Search Results (781)

Background: Breast imaging ends in a management decision, from routine return to biopsy, yet most models are evaluated by accuracy alone and say nothing about when they may be wrong. We present a trust-aware multimodal model that reads paired mammography and ultrasound, recommends one of three BI-RADS-derived management actions, and returns to the radiologist the cases it cannot call. Methods: Two foundation encoders are adapted with low-rank adapters and combined by a mask-aware fusion head. A trustworthiness layer adds calibration, conformal prediction sets, selective deferral, and an atypicality flag. We evaluated our method on the Breast Cancer Multimodal Imaging Dataset (BCMID), comprising 332 cases from 323 patients at a single centre. The reference standard is a three-class management grouping we derive from the reporting radiologist’s BI-RADS assessment, so performance is in concordance with that derived label rather than with pathology, observed patient management or longitudinal clinical outcome. Results: Macro AUROC was 0.759 and balanced accuracy was 0.556. Isotonic calibration reduced calibration error from 0.088 to 0.052, and prediction sets reached an empirical coverage of 0.934 at a mean set size of 2.32. Deferring the least confident 30% by a retrospective ranking of the pooled cohort raised balanced accuracy to 0.631. Two of 63 positive-management cases were under-triaged and 15 routed to additional imaging, with a recall of 0.730; freezing the encoders left macro AUROC at 0.756 but raised the under-triage count to twelve. Conclusions: Error rate and error direction are separable properties, and neither accuracy nor macro AUROC records the direction. The system therefore pairs each recommendation with calibrated probabilities, a conformal set of plausible actions and an explicit defer option, so uncertain cases return to the radiologist. These results establish an internally validated operating profile for radiologist-facing support; clinical safety, deployment readiness and benefit to patients still require external and prospective evaluation. Full article

Soil salinization poses a significant threat to sustainable agricultural development and ecological security worldwide, resulting in considerable crop losses annually. The advent of drone-based hyperspectral remote sensing offers promising solutions for monitoring soil salinity, due to its high spatial resolution and versatile data acquisition capabilities. However, soil moisture alters both the scattering and absorption characteristics of electromagnetic radiation, thereby modifying soil spectral reflectance and masking salinity-related diagnostic spectral features, which can reduce the accuracy of conventional salinity estimation models. This study evaluates six spectral transformation methods—raw reflectance data (Ref), first derivative (FDR), Piecewise Direct Standardization (PDS), Orthogonal Signal Correction (OSC), FDR + PDS, and FDR + OSC—in conjunction with three machine learning algorithms: K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP). A Stacking ensemble model integrating these base learners was further developed to improve soil salinity inversion under moisture interference. The results demonstrated that the Stacking model achieved the highest accuracy and stability among the evaluated models. Additional comparisons with XGBoost and Random Forest (RF) further confirmed the competitive performance of the proposed Stacking framework. The FDR + OSC–Stacking combination achieved the best validation performance, with $R_p^2$ = 0.87, RMSEP = 0.67 mS·cm⁻¹, and RPD = 2.93. Compared with the Ref–Stacking model, $R_p^2$ increased by 0.32 (from 0.55 to 0.87), while RMSEP decreased by 0.58 mS·cm⁻¹ (from 1.25 to 0.67 mS·cm⁻¹). The results showed that PDS had limited effectiveness in correcting moisture-related spectral variation, whereas OSC more effectively mitigated moisture interference while preserving spectral information relevant to salinity estimation. Among the machine learning models evaluated, the Stacking ensemble model achieved better predictive performance than MLP, SVR, and KNN. Furthermore, the FDR + OSC–Stacking combination provided the best performance among the evaluated modeling frameworks and was successfully applied to UAV hyperspectral imagery for spatial mapping of EC_1:5. These findings demonstrate the potential of combining appropriate spectral correction with Stacking for UAV-based soil salinity assessment and provide useful technical support for site-specific salinity management in precision agriculture. Full article

