Search Results (1,001)

Background/Objectives: Ultrasound is widely used for breast cancer screening and diagnosis, particularly in low- and middle-income settings, but its diagnostic reliability is often compromised by speckle noise that degrades lesion margins and tissue texture. This study proposes a compact convolutional neural network architecture that replaces standard max or average pooling layers with wavelet-based pooling using Symlet families, and optionally includes wavelet-domain preprocessing to suppress input noise. Methods: We conducted 108 experiments across six pooling configurations (avg, max, Sym2 ± preprocessing, Sym4 + preprocessing, Sym6 + preprocessing), two network depths, three batch sizes, and three simulated speckle levels (0%, 10%, 20%). Results: The proposed wavelet-based pooling framework showed consistently stronger in-domain performance than conventional pooling strategies across clean and speckle-corrupted settings, with the Sym2 + preprocessing configuration giving the best overall results. The model achieved 93.90% accuracy and 98.89% ROC AUC under clean internal test conditions and maintained stable performance under increased simulated noise levels. However, external validation on the independent BrEaST-Lesions-USG dataset revealed substantial performance degradation, with accuracy decreasing to 53.97% and ROC AUC to 0.4713, indicating limited cross-dataset generalization. Conclusions: These findings suggest that wavelet pooling is an effective architectural modification for improving in-domain robustness under controlled perturbation, although additional strategies are still required before reliable real-world deployment can be claimed. Full article

This paper investigates solar energy production forecasting at a monthly temporal resolution using a pooled neural network framework applied to the Chikalov photovoltaic systems in southwestern Bulgaria. The study considers several related PV installations with unequal time-series lengths and formulates the forecasting task as one-step-ahead prediction of the next monthly total energy yield, measured in kWh, in a global pooled setting. Two complementary neural architectures are compared: a multilayer perceptron (MLP), which serves as a nonlinear feed-forward benchmark based on lagged observations and seasonal descriptors, and a gated recurrent unit (GRU), which explicitly models sequential temporal dependence. In both cases, seasonality is represented through cyclical calendar encodings, while model selection is performed by chronological hyperparameter search using a separate validation block. Forecast accuracy is assessed by RMSE, MAE, coefficient of determination (R²), MAPE, and sMAPE, and uncertainty is quantified through validation residual prediction intervals. The results show that the MLP achieves stronger validation performance, whereas the GRU provides better final out-of-sample generalization after refitting on the combined training and validation data. For both architectures, the best configurations are obtained with a 12-month input horizon, indicating that one full annual cycle contains the most informative memory for forecasting monthly aggregated photovoltaic energy yield in the considered dataset. After refitting on the combined training and validation data, the GRU achieved the best final out-of-sample performance, with RMSE = 296.38 kWh, MAE = 213.16 kWh, R² = 0.9231, MAPE = 7.52%, and sMAPE = 7.49%. Overall, the findings demonstrate that pooled neural modeling is an effective framework for monthly PV production forecasting and can provide practically useful support for sustainable energy planning, monitoring, and optimization. Full article

With the continuous growth of global population and increasing pressure on food security, the transformation toward precise and intelligent agricultural production has become an inevitable trend. In this context, accurate identification of field weeds is crucial for improving crop yields and reducing agricultural inputs. However, agricultural Internet of Things (IoT) edge devices are generally subject to strict constraints in terms of power consumption, storage, and real-time performance. Existing lightweight convolutional neural networks often struggle to simultaneously achieve high accuracy and low resource consumption for fine-grained weed identification tasks. To address this challenge, this paper proposes a hardware aware lightweight convolutional neural network named Weed-Aware Ghost Enhanced Network (WAGENet) for microcontroller deployment. The network synergistically integrates Ghost low-cost feature generation, Mobile Inverted Bottleneck Convolution (MBConv) for deep semantic extraction, Squeeze and Excitation (SE) and Coordinate Attention (CA) dual attention mechanisms for channel space joint calibration, and Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context fusion. It constructs a progressive feature abstraction system from shallow textures to high-level semantics. On the public DeepWeeds dataset, WAGENet achieves 95.71% classification accuracy and 93.80% F1 score with only 0.163 M parameters and 2.43 × 10⁸ multiply accumulate operations (MACC), attaining a parameter efficiency of 587.19%/M and significantly outperforming existing mainstream lightweight models. The model has been successfully deployed on the STM32H7B3I microcontroller development board, achieving a single inference latency of 94.63 ms, an internal Flash footprint of only 686.95 KiB, and a single inference energy consumption of 41.45 mJ. Experimental results demonstrate that WAGENet achieves a trade off among accuracy, latency, and energy consumption under strict resource constraints, providing a reproducible microcontroller deployment paradigm for battery powered field robots, drones, and other agricultural IoT edge devices. Full article

