Search Results (48,729)

Background/Objectives: Large language models (LLMs) are increasingly explored for radiology education, but their role in differential diagnosis learning remains under-investigated. This study evaluates the perceived usefulness of LLM-generated differential diagnoses compared with a standard radiology casebook. Methods: In this multi-center study, radiology trainees at junior (years 1–2) and advanced (years 3–5) levels evaluated 225 cases from a gold-standard casebook spanning nine subspecialties. Participants ranked the usefulness of their personal clinical experience, the casebook, and LLM teaching, and rated the LLM output using a five-point Likert scale across Clarity, Trust, Differential Usefulness, and Diagnostic Usefulness. Results: Thirteen trainees (4 junior, 9 advanced) completed 2425 evaluations. Overall, the casebook was rated most useful (mean rank 1.7 ± 0.2), followed by LLM teaching (1.8 ± 0.3) and personal experience (2.4 ± 0.2; p = 0.023), with no significant difference between LLM and Textbook (p = 0.438). Junior trainees favored LLM teaching more than advanced trainees (first-rank 66.6% vs. 22.1%; p = 0.037). Across subspecialties, the casebook consistently ranked highest, with LLM slightly lower and experience lowest. LLM teaching received high ratings for Clarity (4.4 ± 0.3), Trust (4.3 ± 0.3), Differential Usefulness (4.3 ± 0.4), and Diagnostic Usefulness (4.2 ± 0.4), with no statistically significant difference between domains (p = 0.149). Conclusions: LLM-generated differential diagnoses are clear, trustworthy, and perceived as highly useful for education, nearing the perceived value of a standard casebook, especially for junior trainees. While textbooks remain essential, LLMs hold promise as supplementary tools, but caution is needed due to potential inaccuracies and their inability to replicate image-based teaching. Full article

Background/Objectives: To develop and validate a preoperative nomogram incorporating clinical and ultrasonographic features for the non-invasive differentiation of invasive ductal carcinoma (IDC) and invasive lobular carcinoma (ILC) of the breast. Methods: Preoperative clinical information and ultrasonographic features of patients with pathologically confirmed IDC and ILC were retrospectively collected. A total of 803 patients (600 with IDC and 203 with ILC) were enrolled and randomly allocated to training and validation sets in an 8:2 ratio. Univariate and multivariate logistic regression analyses were performed to identify independent predictors for differentiation. These predictors were subsequently incorporated into a nomogram and a corresponding weight plot. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC), while calibration was evaluated using calibration curves. Clinical net benefit was determined through decision curve analysis (DCA). Results: Significant differences were noted between IDC and ILC in multiple clinical and ultrasonographic characteristics (p < 0.05). Multivariate logistic regression analysis in the training set identified lesion margin, shape, depth, menopausal status, palpability, lesion classification, and internal echo as independent predictors of ILC. Notably, our constructed nomogram exhibited favorable predictive performance, calibration, and clinical utility in both the training and validation sets. Conclusions: A nomogram incorporating seven independent predictors, namely lesion margin, shape, depth, menopausal status, palpability, lesion classification, and internal echo, was developed and validated. This nomogram enables individualized and quantitative prediction of the preoperative probability of ILC and may serve as a non-invasive adjunct to support surgical decision-making. Full article

Artificial intelligence is reshaping employment in smart cities, yet centralized hiring platforms can deepen exclusion for persons with disabilities through privacy risk, biased models, weak multilingual support, and limited accommodation awareness. Because disability-related records are highly sensitive, no single institution holds enough representative data to train fair models, and centralizing such data is rarely permissible across borders. We propose FedAgent-Chain, a framework that integrates federated learning, blockchain-based auditability, multilingual processing, rule-based agentic services, and human-in-the-loop governance, extended with an education-to-employment module that builds individualized, accessible job-readiness pathways. Institutions across Saudi Arabia, the United States, China, and Europe train shared models without exchanging raw data. In a prototype evaluation on synthetic records over five seeds, the framework reached a mean F1 of 0.7207 (95% CI: [0.6506, 0.7909]), comparable to a centralized logistic-regression baseline while preserving data locality, with a formal $(\epsilon =3.2,\delta ={10}^{-5})$ differential-privacy guarantee after 20 rounds. Multi-dimensional fairness regularization lowered disability-category and work-mode disparity by 32.3% and 40.3% relative to local-only training. We report the fairness behavior transparently, including a case where the penalty does not outperform standard FedAvg on disability-category disparity, and we position cross-institutional integration with accountable governance, rather than raw metric superiority, as the central contribution. Full article