Accurate multi-step traffic-flow forecasting requires dynamic spatial modeling and adaptation to heterogeneous temporal and node-level patterns. This study proposes FAMoE-ST, a hierarchically frozen attention network with hybrid-memory experts. Historical flow, temporal context, and node identity are encoded as sensor tokens and processed by a six-layer Transformer initialized from GPT-2. To avoid dependence on arbitrary sensor indexing, the causal mask is replaced by all-to-all bidirectional spatial attention and every sensor uses the same GPT position ID. The lower four blocks are frozen, while the upper two blocks are adapted and their feed-forward networks are replaced by four-expert modules. A linear router is fused with a 32-slot memory router; each token retrieves four slots and activates two experts. With 12 observations predicting the next 12 steps, FAMoE-ST achieves MAE/RMSE/MAPE of 18.55/30.35/12.91% on PEMS04 and 14.57/24.08/9.71% on PEMS08. Relative to ST-LLM, MAE decreases by 6.97% and 7.39%, respectively. Component and capacity-matched controls support the roles of restricted adaptation, sparse expert capacity, and memory routing. A separate PEMS04 initialization control finds no advantage from GPT-2 pretraining over random initialization; the contribution is therefore attributed to the proposed structural adaptation rather than to transferred linguistic knowledge. Full article

Image segmentation enables accurate object identification, a key requirement in computer vision applications. In aquaculture, this technology is essential for monitoring and management of species such as Oncorhynchus mykiss. In the Peruvian highlands, where trout farming is a vital economic activity, robust computational models are needed to automate the estimation of fish size and weight and to optimize the sustainability of production systems. This research proposes ATrUNet, a U-Net-based architecture optimized for accurate segmentation of Oncorhynchus mykiss. ATrUNet improves the information flow between the encoding and decoding layers by incorporating convolutional layers, batch normalization, and activation functions. A dataset of 1166 images was constructed, processed using LabelMe with JSON annotations, and converted into binary masks. Evaluation was conducted using loss, accuracy, and IoU. As a result, ATrUNet showed higher performance than U-Net, achieving a 27.55% reduction in loss, a 0.40% increase in accuracy (0.992), and improvements in overlap metrics such as IoU and GIoU by 2.70% and 8.75%, respectively. Future work includes expanding dataset diversity, exploring instance segmentation, and optimizing for embedded deployment. This research contributes to the application of computer vision to aquaculture, with the possibility of extending it to different species and various research contexts. Full article

Post-quantum cryptography changes the size, timing, and algorithmic context of Transport Layer Security (TLS) handshakes, creating a dynamic normal class for anomaly detectors. This study evaluates an assurance framework, rather than proposing a new classifier, on CIC-PQC_OAV v1: 40,010 sessions represented by 32 encrypted-metadata features. An audit finds 185 exact fingerprints shared across the supplied partitions, affecting 1594 sessions, plus one exact conflicting-label group. We compare the fixed split with five-seed, size-matched sample-stratified, exact-disjoint, raw round-3-disjoint, and IQR-normalized round-3-disjoint protocols, each separating fitting, probability calibration, threshold selection, conformal calibration, and testing. The fixed-split LightGBM $F_1$ is 0.9013; controlled five-seed means cluster at 0.8872–0.8898, showing that its gap is not attributable solely to duplicate control. Condition-disjoint tests reveal heterogeneous anomaly transfer and false-positive rates of 0.9998, 0.9897, and 0.6107 for three unseen valid families. Isotonic calibration yields a Brier score of $0.0221\pm 0.0013$, while five-seed perturbations confirm sensitivity to timing masking and byte scaling. The results support layered, dataset-bounded evaluation combining fingerprint independence, condition holdouts, calibration, selective review, robustness, and explanation. Full article

Large-scale public events generate concentrated passenger demand during egress periods, often overwhelming urban transit systems. This paper proposes a multi-modal evacuation framework that coordinates in-service buses temporarily diverted from existing lines and dedicated shuttle vehicles pre-positioned at depots. The problem is formulated as a two-layer stochastic optimization under travel time uncertainty: the upper layer determines pre-event shuttle fleet sizing, while the lower layer makes real-time dispatching decisions for both modes. We propose an Uncertainty-Aware Reinforcement Learning framework with Categorical DQN (UARL-CD) that learns a robust dispatching policy through a reward function aligned with the lower-level objective, explicitly accounting for travel time uncertainty via distributional value representation and stochastic training, with an action masking mechanism enforcing operational constraints. Simulation experiments based on a realistic stadium evacuation scenario demonstrate that the proposed framework significantly outperforms deterministic optimization and rule-based strategies, achieving a 31.6% reduction in evacuation completion time and a 48.4% reduction in average passenger waiting time compared to shuttles alone, while maintaining robustness to travel time uncertainty with only 4.0% performance degradation and online decisions executed within the 2-min decision interval. Full article