The precise and timely diagnosis of grapevine diseases is paramount for ensuring food security and mitigating economic losses within the viticulture sector. While existing deep learning models offer high accuracy, their computational intensity and hardware requirements often hinder their use in portable or low-power field systems. This study addresses this gap by proposing GrapeLeafNet, a lightweight convolutional neural network optimized for efficient feature extraction. GrapeLeafNet introduces a strategic hybrid approach that combines the low parameter efficiency of models like SqueezeNet with the rapid feature propagation advantages offered by shallow architectures such as AlexNet. By eliminating the sequential processing latency caused by SqueezeNet’s 18-layer deep structure and the excessive 61-million-parameter memory burden of AlexNet, this model establishes a critical balance between low latency and high accuracy through its optimized 7-layer architecture. Characterized by an original integration of standard convolutional layers, batch normalization, and max pooling, GrapeLeafNet achieves high computational efficiency with only 1.6 million parameters and a 6.26 MB memory footprint. This structural optimization enhances deep feature hierarchies, enabling the model to focus on distinctive pathological signs within complex leaf patterns and maximize classification sensitivity by filtering out irrelevant features. The evaluation was conducted using the Niphad Grape Leaf Disease (NGLD) dataset, incorporating data augmentation to mitigate inherent class imbalances. Additionally, data augmentation techniques were employed to mitigate inherent class imbalances within the dataset. Experimental results demonstrate that GrapeLeafNet achieved 97.06% accuracy and a 94.77% F1-score on the original dataset, outperforming recent benchmarks by 2.46%. Following augmentation, performance reached 98.29% accuracy and a 98.16% F1-score, representing a 6.16% higher F1-score than contemporary models. GrapeLeafNet exhibits high robustness against asymmetric class distributions and establishes a significant performance margin over existing architectures. Its lightweight nature, combined with superior accuracy and F1-score metrics, makes it an ideal candidate for integration into mobile devices and real-time agricultural monitoring systems. Full article

Hand gesture recognition (HGR) is a critical area in computer vision that supports intuitive human–computer interaction and sign language communication, yet existing systems remain sensitive to lighting variations, background clutter, and diverse hand postures. This study introduces two contributions to address these limitations: a Gradient-Based Augmentation Validation (GBAV) framework that establishes structurally safe augmentation ranges before training, and a multi-backbone Convolutional Neural Network (CNN) architecture combining ResNet50 and InceptionV3 with optional attention-based pooling. GBAV uses magnitude-weighted gradient orientation histograms with Pearson correlation and Kullback–Leibler divergence thresholds to verify label invariance under spatial transformations, providing a classifier-agnostic pre-training calibration mechanism. The proposed framework is evaluated on three static gesture datasets, Indonesian Sign Language (BISINDO), American Sign Language (ASL), and Hand Gesture 14 (HG14), yielding validation accuracies of 96.87%, 99.92%, and 95.25%, respectively, with 5-fold cross-validation on HG14 confirming result stability (93.51% ± 2.31%). Quantitative attention localization, cross-dataset transfer evaluation, and computational efficiency analysis (26.8 ms per image, ~37 FPS) further support the framework’s robustness and practical deployability. These findings establish GBAV-calibrated augmentation as the principal performance driver, which complements the multi-backbone architecture for robust hand gesture recognition across diverse visual contexts. Full article

Automated defect detection in industrial casting processes is important for improving product quality while reducing the cost of manual inspection. In this work, two deep convolutional neural network (CNN) architectures, InceptionV3 and InceptionResNetV2, are evaluated for the binary classification of defects in submersible pump impellers. A genetic algorithm (GA) is used to optimize key hyperparameters, including dropout rate, learning rate, and dense layer configuration, while model complexity is assessed through Pareto-based analysis. Single-run optimization results show that InceptionV3 achieves high classification accuracy (99.0%) with lower model complexity than InceptionResNetV2 (98.75%). Repeated experiments using different random seeds demonstrate relatively stable performance across runs, with InceptionV3 achieving an accuracy of 0.9913 ± 0.003 and InceptionResNetV2 achieving 0.9860 ± 0.0076. Additional experiments were conducted using random-search baselines and classification-head ablation studies (Flatten vs. Global Average Pooling). These experiments showed that optimization strategy and architectural design choices influence both predictive performance and computational complexity. The environmental impact of the training process is evaluated using CodeCarbon, with energy consumption ranging from 0.083 to 0.098 kWh and carbon emissions ranging from 2.008 to 2.401 g CO₂eq for InceptionV3 and InceptionResNetV2, respectively. Overall, the results suggest that the most effective configuration depends on the evaluated architecture and experimental setting, highlighting the importance of balancing accuracy, model complexity, and computational efficiency in industrial defect detection systems. Full article

