Search Results (63)

Automated analysis of peripheral nerve ultrastructure is bottlenecked by heterogeneous electron microscopy (EM) datasets, where varying staining protocols and resolutions create domain shifts that confound deep learning. To address this, we developed a generalized segmentation pipeline. Using a custom pre-processing workflow (CLAHE and noise suppression) integrated into ZEISS Arivis Pro, we standardized inputs across three disparate domains: traditional osmium-based Palade, lanthanide-based “green” Uranyl-free method, and low-resolution Ellisman preparations. A U-Net trained on a highly constrained 15-image composite dataset achieved peak internal Intersection over Union (IoU) scores >0.95 for myelin and Schwann cells. Crucially, during open-world, zero-shot inference on an expanded independent testing cohort (N = 40), the model sustained robust Dice Similarity Coefficients of 0.854 for myelin and 0.597 for mitochondria. This demonstrates that integrating classical image standardization with deep learning effectively mitigates EM domain gaps, enabling comprehensive 3D multi-organelle reconstructions from challenging data. To ensure transparency and community utility, the pre-trained models and standardization scripts are provided in a public, open-access repository. Ultimately, this pipeline supports the transition to sustainable, non-toxic EM protocols and provides a robust pathway for unlocking historical clinical archives for automated organellomics. Full article

This work investigates a hardware-aware, circuit-level emulation of BB84-like circuit primitives on noisy intermediate-scale quantum (NISQ) processors. The motivation is to evaluate whether BB84-like basis sifting and intercept–resend-induced QBER behavior remain observable when selected BB84 operations are mapped to parallel single-qubit circuits on gate-based devices. The proposed mapping represents Alice’s preparation, optional Eve intercept–resend emulation, and Bob’s measurement as processor-internal circuit layers; it is therefore an on-chip emulation and not an end-to-end optical QKD implementation. Experiments combine real IBM superconducting processors with Qiskit, Cirq, and Azure/Q# simulator-based or noise-modeled evaluations. Baseline QBER was first calibrated for each backend, and intercept–resend experiments then produced a clear QBER separation from the no-eavesdropper condition. The observed sifted-bit utilization was close to the expected 50% BB84 basis-matching reference, while the constant-depth circuit structure supported scalable raw/sifted-bit generation before any classical post-processing. These observations are treated as implementation-level consistency checks and backend-dependent experimental metrics, rather than as new BB84 protocol-level results. Finite-shot uncertainty, calibration drift, and backend-specific noise are treated as limitations of the proposed QBER-based evaluation rule rather than as deployment-level security guarantees. Because the study does not implement a physical quantum channel, authenticated classical communication, error correction, privacy amplification, finite-key security analysis, or general QKD attack models, the reported metrics should be interpreted as raw/sifted-bit experimental metrics and QBER-based disturbance evaluation for BB84-like NISQ emulation, not as secure key rates, secure throughput, or practical QKD deployment results. Full article

Identifying village functional differentiation within rural territorial systems is essential for differentiated rural revitalization and place-based governance. However, existing approaches still lack effective analytical pathways for translating complex rural territorial relations and sparse planning labels into fine-grained measures of rural functional intensity. To address these gaps, this study develops a Few-Shot Hierarchical Graph Representation Learning (FH-GRL) framework. By integrating a Hierarchical Graph Infomax (HGI) model to capture cross-scale village–township–city relational dependencies and an Evidential Deep Learning (EDL) mechanism to map high-dimensional representations into class-specific evidence and Global Percentile Ranks (GPR), the framework supports fine-grained classification and continuous grading of rural functions. Empirical analysis in Pingdingshan City yields three main findings. First, within the present case study, FH-GRL shows more stable performance than traditional flat clustering and local graph models in identifying complex rural functions under limited labeled samples. Second, hierarchical context serves as a spatial calibration mechanism, reducing locally generated noise and improving the identification of village functional differentiation under spatial heterogeneity. Third, rural functional differentiation reflects the combined effects of place-based conditions and potential flow-related interaction conditions. In particular, Center villages show differentiated trajectories between endogenous production or service centers in agricultural plains and exogenous service centers along urban development axes. Overall, this study provides a planning-oriented quantitative framework for diagnosing rural functional differentiation under label scarcity and spatial heterogeneity. The GPR-based outputs can support the identification of high-intensity functional carriers, transitional villages, and general reserve areas, thereby providing diagnostic evidence for differentiated governance and tiered resource allocation. Rather than replacing formal planning judgment, the framework offers geospatially informed support for classified rural governance and more evidence-informed territorial planning. Full article