Underground mining robots are increasingly modeled for planning, operator training, and digital-twin workflows, where reliable actuator-level kinematics is needed to reduce hazardous in situ trials. Unlike typical open-chain industrial manipulators, representative mining machines are often linear-actuator-driven closed-chain mechanisms with planar four-bar linkages, making reusable kinematic modeling and real-time FK/IK solving challenging. We present MineRobot, an actuator-centered framework for modeling and solving the kinematics of this representative mechanism class. MineRobot introduces the Mining Robot Description Format (MRDF), a domain-specific representation that parameterizes mining-robot kinematics with native semantics for actuators and loop closures. It then contracts planar four-bar substructures into generalized joints and extracts, for each actuator, an Independent Topologically Equivalent Path (ITEP) classified into four canonical types. Based on this decomposition, per-type solvers are composed into a sequential forward-kinematics (FK) pipeline, while inverse kinematics (IK) is formulated as a bound-constrained actuator-length optimization solved by a Gauss–Seidel-style update scheme. By converting coupled closed-chain kinematics into small topology-aware solves, MineRobot reduces robot-specific hand derivations and supports efficient repeated FK/IK computation without treating each query as a full coupled constraint-solving problem. Experiments on representative underground mining robots demonstrate real-time FK performance and robust IK convergence within the tested operating ranges, supporting the use of MineRobot as an actuator-centered kinematic layer for planning, training, and digital-twin workflows. Full article

Modeling thermal environments across scales is crucial for climate-adaptive design and energy management. Traditional numerical methods (e.g., CFD) offer high accuracy and physical consistency, but they are computationally expensive. In contrast, purely data-driven models, though efficient, lack physical consistency and generalization capability. This review systematically examines Physics-Informed Neural Networks (PINNs), a hybrid paradigm in which physical prior knowledge is embedded directly into the neural network training process. A structured keyword search of the Web of Science Core Collection was performed, and 94 peer-reviewed journal articles were analyzed. The evolution from numerical simulations and data-driven surrogate models to PINNs is outlined. PINN methods are classified according to the stage at which physical prior information is integrated (i.e., dataset development, model construction, or loss function formulation). Current research remains heavily focused on loss function constraints, whereas systematic integration into data augmentation and model construction remains limited. Application domains span indoor environments, outdoor environments, and building systems, with each domain exhibiting unique prior integration strategies tailored to specific problems. Future PINN modeling should evolve toward multi-physics coupling, adaptive loss balancing, cross-scenario transfer learning, and unified evaluation benchmarks. PINNs in this field are promising but remain at an early stage, especially for complex urban-scale deployment. This review synthesizes existing research around the three stages of dataset development, model construction, and loss function formulation, summarizes the prior integration strategies adopted in the domain of building thermal environments, and provides a practical workflow for embedding physical prior knowledge at different stages of model development. Full article