Background: An accurate D’Amico risk stratification is mandatory for prostate cancer (PCa) management. The purpose of this proof-of-concept study was to establish a methodological framework for integrating validated clinical nomograms with strict data-quality governance in order to generate reliable artificial neural networks (ANNs), even when the sample is small. Methods: We performed a retrospective analysis of a curated cohort of 49 patients from one centre. A multilayer perceptron (MLP) was trained using 11 variables, including the ISUP biopsy grade and Briganti nomogram. Model development was guided by a proactive data-quality protocol based on FAIR principles—the DQG-AI framework (data quality governance for AI-readiness, developed at Clínica Universidad de Navarra)—with stringent checks for accuracy, consistency and validity to ensure data were “AI-ready”. A sensitivity analysis was conducted on three data partitioning scenarios (20/80, 34/66 and 39/61). Results: From a starting pool of 76 patients, the DQG-AI framework was applied to create a highly selected cohort of 49 patients. A multilayer perceptron (MLP) trained on this “AI-ready” dataset achieved, on the 20/80 configuration, mathematically perfect discrimination (AUC 1.000; 100% accuracy) for High vs. Intermediate risk groups on a very small refined internal test set (N = 9), a figure we interpret as a methodological artefact of the curated dataset and validation constraints rather than as an indicator of true model performance. This complete accuracy is not, however, presented as evidence of generalizable clinical utility: it is a best-case figure obtained on a single, very small test subset (N = 9) after necessary validation-related exclusions, and the wide confidence interval (66.4–100%), together with the software-driven removal of test cases carrying factor levels absent from the training set (detailed in the Methods section), explicitly preclude any inference about real-world performance. Accordingly, the deliverable of this proof-of-concept study is the DQG-AI framework itself, not the model’s reported accuracy. Conclusions: The main contribution of this proof-of-concept study is the effective illustration of the DQG-AI framework as a strict, repeatable approach for producing “AI-ready” urological datasets. Although the MLP demonstrated a robust internal signal for risk discrimination, its flawless accuracy is an ideal, non-generalizable situation. The most important deliverable that needs external validation is the DQG-AI framework, not the model’s performance metrics. A pre-specified three-phase multi-institutional validation roadmap (single-centre cohort expansion → within-system between-site validation → Spanish multi-centre external validation), with a minimum target of ~220 evaluable patients derived from a 10-events-per-predictor floor, is provided to operationalise this external validation. Full article

Early mild cognitive impairment (EMCI) is the earliest intervenable stage of Alzheimer’s disease (AD). Although graph neural networks (GNNs) have begun to exploit brain network topology, traditional fMRI-based diagnostic methods often neglect these structural patterns by relying on vectorized features. Furthermore, existing GNNs frequently disregard inter-regional functional heterogeneity and group-level discriminative patterns, leading to limited accuracy and biomarker interpretability. To address these challenges, we propose HF-BrainGNN, an end-to-end hierarchical graph learning framework for EMCI identification. Our method introduces a functional affinity region convolution (FAR-Conv) layer to learn region-adaptive kernels, a Differential Focus Pooling (DF-Pool) module to identify disease-salient brain regions by maximizing inter-group distinctiveness, and a hierarchical integration classifier (HIC) to fuse multi-level graph representations. The framework is optimized using classification, focus separation, and consistency regularization losses. Experiments on the ADNI dataset (104 EMCI, 114 Cognitively Normal) show that HF-BrainGNN achieves 86.78% accuracy, outperforming the best baseline (Hi-GCN) by 4.64%. Furthermore, the automatically identified regions, such as the bilateral hippocampus and default mode network hubs, align with established EMCI biomarkers. Ultimately, HF-BrainGNN provides an efficient, interpretable artificial intelligence tool for precise brain network characterization and early AD intervention. Full article