Landslide detection has largely relied on supervised learning or DEM-based representations, which can limit rapid deployment and generalization across heterogeneous terrain. In this work, we present a zero-shot, fully unsupervised framework that identifies landslide-like geomorphic instability candidates from raw UAV-mounted LiDAR, removing the need for labeled data, pre-event baselines, or rasterized terrain abstractions. Our approach is motivated by the observation that landslides manifest as localized geometric inconsistencies in the terrain surface. We capture this through a multi-scale formulation that combines point-level and cluster-level indicators of instability. At the point level, a PCA-based residual depth metric reduces slope-induced bias and highlights surface discontinuities, while local concavity captures terrain depletion patterns. At the cluster level, geomorphometric descriptors such as curvature concentration, surface roughness, elevation discontinuity, and slope variation are extracted using density-aware 3D clustering and integrated through adaptive feature fusion. The resulting probabilistic instability field enables spatially coherent delineation of landslide scars, including rupture boundaries, displaced material, and emerging failure regions. In addition, the detected patches provide useful priors for post-event susceptibility analysis without requiring temporal observations. Experiments across diverse geomorphic settings show that the proposed method improves detection of subtle terrain disturbances compared to DEM-based pipelines and supervised learning approaches, while remaining robust to noise and terrain variability. Overall, this work demonstrates that geometry-driven, unsupervised inference on raw 3D data can serve as a practical and scalable alternative for near real-time landslide detection using UAV-based systems. Full article

Reliable evaluation resources for semantic deduplication, semantic textual similarity (STS), and retrieval remain limited for low-resource and morphologically rich languages. This study introduces KazakhTextDuplicates, a Kazakh-language benchmark for controlled evaluation of semantic duplication under varying levels of semantic preservation and surface-form distortion. The benchmark includes two complementary versions. KazakhTextDuplicates v1.0 is a diagnostic dataset derived from naturally occurring duplicate and near-duplicate pairs for analyzing relationships between duplication labels and lexical overlap. KazakhTextDuplicates v2.0 is a controlled benchmark created using deterministic transformations that define seven semantic duplication regimes: exact duplication, contextual reformulation, paraphrasing, partial semantic overlap, and three levels of character-level noise. Each pair is assigned both a regime label and a predefined similarity score, enabling evaluation across duplicate classification, STS, and retrieval tasks. Five embedding models (BGE-M3, LaBSE, Multilingual-E5-Large, KazEmbed-v5, and OpenAI text-embedding-3-large) were evaluated in a zero-shot setting. Results show that v1.0 is strongly affected by surface-form similarity and is therefore more suitable for diagnostic analysis. In contrast, v2.0 provides a more challenging and informative evaluation environment. OpenAI text-embedding-3-large achieved the strongest STS performance (Pearson = 0.510, Spearman = 0.652, MSE = 0.075) and the best duplicate regime classification results (Accuracy = 0.155, Macro-F1 = 0.066). Retrieval performance remained strong at higher cutoffs despite lower first-rank stability. The results demonstrate that benchmark design substantially affects semantic similarity evaluation and emphasize the need for controlled assessment in low-resource agglutinative languages. Full article

