survey provides a comprehensive and forward-looking reference for practitioners and researchers developing robust, privacy-preserving, and adaptive GAN-based intrusion detection systems.

Early rehabilitation in neurocritical care is often underutilized due to fragmented workflows, interdisciplinary coordination challenges, and the absence of structured digital decision support. Traditional clinical decision support systems (CDSS) often address single domains and lack adaptability to the dynamic, multiprofessional workflows of intensive care units (ICUs). To develop and evaluate the usability of the ERATbi App (Early Recovery After Traumatic Brain Injury App), a modular, tablet-based CDSS was designed to streamline early rehabilitation planning and strengthen interdisciplinary coordination for patients with moderate-to-severe traumatic brain injury (TBI) in intensive care settings. The ERATbi app integrates four functional modules-delirium risk management, precision nutrition, stepwise early mobilization, and respiratory care for rib fractures-into a unified interface. A simulation-based usability study was conducted with 18 ICU clinicians. Evaluation metrics included System Usability Scale (SUS) scores, task completion rates, error rates, and task durations. Additional user feedback was collected via a 5-point Likert satisfaction survey and semi-structured qualitative interviews. The app demonstrated high usability (mean SUS score 83.6, SD 7.4), a 100% (18/18 participants) task completion rate, and a low error rate (4.2%). Average module completion time was 6.5 minutes, and user satisfaction was high (mean 4.7, SD 0.5). Users highlighted the value of the app's visual logic, real-time alerts, adaptive thresholds, and modular workflow integration for enhancing team coordination and decision consistency. The ERATbi app demonstrated excellent usability, high user satisfaction, and clinical relevance in simulated ICU workflows. Its logic-driven, workflow- integrated design may support scalable, interdisciplinary implementation of early rehabilitation in neurocritical care settings.