This study proposes a stress field data-driven prediction method that combines a finite element thermo-mechanical coupling model with a multi-machine learning framework. This method takes the inversion of stress based on the shrinkage behavior of deposition layers as the core logic, extracts the node displacement shrinkage during the cooling to solidification process of the melt pool in the thermal coupling simulation as the key feature input, and constructs extreme gradient boosting (XGBoost), Gaussian process regression (GPR), and deep convolutional neural network (DCNN) models, respectively, to achieve accurate prediction of nodal effect stress and triaxial stress in the laser directed energy deposition (L-DED) node process. The experimental results show that the XGBoost algorithm performs the best in various stress prediction indicators, and its generated stress distribution cloud map is highly consistent with the thermal coupling simulation results, suggesting a strong correlation between deposition layer shrinkage behavior and the stress field under the investigated conditions. In addition, compared to traditional finite element simulations, this method significantly improves computational efficiency while ensuring prediction accuracy, providing a new approach for rapid assessment of residual stresses. Full article

A novel method based on noise-assisted multivariate empirical mode decomposition (NA-MEMD) combined with a modified multiple branch convolutional neural network (MMBCNN) is designed to detect fault events in distribution networks and to classify various faults in a distribution system. Given the presence of noise components in transient voltage signals, a moving time window technique integrated with the NA-MEMD method is employed to process high-frequency sampling and long-term series signals. This method is also utilized to reliably identify noise components in modal components through permutation entropy. On this basis, the Clarke transform is employed to convert transient voltage signals into the d–q axis, and three-phase voltage waveforms are transformed into a ring image. Moreover, an MMBCNN is developed to accurately detect and classify distribution network faults, and a modified pooling function is introduced to improve feature extraction ability and model convergence performance. Finally, the accuracy and effectiveness of the proposed algorithm are estimated and analyzed using measurement and fault simulation data from distribution networks. Full article

COPD remains a prevalent and debilitating respiratory condition, necessitating early and accurate diagnosis for optimal clinical intervention. In this study, we propose a novel deep learning-based diagnostic framework that employs the ECAPA-TDNN (Emphasized Channel Attention, Propagation and Aggregation—Time Delay Neural Network) architecture to classify respiratory sound signals from the ICBHI dataset. Originally designed for speaker verification, ECAPA-TDNN introduces channel attention and multi-scale feature aggregation, which we adapt for the first time to the domain of medical acoustic analysis. This architecture allows the model to effectively capture subtle and discriminative patterns in pathological breathing sounds, overcoming the limitations of conventional CNN-based methods. Our methodology integrates rigorous signal preprocessing, log-Mel spectrogram extraction, and data augmentation to enhance robustness and generalization. An Attentive Statistics Pooling mechanism is employed for temporal feature summarization, while Grad-CAM-based explainability is incorporated to improve the interpretability of the diagnostic predictions. The model is rigorously validated using a five-fold cross-validation scheme, achieving a mean validation accuracy of 96.8% with consistently high F1-scores and recall rates across all folds. Comparative analysis with prior methods highlights the superiority of our ECAPA-TDNN-based approach in terms of diagnostic precision, robustness, and potential clinical applicability. To the best of our knowledge, this is the first work to adapt ECAPA-TDNN for COPD detection from respiratory sounds, establishing a new benchmark in interpretable and high-performance acoustic-based respiratory disease screening. Full article

Aiming at the issues of low recognition accuracy and high model computational complexity for power quality disturbances (PQDs) in strong-noise environments, this paper proposes a novel lightweight PQD-recognition method that integrates a hybrid architecture of variational mode decomposition (VMD), convolutional neural network (CNN), and transformer. Firstly, a hybrid optimization algorithm named the monkey–genetic hybrid optimization algorithm (MGHOA) is proposed to optimize VMD parameters for denoising disturbance signals, thereby enhancing recognition accuracy in noisy environments. Secondly, to fully extract disturbance signal features and reduce the computational complexity of the model, a lightweight CNN–transformer model is designed. Depthwise separable convolution (DSC) is employed to extract local features and the multi-head attention mechanism of transformer is utilized to mine the long-distance dependence and global features, thereby enhancing the feature representation. Thirdly, a multitask joint-learning method is proposed to collaboratively optimize classification accuracy and temporal localization tasks, enhancing the discrimination of similar disturbances. Additionally, a dual-pooling global feature fusion strategy is designed to further enhance the model’s ability to discriminate complex disturbances. Comparative experiments on 16 typical PQD types demonstrate that the proposed method achieves excellent performance in recognition accuracy, model robustness, and computational efficiency. The integration of the MGHOA–VMD module improves recognition accuracy by 1.08%, while the multitask joint-learning method contributes an additional 0.55% improvement. When achieving recognition accuracy comparable to complex models, the training time of the proposed method is 36.51% of that required by DeepCNN and merely 5.90% of that required by bidirectional long short-term memory (BiLSTM), with a 31.22% reduction in parameter scale. This work provides a novel solution for intelligent power quality disturbance recognition. Full article