Speech-based non-invasive screening offers a cost-effective and scalable approach for the early detection of Alzheimer’s disease (AD). However, the clinical utility of deep learning models remains severely constrained by the scarcity of labeled speech data in low-resource languages, necessitating cross-lingual transfer learning. Conventional domain adaptation paradigms typically assume semantically consistent feature domains and focus heavily on aligning marginal distributions; however, they suffer catastrophic performance degradation when applied to cross-lingual pathologic speech. By analyzing disease-associated representation vectors within a self-supervised HuBERT space, we uncover a systematic mechanism driving this failure, a phenomenon we term cross-lingual polarity flip, where the direction of disease-relative-to-control feature offsets fundamentally reverses between languages. While prior multilingual studies have largely discarded such dimensional inconsistencies as ungeneralizable noise, a 500-round Monte Carlo stability analysis demonstrates that these flips occur in a highly stable, structural manner across 18.3% of top discriminative dimensions. Leveraging this insight, we introduce Monte Carlo Polarity Flip Calibration (MC-PFC), a few-shot framework designed to explicitly rectify flip orientations. Requiring only five labeled support samples per class from the target domain, MC-PFC robustly estimates direction flips via a separability-weighted ensemble voting mechanism. Evaluated on a strictly held-out Chinese blind test set, MC-PFC achieves an area under the receiver operating characteristic curve (AUC) of 0.871, recovering 99.5% of the performance achieved by a full in-domain trained upper bound (AUC = 0.875). Ablation experiments confirm that direction calibration yields a substantial +0.361 AUC gain, vastly outperforming standard distribution alignment (+0.081). This work establishes a data-efficient paradigm for cross-lingual medical analysis, shifting the clinical AI focus from discarding cross-lingual discrepancies to actively modeling and calibrating them. Full article

Few-shot fine-grained image classification aims to recognize subcategories with high visual similarity using only a limited number of annotated samples. Existing metric learning-based methods typically rely solely on spatial-domain features. Confined to this single perspective, models inevitably suffer from inherent texture biases, entangling essential structural details with high-frequency background noise. Furthermore, lacking cross-view geometric constraints, single-view metrics tend to overfit this noise, resulting in structural instability under few-shot conditions. To address these issues, this paper proposes the Dual-Subspace Network (DSNet). Specifically, DSNet utilizes the discrete cosine transform (DCT) and a low-pass filtering mechanism to explicitly isolate low-frequency global structural components from spatial features, thereby suppressing background interference. Truncated Singular Value Decomposition (SVD) is employed to construct independent, low-rank linear subspaces for both spatial texture and frequency structural features. An adaptive gating mechanism is designed to dynamically fuse the projection distances from these dual views. This strategy leverages the structural stability of the frequency subspace to prevent the spatial subspace from overfitting to background features. Extensive experiments on four benchmark datasets—CUB-200-2011, Stanford Cars, Stanford Dogs, and FGVC-Aircraft—demonstrate that DSNet exhibits excellent classification performance and robustness, achieving highly competitive results compared to existing metric learning algorithms. Complexity analysis further confirms that the proposed network achieves a favorable balance between high accuracy and computational efficiency, providing an effective new paradigm for few-shot fine-grained visual recognition. Full article