Aspartame is a widely used artificial sweetener, but its possible relationship with rheumatoid arthritis (RA) remains insufficiently understood. This study aimed to explore, rather than prove, potential molecular links between aspartame-related targets and RA-associated gene networks. Three public RA transcriptomic datasets (GSE55235, GSE55457, and GSE77298) from the Gene Expression Omnibus (GEO) database were integrated as discovery/training data. Because these datasets included different tissue origins, batch correction was used to reduce dataset-level technical variation, whereas tissue-origin-related biological variation was not assumed to be fully removable. After differential expression analysis, RA-associated differentially expressed genes (DEGs) were identified. The single-cell dataset GSE200815 was used for cell annotation and cellular expression visualization; because its comparator group consists of psoriatic arthritis (PsA) samples rather than healthy controls, single-cell results were interpreted as RA-vs-PsA observations and were not treated as disease-versus-healthy-control evidence. Potential targets of aspartame were retrieved from ChEMBL, SwissTargetPrediction, and the Similarity Ensemble Approach (SEA), and were intersected with RA-related DEGs to construct an aspartame-gene-RA regulatory network. Diagnostic models were developed using 113 machine-learning algorithm combinations to determine an optimal multigene model and its core genes. HLA-DRB1 was selected for exploratory scTenifoldKnk-based virtual knockout mainly because it was included in the optimal model and has a well-established role in RA immunogenetics; the single-cell analysis was used only to describe cellular distribution in the RA/PsA dataset. Molecular docking was then used to evaluate the possible interaction between aspartame and HLA-DRB1. Forty-four intersected genes linked the predicted aspartame targets with RA DEGs. The random forest plus partial least-squares generalized linear model (RF + plsRglm) identified 16 core genes. Network-level interpretation indicated that these genes were distributed across immune/antigen-processing, inflammatory-signaling, protease/extracellular-matrix-remodeling, adhesion, metabolic, and proliferation-related modules; therefore, HLA-DRB1 was treated as a prioritized immune-module candidate rather than as the sole driver of the network. Following virtual knockout of HLA-DRB1, affected genes were enriched in extracellular matrix organization, extracellular structure organization, extracellular matrix, collagen trimer, extracellular matrix structural constituent, and collagen binding. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways included integrin signaling, focal adhesion, proteoglycans in cancer, cytoskeleton in muscle, and phosphoinositide 3-kinase/protein kinase B (PI3K/AKT) signaling. Molecular docking showed a minimum binding energy of −6.7 kcal/mol, which was more negative than the preset stability criterion of −5.0 kcal/mol, and the docking pose suggested contacts around ARG-146. This integrative analysis suggests a hypothesis-generating association between aspartame-related predicted targets and RA-relevant molecular networks involving HLA-DRB1 and other core genes. The findings do not establish causality and require experimental, epidemiological, biophysical, and tissue-stratified validation before any causal or clinical inference can be made. Full article

Background: Sexuality education is essential for adolescent health and well-being, yet in Italy it is not included in a mandatory national curriculum, resulting in heterogeneous implementation across regions. Within the ESPRIT project, a multidisciplinary training-of-trainers (ToT) model was developed to prepare professionals to support school-based peer-education pathways. This study aimed to describe the training model and perform a pilot evaluation of short-term knowledge outcomes among trained participants. Methods: A pilot non-randomized controlled comparative study was conducted within the ESPRIT project framework. A multidisciplinary Training Team developed a structured ToT pathway based on WHO guidance, national recommendations, and peer-education models. Ten advanced public health residents in Hygiene and Preventive Medicine attended a three-day residential training course. One month later, a 10-item knowledge questionnaire was administered to trained participants (n = 10) and untrained advanced public health residents (n = 10). Results: Trained participants achieved higher questionnaire scores than the comparator group (median score 8 [IQR 2] vs. 3.5 [IQR 2]; p < 0.0005). Conclusions: Structured ToT programmes may represent a promising approach for strengthening professional preparation in sexuality education. Larger studies with longer follow-up are needed to evaluate sustainability and real-world implementation. Full article