Fetal electroencephalogram (fEEG) recording could address the gap of functional fetal brain testing and enable direct assessment of fetal brain status during pregnancy and labor. However, the fetal EEG signal (~1 µV) is masked by roughly 80 dB (about 10,000×) of maternal ECG and other bioelectric interferences. This synthetic-data feasibility study presents whether the fetal EEG can be recovered from the maternal abdomen surface by using the following three sensing architectures on physiologically scaled synthetic data: single electrode, dual electrode, and a concentric 28-electrode array centered over the fetal head. We show that exploiting a known geometric attenuation steering vector with a minimum-variance distortionless-response (MVDR) beamformer raises recovery correlation from ~0 to ~0.28 and improves signal-to-noise ratio (SNR) by ~67 dB. A follow-up validation replaces the synthetic EEG generator with actual neonatal EEGs recorded from the scalp, obtained from OpenNeuro ds004577 and the Helsinki Zenodo corpus. Comparable recovery performance across both datasets confirms that the synthetic feasibility conclusion generalizes to real neonatal EEG composition and behavioral characteristics. The primary evaluation is based on waveform-level engineering metrics (correlation, SNR, and RMSE), while exploratory secondary analyses assess aEEG envelopes, band-power trends, and burst detection. These analyses are not intended as clinical validation. Independent per-electrode sensor noise remains the dominant residual limiter under the homogeneous geometric model. The reported r ≈ 0.28 should not be interpreted as expected in vivo performance at all gestational ages: an intact vernix layer can add approximately 35 dB of attenuation and makes the required noise floor substantially more stringent. This analysis adds a five-layer volume-conductor analysis, misalignment, depth and impedance-drift stress tests, recording-level statistics, comparisons with classical extraction families (PCA, ICA, adaptive cancelation, and multichannel Wiener filtering), an interference-alignment sensitivity analysis, and exploratory biomarker-level analyses, which together bound the idealizations of the forward model. Full article

Contemporary warehouses operating under the Warehouse 4.0 paradigm are complex cyber-socio-technical systems, in which risk management maturity assessment is typically conducted in isolation from its impact on the resilience of the intralogistics handling process. The aim of this article is to develop the conceptual framework of a three-layer RMMA-W4.0 method, supporting the resilience of such systems by integrating a map of risk sources, a maturity assessment model covering seven risk management dimensions, and six resilience dimensions of the intralogistics handling process. These dimensions were linked through cross-mapping analysis, enabling a resilience profile to be derived from a maturity profile. The application of the method is illustrated through a case study of a highly automated AS/RS-class warehouse. The results showed that an aggregated maturity index may mask a significant imbalance in the profile—a high level of technical maturity did not translate into most resilience dimensions, particularly degraded-mode operation and recovery, which made it possible to formulate targeted improvement recommendations. The conclusions indicate that decomposing the assessment into a maturity profile and a resilience profile derived from it enables the identification of areas requiring priority reinforcement, while a full gap analysis remains a direction for future research. Full article

Unmanned Aerial Vehicle-borne Ground-Penetrating Radar (UAV-GPR) has been used for ice thickness monitoring in lakes and rivers due to its non-contact measurement, high resolution, and operational flexibility. Existing algorithms can extract ice layer boundaries by tracking continuous bottom reflections in GPR images. However, they fail when the radar signal lacks a clear bottom reflection—a common condition in ice layers containing unfrozen water—and manual interpretation remains time-consuming. To address this limitation, this paper builds a freshwater ice GPR dataset covering both fully frozen and unfrozen water-bearing zones, and proposes a method for ice thickness identification based on the DeepLabv3+ neural network. The model performs pixel-level binary classification, labeling each pixel as ice layer or background, and generates a segmentation mask that constrains the subsequent thickness calculation to valid ice regions only. Field validation against drilling measurements demonstrates that the model achieves Intersection over Union (IoU) of 97.12% and an F1-score of 98.54% for ice layer identification, with a relative error in ice thickness measurement below 3% based on five borehole measurements. Field tests in two reservoirs across Tibet and Jilin, China, demonstrate that the proposed method can accurately characterize the distribution and thickness of the ice layer while effectively eliminating the interference of unfrozen water zones. The results demonstrate that the proposed method can provide automated, accurate ice thickness estimates for UAV-GPR surveys of freshwater ice. Full article