Intelligent fault diagnosis of rotating machinery is essential for manufacturing reliability and predictive maintenance, yet deployment of deep learning models is limited by data scarcity: fault samples are rare, costly, and hazardous to obtain. Conventional synthetic data methods such as Generative Adversarial Networks and Variational Autoencoders often exhibit mode collapse, spectral distortion, and limited physical interpretability. This work presents MechaForge, a multi-strategy framework that employs Large Language Models (LLMs) as physics-guided generators for bearing fault time-series data. The approach is grounded in bearing kinematics, Motor Current Signature Analysis (MCSA), and the interpretation of in-context learning as implicit Bayesian inference. Within MechaForge, four progressively constrained tracks are defined: a real-data baseline, few-shot LLM mimicry, multi-stage semantic reasoning, and physics-guided generation with constraints on root mean square, kurtosis, and fault-band spectral energy. For direct benchmarking, conventional VAE- and GAN-based augmentation baselines are additionally evaluated under the same dataset split, synthetic-data budget, downstream CNN architecture, and evaluation metrics. Experiments on the Paderborn bearing dataset show that the Basic LLM track achieves the strongest performance under the present protocol (0.7862 accuracy, 0.7648 macro-F1), exceeding the added VAE and GAN baselines (both 0.7428 accuracy; 0.7202 and 0.7257 macro-F1, respectively), while a control experiment confirms that synthetic data provides discriminative structure rather than labeled noise. These results indicate the promise of LLM-based diagnostic augmentation under data scarcity in the present Paderborn setting, rather than a definitive demonstration of broad transferability across fault-diagnosis scenarios. Full article

Background/Objectives: To evaluate and compare the qualitative and quantitative image performance of multiplexed sensitivity-encoding diffusion-weighted imaging (MUSE-DWI) against conventional single-shot (ss-DWI) and high-resolution single-shot (HR-ssDWI) sequences in patients with esophageal cancer. Methods: Twenty patients who underwent esophagus MRI, including ss-DWI, HR-ssDWI and MUSE-DWI, were retrospectively enrolled. Image quality, esophageal contour, lesion conspicuity and image distortion were independently graded by two radiologists using a five-point scale and compared between the three sequences. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of esophageal tissue were measured and compared between the three sequences. Results: After Bonferroni correction (p < 0.017), MUSE-DWI had significantly higher scores than HR-ssDWI in image quality, esophageal contour delineation and lesion conspicuity, and all three sequences had statistically significant differences in image distortion scores with MUSE-DWI performing the best. Quantitative analysis revealed that MUSE-DWI had the highest SNR and CNR values; significant differences were found in SNR between ss-DWI and HR-ssDWI (p < 0.001), and in both SNR and CNR between HR-ssDWI and MUSE-DWI (p < 0.001), while no significant differences were observed in SNR and CNR between ss-DWI and MUSE-DWI (p > 0.017). Conclusions: MUSE-DWI outperforms ss-DWI and HR-ssDWI in reducing image distortion, with comparable quantitative image quality metrics to ss-DWI. It represents a valuable optimized DWI technique for esophageal clinical imaging. Full article

The accurate and timely identification of urban building footprints is critical for sustainable urban planning and disaster management. Traditional remote sensing methods for this task often face limitations in scalability, accuracy, and adaptability to complex urban morphologies. This paper addresses these challenges by developing and evaluating a novel data-centric framework that synergistically integrates Graph Neural Networks (GNNs) with zero-shot superpixel segmentation derived from the Segment Anything Model (SAM) applied to Sentinel-2 imagery. A cornerstone of our methodology is a rigorous assessment of OpenStreetMap (OSM) data, refined through temporal NDVI stability analysis to generate high-quality ground truth. We propose an optimized UrbanGraphSAGE model, enhanced with spectral data augmentation and trained using a robust loss function with label smoothing to mitigate label noise. In the complex urban landscape of Algiers, Algeria, our approach achieves a Test F1-Score of 0.7131, demonstrating highly competitive performance with standard pixel-based baselines like U-Net while offering significant topological and computational advantages. Specifically, our model operates with merely 19,585 parameters—orders of magnitude fewer than pixel-based CNNs. A rigorous Gold Standard evaluation against manually labeled imagery confirms the model’s high recall (0.8484) and reliability for automated urban monitoring. Full article