Flow-field parameters of bow shocks in high-altitude rarefied flow are fundamental for seeker radiation noise evaluation and thermal-protection design. The conventional direct simulation Monte Carlo (DSMC) method is computationally expensive, making it difficult to achieve real-time prediction and massive sample generation of flow-field parameters. This paper presented a surrogate model adopting a convolutional neural network (CNN) to rapidly predict bow-shock reactive flow-field parameters. A blunt body with a nose radius of 0.1–1.0 m was investigated. The Latin hypercube sampling methodwas used to construct a sample space spanning altitudes of 80–150 km and Mach numbers of 15–35. DSMC-calculated data was segmented into training and test sets at a ratio of 4:1 and verified by the bow-shock ultraviolet experiments. An encoder–decoder CNN with a parallel decoder strategy was established to develop a bow-shock reactive flow surrogate model (CNN-BS) and conduct error evaluation. The results show that the mean absolute percentage errors for temperature, velocity, pressure, and nitric oxide number density are below 8%, with coefficients of determination close to 1. The average prediction time is 0.5 s, enabling online data generation. The CNN-BS model provides efficient support for radiation-noise evaluation and thermal-protection design of hypersonic blunt bodies. Full article

Background/Objectives: Chronic kidney disease (CKD) remains underdiagnosed until advanced stages, motivating reliable, clinically deployable screening models that pair high discrimination with an explicit operating threshold and transparent explanations. Methods: In this study, we propose a CKD detection framework that integrates structured preprocessing, class imbalance handling, stratified 10-fold cross-validation with out-of-fold (OOF) prediction, and clinically oriented threshold selection via the Youden index, followed by explainability using SHAP and LIME. Experiments were conducted on two datasets. Across a broad panel of ten machine learning models, gradient boosting methods consistently dominated. Results: LightGBM achieved the best overall clinical composite performance on both datasets. On Dataset 1, LightGBM delivered near-ceiling OOF discrimination (ROC-AUC = 99.98, PR-AUC = 99.98) and an excellent clinically balanced performance at the best Youden threshold (0.41), reaching sensitivity = 99.20, specificity = 99.60, accuracy = 99.40, F1 = 99.40, and MCC = 98.80, with robust cross-validation stability (CV AUC = 99.99 ± 0.04; CV sensitivity = 99.10 ± 1.81; CV specificity = 99.46 ± 1.42; CV MCC = 98.59 ± 2.19), strong calibration (Brier = 0.006), and fast training (0.078 ± 0.019 s/fold). On Dataset 2, LightGBM maintained high generalization (ROC-AUC = 99.72, PR-AUC = 99.64) and clinically deployable balance at the best Youden threshold (0.35), achieving sensitivity = 98.10, specificity = 98.03, accuracy = 98.06, F1 = 98.06, and MCC = 96.13, with consistent fold-wise performance (CV AUC = 99.69 ± 0.25; CV sensitivity = 97.25 ± 1.25; CV specificity = 98.11 ± 1.02; CV MCC = 95.37 ± 1.56), acceptable calibration (Brier = 0.0173), and practical training time (0.742 ± 0.144 s/fold). Conclusions: Finally, SHAP and LIME explanations confirmed that model decisions align with clinically meaningful renal function and symptom/biomarker patterns at both population and patient levels, supporting safer translation of the proposed framework into CKD screening and decision-support workflows. Full article

Infra-hepatic inferior vena cava (IVC) balloon occlusion is an effective strategy for reducing intraoperative bleeding during precision liver surgery, yet rapid balloon inflation can produce abrupt transient deviations in downstream venous pressure that are not yet quantitatively characterized. Current practice relies on operator experience, with no quantitative framework to balance occlusion efficacy against downstream pressure safety. A computational fluid dynamics (CFD) model of the balloon-occluded IVC was developed in ANSYS 2025 R2 with two-way fluid–structure interaction (FSI), Carreau–Yasuda blood rheology, and a balloon described by an Ogden hyperelastic model; the flow regime was laminar (Re ≈ 254). Reduced-order ARX models of four input–output subsystems were identified from CFD-generated data, and a model predictive control (MPC) strategy was formulated to penalize downstream pressure overshoot through a weighted cost function. The identified models achieved training normalized root-mean-square errors of 0.0363 to 0.1164 and out-of-sample validation errors of 0.1224 to 0.2381. Conventional sigmoid inflation induced a 45.82% overshoot in downstream pressure (P_aft); the optimal input signal (q = [⁰,¹,⁰,⁰], λ = 0.1) reduced this to 6.05%, a reduction of 39.77 percentage points, while preserving >90% flow occlusion at U_F = 3 × 10⁴ Pa. The proposed framework offers a quantitative basis for balloon-occlusion device design that limits downstream pressure overshoot, motivating subsequent benchtop, ex vivo, and in vivo validation. Full article