Advancements in intelligent manufacturing require solutions to the conventional flexible job-shop scheduling problem (FJSP) to accommodate increasingly intricate constraints, particularly in the semiconductor and electronic component sectors, where batch-processing machines (BPMs) significantly intensify scheduling complexity. To address this challenge, this study formulates an extended FJSP with multiple BPMs and proposes an end-to-end two-layer multi-agent deep reinforcement learning framework. Job and machine agents perform decentralized action mapping, while workshop states are encoded using a heterogeneous disjunctive graph and a dual-graph attention network. Unlike standard FJSP learning methods that primarily address operation–machine decisions, the proposed framework jointly models machine assignment, operation sequencing, variable-length batch formation, and BPM allocation within a unified policy, with a pointer network-based batching agent and an equipment-selection agent that handle batch-processing decisions under feasibility masking. The framework was validated using plant-derived production data and multi-scale synthetic instances. Numerical results show that the proposed method achieves competitive performance across the tested batching and standard-FJSP settings. In standard-FJSP comparisons, relative performance was scenario-dependent: DANIEL performed better in S1, whereas both proposed variants outperformed all comparators in S2. These results support the framework as an effective scheduling approach for deterministic FJSP with BPMs and indicate cross-scale generalization across evaluated instances. Full article

This study develops a SCHISM-based hydrodynamic model and an offline Lagrangian virtual-particle workflow for the Zadar Channel, a geometrically complex island–mainland passage in the eastern Adriatic. Independent hourly observations from the MP Zadar tide gauge operated by the Hydrographic Institute of the Republic of Croatia (HHI) were used to evaluate the modelled free-surface response. After exclusion of the first 24 h ramping period, 192 matched hourly pairs gave a Pearson correlation of 0.913, a mean bias of 0.012 m, a mean absolute error of 0.051 m, and a root-mean-square error of 0.070 m; cross-correlation was maximized at zero lag. The model reproduced the timing of the observed oscillations but underestimated their amplitude, with simulated and observed standard deviations of 0.117 and 0.157 m, respectively. The adopted unstructured mesh contains 16,962 triangular elements and 9081 nodes. In four 24 h particle-sensitivity tests, maximum reach ranges from 8.4 to 14.2 km; a 15-fold change in horizontal diffusivity affects reach less than sampling a lower model layer, which reduces reach by 32.8%. In the June 2025 event calculation, cumulative numerical shoreline contact increases from zero to all 1000 particles. The approximately 4 km Copernicus regional product masks the narrow interior passages and is therefore used only to assess spatial representativeness, not to validate channel currents. The tide-gauge comparison supports the modelled sea-level response and its timing at one station, but does not constitute direct validation of local current velocities. The reported trajectories are current-driven passive-particle diagnostics; wave–current coupling, Stokes drift, and material-specific fate processes are not represented. Full article

Floods degrade road networks at the same time as demand for emergency medical services (EMSs) rises, yet national EMS command systems rarely receive any information on flood-induced road barriers. This paper presents Ambulance STARS (SaTellite-assisted Ambulance Routing System), a service-oriented framework that links satellite observation with ambulance dispatch. A cloud-based flood detection service derives flood extent from Sentinel-1 SAR amplitude change detection executed in a cloud-based Earth observation data and compute backend and translates it into road passability layers. A routing engine then maintains an in-memory road graph whose travel times are calibrated with empirical ambulance speed models built from four years (2020–2023) of GPS records of an EMS fleet in southern Poland, with separate speeds for driving with and without emergency signals (61.8 and 37.2 km/h, respectively). An API gateway with single-file tile delivery, a replicated relational data tier, and an observability stack complete the architecture, and a web client offers dispatchers live routing and multi-unit incident simulation. The framework was tested on the September 2024 flood in the Municipality of Nysa, Poland. The SAR module delineated 665 ha of inundation and marked 8.5 km of the 656.7 km routing network as impassable (508 barrier points), and the same procedure applied to a reference optical mask of 18 September yielded 17.3 km and 1006 points. Because the SAR and optical acquisitions captured different phases of the flood wave, agreement on the rare impassable-road class was low, and the two products were, therefore, used to bracket operational uncertainty rather than to define a single ground truth. Applied without local retuning to Lewin Brzeski, the same flood detection workflow showed consistent performance against the CEMS reference product. The routing module produced statutory 8/15/20 min accessibility maps in 12–34 s under warm-cache benchmark conditions. With SAR-derived barriers, the share of the network reachable within 15 min fell from 88% to 80%, and 2 villages with 938 inhabitants lost road access to EMS entirely. With barriers derived from the optical mask, the 15 min share fell to 39.8% and seventeen settlements lost road access entirely, underlining how strongly the barrier source shapes the operational picture. Post-acquisition processing completes in under one minute under warm-cache conditions with road data preloaded, and satellite-derived road passability is fast enough to support near-real-time decision-making, subject to the constellation revisit time and to integration with EMS command systems. Full article