This study presents an integrated optimization framework that combines the Design of Experiments (DOE) approach with Machine Learning (ML) techniques to minimize defects in the injection molding of Fuel Vapor Charcoal Canisters. The research focuses on five critical process parameters—melt temperature, mold temperature, filling time, pressure holding time, and pure cooling time—whose combined influence on major molding defects (warpage, shrinkage, shear stress, residual stress, and short shots) was systematically investigated. A Taguchi L25 orthogonal array was employed to structure the experiments and identify the optimal parameter levels through signal-to-noise (S/N) ratio analysis using the “smaller-the-better” quality criterion. The Taguchi results revealed that pressure holding time was the most influential factor, followed by mold temperature and melt temperature. Simulation results from SolidWorks Plastics confirmed the reduction in major defects under the optimized settings. To further validate and generalize the DOE findings, a Random Forest regression model was trained on the same dataset to capture nonlinear interactions between parameters. The model achieved an average RMSE of 2.451 ± 0.591 in five-fold cross-validation, demonstrating strong predictive accuracy. Feature importance analysis indicated that pressure holding time accounted for approximately 77.5% of the variance in the defect index, reaffirming its dominant role. A 3D response surface of the global parameter space (mold temperature vs. pressure holding time) revealed a distinct minimum defect region, consistent with the DOE-optimized settings. The Taguchi analysis identified the optimal parameter settings as Melt Temperature at Level 2, Mould Temperature at Level 3, Filling Time at Level 4, Pressure Holding Time at Level 5, and Pure Cooling Time at Level 4, which collectively produced the highest S/N ratios and the lowest overall defect index. The overall discrepancy between DOE and ML predictions was only 12.5%, confirming methodological consistency. The integration of DOE and ML not only enhances parameter interpretability and defect prediction accuracy but also provides a scalable, data-driven approach for intelligent process control and quality assurance in automotive injection molding. Full article

Sand production in an offshore hydrocarbon wells poses significant operational and integrity challenges, particularly in deviated wells, where complex flow geometries intensify particle transport and erosion risks. The traditional sand-monitoring method utilizes stationary acoustic sensors attached to the production flowline at the surface. However, these sensors provide limited spatial coverage and intermittent measurements, restricting their ability to detect early sanding onset or precisely localize sanding intervals. By combining with vertical seismic profiling (VSP), Distributed Acoustic Sensing (DAS) delivers continuous, high-density data along the entire length of the wellbore and is increasingly recognized as a powerful diagnostic tool for real-time downhole monitoring. This study presents a field application of DAS-VSP for detecting and characterizing sand transport in a deviated offshore production well equipped with 350 distributed fiber-optic channels spanning 0–1983 m true vertical depth (TVD) at 8 m spacing. A multistage workflow was developed, including SEGY ingestion and shot merging, channel and time window selection, trace normalization, and low-pass filtering below 20 Hz. Multi-domain signal analysis, such as RMS energy, spike-based time-domain attributes, FFT, PSD spectral characterization, and time–frequency decomposition, were used to isolate the characteristic im-pulsive low-frequency (<20 Hz) signatures associated with sand impact. An adaptive thresholding and event-clustering scheme was then applied to discriminate sanding bursts from background noise and integrate their acoustic energy over depth. The processed DAS section revealed distinct, depth-localized sand ingress zones within the production interval (1136–1909 m TVD). The derived sand log provided a quantitative measure of sand intensity variations along the deviated wellbore, with normalized RMS amplitudes ranging from 0.039 to 1.000 a.u., a mean value of 0.235 a.u., and 137 analyzed channels within the production interval. These results indicate that sand production is highly clustered within discrete depth intervals, offering new insights into sand–fluid interactions during steady-state flow. Overall, the findings confirm that DAS-VSP enables continuous real-time monitoring of the sanding behavior with a far greater depth resolution than conventional tools. This approach supports proactive sand management strategies, enhances well-integrity decision-making, and underscores the potential of DAS to evolve into a standard surveillance technology for hydrocarbon production wells. Full article