To address the low efficiency and high subjectivity of manual inspection of Astragalus slices, as well as the limited fine-grained detection accuracy caused by the visual similarity between the characteristic radial “chrysanthemum heart” texture and minor defects such as insect damage and mold, this study proposes a lightweight intelligent detection model named Astra-YOLO. A dataset consisting of 622 original Astragalus slice images from four categories was divided into training, validation, and test sets at a ratio of 8:1:1. Data augmentation was applied exclusively to the training set, resulting in a total of 3110 images. Based on YOLOv11n, three targeted improvements were introduced: GhostConv lightweight convolution was employed to reduce model parameters and computational cost; the parameter-free SimAM attention mechanism was integrated to suppress interference from complex textures and enhance defect feature representation; and Wise-IoU v3 was adopted to improve bounding box regression for precise localization of small defects. The experimental results demonstrate that Astra-YOLO achieves superior performance with only 2.53 million parameters and 6.20 GFLOPs. The model attains an mAP@0.5 of 92.7%, an mAP@0.5:0.95 of 73.8%, a precision of 92.4%, and a recall of 92.1%. These results indicate that Astra-YOLO effectively balances lightweight design and detection accuracy, outperforming the baseline model and other improved variants, thereby providing reliable technical support for industrial online inspection and automated quality grading of Astragalus slices. Full article

Efficient deep learning systems are increasingly essential for automated quality inspection in the agri-food industry. This work systematically investigates the impact of optimized data-loading and augmentation pipelines on both training efficiency and predictive performance of convolutional neural networks for fruit defect classification. Benchmark experiments compare CPU-based preprocessing, multithreaded tf.data pipelines, GPU-accelerated workflows, and NVIDIA DALI, showing up to a 16× reduction in training time together with significantly improved GPU utilization. Building on these findings, the optimized pipeline is deployed on a large-scale industrial dataset comprising more than 3 million tangerine image patches. Carefully designed augmentation strategies—including geometric transformations and color perturbations—are introduced to enhance data diversity while preserving the intrinsic visual characteristics of the product. The substantial reduction in training time enables a more efficient exploration of candidate architectures through a tailored Neural Architecture Search (NAS) framework designed for resource-constrained industrial settings. The proposed framework explores internal CNN hyperparameters while preserving architectural depth to satisfy real-time inference constraints. To reduce the computational cost of NAS, a Random Forest–based performance predictor is trained on early-epoch indicators such as the F1-score and used to rapidly screen candidate models. A genetic algorithm is then employed to efficiently explore the search space and identify high-performing configurations. Experimental results demonstrate that the proposed end-to-end workflow significantly accelerates the model development cycle while maintaining or modestly improving classification accuracy. While the reductions in training time are substantial, the predictive-performance improvements observed through NAS are comparatively modest and should be interpreted primarily as evidence that the proposed framework can identify competitive configurations under industrial deployment constraints. The resulting framework provides a practical and scalable workflow for developing and deploying automated visual inspection systems in industrial agri-food production lines. Full article