Existing readiness indices for Connected and Automated Vehicles (CAVs) produce a single composite score per city. This conflates use cases with very different territorial requirements and, being compensatory, allows strong performance on one criterion to mask a deficiency that is, in practice, a hard prerequisite. This paper proposes and demonstrates an assessment framework with four elements. First, the verdict is stratified by Operational Design Domain (ODD) profile, so that one territory yields one verdict per use case. Second, each profile is assessed only on its eligible service area, the roads on which the corresponding service could actually operate. Third, a non-compensatory veto layer is overlaid on a compensatory segment-level score, which separates the readiness score from the fit share, the proportion of network length that satisfies every critical threshold. Fourth, the diagnosis is extended to specific corridors through a readiness-aware routing formulation. The framework is demonstrated on the Versailles, Satory, and Velizy perimeter of about 92 square kilometes, using a real OpenStreetMap network of 4834 nodes and 733 kilometres, combined with an explicitly illustrative synthetic indicator calibration. Restricting each profile to its service area changes one verdict outright. The campus shuttle and urban robotaxi profiles obtain near-identical scores, both close to 0.51, but fit shares differing seven-fold, a distinction that a single score cannot express. Readiness-aware routing reroutes up to 42 percent of origin and destination pairs onto better-equipped itineraries for about 1 percent additional travel length. Monte Carlo analysis and cross-method tests using TOPSIS and PROMETHEE-II leave all verdicts unchanged. We position the work as a methodological demonstration rather than a validated deployment tool: the indicator values are synthetic, and empirical calibration, expert weight elicitation, and field validation remain prerequisites for operational use. Full article

Plantations play an essential role in regional ecological security and the sustainable use of forest resources, making accurate knowledge of their spatial distribution and management boundaries critical for forest inventories, afforestation assessment, carbon-accounting support, and stand-level management. To address blurred stand boundaries, complex background interference, and regional spectral variation, we propose a plantation segmentation framework termed the Kernel-Guided Forest Segmentation Network (KGF-Net). The method combines a spectral-index aligned fusion module (SIAF), a structure-aware boundary enhancement module (SABE), a cross-domain discriminative geometry-constrained coupling module (CDGC), and an Adam-compatible forest-aware responsive spectral control strategy (FReSCO). In the source-domain stage, KGF-Net is pretrained on the North American subset using aligned RGB, NDVI, and EVI inputs, enabling the model to learn vegetation-sensitive spectral and structural representations. For Asia, Europe, Africa, and public FAIS target datasets, no target-domain vegetation-index layers are used; instead, the source-domain weights are transferred and the model is fine-tuned separately using target-domain RGB images and plantation masks. Thus, the cross-region experiments evaluate source-to-target transfer learning with RGB-based target adaptation rather than zero-shot generalization or independent multimodal training in every region. Experiments show that KGF-Net achieves competitive segmentation and boundary-delineation performance while maintaining a compact computational profile. The resulting maps provide spatially explicit information on plantation extent, patch configuration, and management boundaries, supporting regional plantation monitoring and forest-management applications. Full article

Objective: To develop and evaluate EDBERT (Emergency Department Bidirectional Encoder Representations from Transformers), a BERT-based architecture built around a multi-input fusion network that enriches free-text triage notes with the additional context of presenting complaints and patient age to predict Emergency Department (ED) disposition, supporting early decision-making and reduced ED length of stay (LOS). Methods: A retrospective cohort of 570,143 ED presentations to the Royal Melbourne Hospital, Australia, was used. At the core of EDBERT, a fusion network integrates three complementary input sources—triage notes, presenting complaints, and patient age—through learnable weights, with the two free-text inputs encoded by customised eight-layer BERT encoder stacks pre-trained on triage text with Masked Language Modelling (MLM). The resulting architecture is also lightweight, containing 66 million parameters, substantially fewer than the 110 million of BERT-BASE. Eighty percent of the dataset was used for training, and twenty percent for testing. Performance was benchmarked against BERT-Base configurations of varying depth. Results: EDBERT achieved an accuracy of 83.26%, a macro-averaged F1-score of 81.99%, and an area under the receiver operating characteristic curve of 0.91, outperforming all single-input BERT-Base variants, including a depth-matched eight-layer model (82.08%), indicating that the gain derives from the fusion of complementary inputs combined with domain-adaptive pre-training rather than from model capacity. Conclusions: A fusion-based BERT architecture that supplements the triage narrative with presenting complaint and age context predicts ED disposition more accurately than larger single-input BERT models, while requiring substantially less memory and computation, making it practical for deployment in hospitals with limited computing resources. Full article