Continuous-variable quantum key distribution (CV-QKD) enables information-theoretically secure key generation between legitimate parties. To further enhance system performance, an improved two-way CV-QKD protocol has been proposed, which is accessible in practice and exhibits increased robustness against excess noise. However, in practical implementations, device nonidealities inevitably drive the optical field from the single-mode regime into the continuous-mode regime. In this work, we introduce temporal modes to characterize the evolution of optical fields in the improved two-way protocol and establish a security analysis framework for the continuous-mode scenario based on adaptive normalization with calibrated shot-noise unit. In addition, finite-size effects are taken into account in the analysis. Our results demonstrate that the improved two-way protocol retains a performance advantage over its one-way counterpart. The analysis provides useful guidance for the practical implementation and performance optimization of improved two-way CV-QKD systems. Full article

Stance detection has emerged as an essential tool in natural language processing for understanding how individuals express agreement, disagreement, or neutrality toward specific targets in social and online discourse. It plays a crucial role in bilingual and multilingual environments, including English-Arabic social media ecosystems, where differences in language structure, discourse style, and data availability pose significant challenges for reliable stance modelling. Existing approaches often struggle with target awareness, cross-lingual generalization, robustness to noisy user-generated text, and the interpretability of model decisions. This study aims to build a reliable, explainable target-aware bilingual stance-detection framework that generalizes across heterogeneous stance formats and languages without retraining on a dataset specific to the target language. Thus, a unified dual-encoder architecture based on mDeBERTa-v3 is proposed. Cross-language contrastive learning offers an auxiliary training objective to align English and Arabic stance representations in a common semantic space. Robustness-oriented regularization is used to mitigate the effects of informal language, vocabulary variation, and adversarial noise. To promote transparency and trustworthiness, the framework incorporates token-level rationale extraction, enables fine-grained interpretability, and supports analysis of hallucination. The proposed model is tested on a combined bilingual test set and two structurally distinct zero-shot benchmarks: MT-CSD and AraStance. Experimental results show consistent performance, with accuracies of 85.0% and 86.8% and F1-scores of 84.7% and 86.8% on the zero-shot benchmarks, confirming stable performance and realistic generalization. Ultimately, these findings reveal that effective bilingual stance detection can be achieved via explicit target conditioning, cross-lingual alignment, and explainability-driven design. Full article

This article discusses the fundamental limitations of Light Fidelity (Li-Fi) systems, an emerging visible light communication technology that is constrained by line-of-sight dependency and optical attenuation. Unlike existing adaptive modulation approaches that focus solely on improving signal processing, we present an integrated framework that combines three key contributions: (1) an adaptive modulation optimization algorithm that selects among OOK, PAM, and OFDM schemes based on instantaneous signal-to-noise ratio thresholds, achieving a 30–40% range extension compared to fixed modulation references; (2) a method for spatial optimization of access points (APs) using the L-BFGS-B algorithm to determine the optimal location of APs, taking into account lighting constraints and coverage uniformity; and (3) comprehensive system-level modeling incorporating shot noise, thermal noise, inter-symbol interference, and dynamic shadowing effects for realistic performance evaluation. Through extensive simulations on multiple room geometries (6 m × 5 m to 20 m × 15 m) and AP configurations (one to six APs), we demonstrate that the proposed adaptive system achieves an average throughput 60% higher than that of fixed OOK, while maintaining 98.7% coverage in a 10 m × 8 m environment with two optimally placed APs. The framework provides practical design guidelines for Li-Fi deployment, including an analysis of computational complexity $O(M\times N)$ for coverage assessment, $O(I\times D^3)$ for access point optimization) and a characterization of convergence behavior. A comparative analysis with state-of-the-art techniques (optical smart reflective surfaces, machine learning-based blockage prediction, and Li-Fi/RF hybrid configurations) positions our lightweight algorithmic approach as suitable for resource-constrained deployment scenarios, where system-level integration and practical feasibility take precedence over innovation in individual components. Full article