Parkinson’s disease is a progressive neurodegenerative disorder characterized by motor impairments that significantly affect handwriting and fine motor control. Recent advances in artificial intelligence have enabled the non-invasive analysis of handwritten patterns as reliable digital biomarkers for early Parkinson’s disease detection. However, existing deep-learning approaches often struggle with diagnostic uncertainty and lack interpretability, limiting their clinical reliability and practical adoption. Moreover, models trained on single datasets frequently exhibit poor generalization across heterogeneous handwriting sources. This study uses two image-based handwriting datasets and one CSV-based HandPD feature dataset, including the Parkinson’s Augmented Handwriting Dataset, Parkinson’s Drawings Dataset, and HandPD Spiral/Meander feature records. A Transformer-based architecture is employed to learn global motor patterns from handwriting images, followed by a fuzzy-logic-based decision layer to handle uncertainty and improve robustness. The novelty of this work lies in integrating Transformer-driven deep feature learning with fuzzy clinical reasoning, supported by an AIC-based handcrafted feature analysis for interpretability. The model performance is evaluated using accuracy, precision, recall, F1-score, MCC, and AUC metrics. The experimental results demonstrate that the proposed Transformer–Fuzzy framework consistently outperforms CNN and Transformer-only baselines, achieving superior classification performance and robust generalization across all datasets, thereby establishing its effectiveness for reliable and interpretable Parkinson’s disease screening. Full article

Gain-scheduled and cascaded proportional–integral–derivative (PID) autopilots remain common practical baselines for fixed-wing unmanned aerial vehicles (UAVs), but training one shared learned controller for heading, altitude, and true airspeed across several maneuvers remains difficult. We study this problem under a strict reach-then-hold benchmark in which all the active channels must enter prescribed green bands and remain there for a terminal hold window. The proposed training recipe combines proximal policy optimization (PPO) with a tri-band maneuver-tracking reward and an outer bounded reward reweighting (BDR) step that updates the base reward weights from recent violation summaries under a Kullback–Leibler (KL) gate. In the JSBSim F-16 six-degree-of-freedom dynamics model, used here as a challenging surrogate benchmark for fixed-wing UAV autopilot learning, the learned controller transfers across a fixed five-lesson sequence, reaches strict success rates of 0.966 on turn and 0.921 on climb, and issues substantially smaller executed-command updates than the shared fixed-gain PID reference used here. Under the reported lesson sequence and step budget, fixed-weight PPO and a reweighting-only variant stall under the same envelopes, while speed remains the main bottleneck for both controllers. We further report exploratory long-horizon tracking, difficult-command stress checks, and an added command-filtered nonlinear dynamic-surface-control (CF-DSC) reference without retraining the learned policy. The CF-DSC results confirm that advanced non-reinforcement-learning (non-RL) controllers can be strong reference methods; therefore, within this reported simulator setup, BDR should be read as a practical and inspectable reward-scheduling heuristic for shared triad tracking rather than as a proof of superiority over all classical, nonlinear, or model-based controllers. Full article

Accurate recognition of partially occluded faces remains challenging in unconstrained and real-time environments, especially under masks, partial occlusions, pose variation, and illumination changes. This study presents a controlled comparison of three hybrid multi-face recognition pipelines for robust occluded face recognition. For fair evaluation, all pipelines use the same SCRFD face detector, preprocessing protocol, Linear SVM classifier, and 60% unknown rejection threshold, while varying only the feature extractor: ResNet29, ConvNeXt, and ResNet100 with ArcFace embeddings. To reduce data leakage, models are trained only on normal, non-occluded faces and tested on unseen partially occluded faces. Evaluation is performed on a custom dataset and the public Real-World Occluded Faces dataset, alongside three existing paper methods with publicly available code tested under the same experimental protocol. The SCRFD with ArcFace ResNet100 and Linear SVM pipeline achieved the best results compared to existing papers and our other pipelines, reaching 97.475% real-time accuracy for five faces and over 99% confusion-matrix-based accuracy on the custom dataset. On the ROF dataset, it also achieved closed-set accuracies of 98.66% for sunglasses and 97.92% for masks, with threshold-based accuracies of 96.35% for the sunglass test and 95.14% for the mask test. Furthermore, it obtained EER values below 0.007 and AUC values above 99%. In real-time testing, it achieved 29.25 FPS with 34.18 ms/frame latency on a GPU-enabled laptop and approximately 5 FPS with 273.4 ms/frame latency on a Raspberry Pi 4. Full article