This study aimed to develop and validate a physics-informed neural network (PINN) framework for data-efficient and physically consistent process optimization in the laser powder bed fusion (LPBF) of Inconel 718 (IN718) superalloy. Laser powder bed fusion (LPBF) is widely adopted for fabricating Inconel 718 (IN718) components in aerospace and energy applications; however, navigating its high-dimensional, nonlinear process parameter space remains a central challenge. High-fidelity finite element simulations are computationally prohibitive for extensive parameter sweeps, whereas purely data-driven machine learning (ML) models are limited by data scarcity and unphysical extrapolation behavior. This study presents a physics-informed neural network (PINN) framework that embeds the transient heat conduction equation and Goldak double-ellipsoidal heat source model directly into the neural network training loss, enforcing thermophysical consistency simultaneously with data fidelity. The model was trained on a curated, multi-source dataset of LPBF IN718 parameter combinations drawn from peer-reviewed experimental studies and validated finite element simulation outputs, spanning the laser power (70–400 W), scan speed (200–2000 mm/s), hatch spacing (50–140 µm), and layer thickness (20–50 µm). The PINN predicted the melt pool width, depth, peak temperature, and relative density with mean absolute percentage errors (MAPE) of 3.8%, 4.7%, 3.1%, and 1.9%, respectively, outperforming a baseline artificial neural network (ANN) with an identical architecture. The framework correctly identified the optimal volumetric energy density (VED) window of 55–105 J/mm³, yielding relative densities ≥99.5%, consistent with the published experimental thresholds for IN718. A data efficiency analysis demonstrated that the PINN with 25% training data achieves a performance equivalent to that of the fully trained ANN with 100% data, confirming an approximately four-fold data efficiency improvement attributable to physics-informed regularization, consistent with theoretical predictions. Sensitivity analysis via automatic differentiation confirmed that laser power and scan speed were the dominant parameters (~85% combined variance), which is in agreement with previous studies. This study provides a computationally efficient, interpretable, and physically consistent ML pathway for the accelerated process qualification of IN718 components for aerospace and energy applications. Full article

Graph neural networks (GNNs) have achieved great success in graph classification, with graph pooling methods being widely adopted for related tasks. Existing approaches typically rely on node ranking or clustering to coarsen graphs, but often fail to effectively leverage global structural information, leading to loss of critical substructures and limited interpretability—key limitations in molecular analysis and social network mining. To address these issues, we propose SparsePool, a graph pooling method that integrates node features and structural patterns through atomic decomposition. By dynamically decomposing graphs into interpretable atomic units via Boolean matrix factorization, SparsePool preserves semantically meaningful substructures while providing transparent evidence of retained patterns. We further introduce an Atomic Pooling Neural Network (APNN) for graph representation learning. Extensive experiments on relevant benchmarks including biochemical and social network datasets demonstrate that SparsePool outperforms state-of-the-art pooling methods, achieving an average classification accuracy improvement of 1.03% over baseline models while reducing structural information loss. We also discuss its compatibility with emerging quantum computing paradigms, such as quantum-accelerated sparse decomposition, as a promising direction for scaling graph processing in industrial contexts. Full article

Recent automated search methods have improved lower bounds for several Ramsey numbers, but the strongest gains often depend on structured seeding and cell-specific heuristic discovery. This leaves open a more fundamental question: Can a useful search structure be transferred across related Ramsey cells rather than rediscovered independently for each target instance? This work proposes a teacher–student framework for transferable structural search in Ramsey graph construction, inspired by the structure-distillation logic of Physics Structure-Informed Neural Networks (Ψ-NNs). The framework builds compressed structural representations from teacher witnesses and search traces, extracts reusable motifs and relations, and reconstructs transfer candidates. These are refined by balanced search and, for weak R(3, s) cells, by exact small-cell supervision. The framework is evaluated as a proof of concept across five Ramsey cells under transfer, matched-compute, search, ablation, and interpretability settings, including a proportional shift-scaling baseline and a greedy triangle-closing baseline that probe the structure-validity frontier from complementary directions. Supplementary experiments cover seed robustness, budget sensitivity, transfer-neighborhood variation, structural-resolution changes, stronger exact supervision, cross-r teacher pooling, single-teacher configurations, and scaling behavior across graph sizes. The results show that the portfolio version of the framework is the strongest balanced transfer method in the current study, while a structure-dominant oracle achieves stronger witness-shape agreement but worse Ramsey-valid construction. These findings reveal a clear structure-validity frontier and suggest that transferable Ramsey search should be evaluated by how well structural priors survive the validity constraints of new cells. Full article