Manual monitoring of lesions in Multiple Sclerosis (MS), which affects the central nervous system, imposes a significant workload on radiologists. In this study, various learning rate (η) scheduling strategies were comparatively analyzed to enhance the performance of a deep learning-based 3D U-Net model in MS lesion segmentation. During the training process, six different LR strategies—Time-Based, Step, Exponential, Adaptive Plateau, Cosine Annealing, and Cyclical—were tested on the MSLesSeg dataset. Model performance was evaluated using Dice Score (DSC), Hausdorff Distance (HD95), TPR, and PPV metrics. Experimental results demonstrated that the Adaptive Plateau (Reduce LR on Plateau) method achieved the highest segmentation success (0.622 DSC) and the lowest geometric error margin (21.695 HD95). The findings confirm the critical importance of hyper-parameter optimization on the generalization capacity and clinical accuracy of models in 3D medical image analysis.

Cloud manufacturing can meet the personalized needs of users by coordinating heterogeneous manufacturing resources and providing widely available networked intelligent manufacturing services. In a complex cloud man ufacturing environment, the trust mechanism provides users with prior knowledge when faced with numerous cloud manufacturing services that have similar functions but different non-functional attributes. Establishing a reliable trust mechanism can prevent service providers from making false claims to attract more users. Unlike earlier SLA-or neighborhood-based schemes [1,2], our framework explicitly fuses capability evidence and ex perience evidence, filters malicious feedback using multi-round voting with fuzzy reputation, and learns on the resulting user-service trust subgraph. Recent surveys show that cloud manufacturing (CMfg) platforms increas ingly integrate IIoT/CPS, data-driven analytics, and AI, while still facing challenges in interoperability, resource virtualization, reliability and trustworthy service selection [3,4]. In parallel, trust modeling has evolved from static reputation or SLA-centric formulations to graph-based and temporal learning frameworks that explicitly address trust dynamicity and malicious behaviors (e.g., TrustGuard and MATA) [5,6]. Furthermore, multi-aspect graph attention and trust-aware collaborative frameworks have been shown to be effective for trust/QoS assess ment in large-scale service graphs [7,8]. Motivated by these advances, this paper revisits hybrid trust management for CMfg and strengthens the discussion and evaluation against recent literature. This paper proposes a multi-source hybrid trust management framework for cloud manufacturing environ ments, aiming to help users identify trustworthy services. First, an adaptive fuzzy hierarchical weighted capability trust estimation method is designed. This method assigns objective weights to QoS attributes based on the degree of service SLA fulfillment, without requiring users to participate in attribute importance comparisons. Second, an experiential trust evaluation method is proposed based on a trust subgraph. It constructs a trust subgraph by filtering out malicious evaluation information and learns the hidden features of users and services through Graph Convolutional Networks (GCN) for experiential trust estimation. This mechanism combines capability trust and experiential trust to form a hybrid trust score, quantifying the trust relationship between users and services. The experimental results demonstrate that the proposed model is effective and reliable in evaluating service trust.

Lower limb motion recognition plays a vital role in intelligent prosthetic control and wearable assistance systems. While traditional methods typically utilize multichannel surface electromyography (sEMG) signals to ensure recognition performance, such systems often involve complex hardware setups, are uncomfortable to wear, and require high computational resources. To address these challenges, this study aims to develop an efficient, low-complexity motion recognition approach based on single-channel sEMG signals. The proposed method integrates fast iterative filtering decomposition (FIFD) with a hybrid deep learning framework. FIFD is employed to decompose raw single-channel sEMG signals into multiple subcomponents, allowing the extraction of informative intrinsic features. These features are then processed by a deep neural network that combines convolutional neural networks (CNNs), attention mechanisms, and long short-term memory (LSTM) units to jointly capture spatial and temporal characteristics of the sEMG signal. Experiments were conducted on a self-constructed dataset comprising sEMG recordings of lower limb movements. The proposed method achieved a recognition accuracy of 99.35%, outperforming conventional multichannel approaches and other baseline models. The model demonstrated strong robustness and generalizability using only single-channel input. This study presents a novel single-channel sEMG motion recognition method that significantly reduces system complexity while maintaining high recognition accuracy. The approach offers a promising solution for developing low-cost, efficient, and user-friendly wearable systems, particularly suitable for lower limb rehabilitation and humancomputer interaction in resource-constrained environments.

Epilepsy is a prevalent neurological disorder that demands accurate and efficient detection for effective clinical management. Electroencephalogram (EEG) signals provide a noninvasive, low-cost, and high-temporal resolution means of monitoring brain activity, which makes it particularly suitable for epileptic seizure detection. However, conventional detection methods that rely on manual feature extraction or expert interpretation are often labor-intensive, subjective, and lack generalization capability. To address these limitations, this study proposes a novel automatic epileptic seizure detection framework that integrates deep feature extraction with machine learning. First, short-time Fourier transform (STFT) is applied to perform time-frequency analysis and capture local dynamic features of EEG signals. Then, pretrained deep learning models are employed to extract deep features, which enhance the representation of the EEG signals. Finally, these deep features are fed into machine learning for epileptic seizure detection. Experiments are conducted on two publicly available EEG datasets, the Bonn EEG dataset and the New Delhi EEG dataset. The results indicate that fusing deep features from multiple pretrained deep learning models enhances the overall detection accuracy. This study presents an effective and novel method for automatic epileptic seizure detection using EEG signals. The combination of deep feature extraction and machine learning provides new insights into the detection of neurological disorders through deep learning techniques.

ABSTRACT Driver fatigue has been identified as one of the primary causes of traffic accidents. As long‐duration and high‐load driving becomes increasingly common, the risks of delayed reactions and impaired distance judgment continue to rise. Traditional behavior‐based methods for detecting driver fatigue often exhibit limited stability in complex driving environments. In contrast, electroencephalography (EEG) offers a more reliable detecting method by directly capturing central nervous system activity. This work focuses on fatigue driving detection based on deep learning and EEG, which outlines commonly used public datasets, key preprocessing techniques, feature extraction techniques, performance evaluation metrics, and mainstream deep learning architectures. Based on research progress over the past three years, the use of datasets, published journals, research challenges, and limitations of current methods were analyzed. Future research should improve the model's generalization ability and robustness, introduce richer brain network features, and construct a larger‐scale, high‐quality dataset that closely resembles the real driving environment.

Traditional deep learning methods for traffic spatial-temporal prediction often require manual architecture adjustments by experts, which is both time-consuming and inefficient. Neural Architecture Search (NAS) technology offers a promising solution by automating the discovery of optimal network architectures for specific tasks. However, existing NAS approaches face significant limitations, particularly in handling multi-period features and achieving a balanced extraction of spatial and temporal features. To address the above problems, we propose a novel Traffic Flow Prediction Model based on Multi-Period Spatial-Temporal Stepwise Search (MPSTSS). First, the neural architecture search space is expanded by incorporating more classic traffic spatial-temporal feature extraction networks. Second, specialized spatial-temporal search networks are designed to handle traffic data with various periodic patterns, including weekly, daily, and recent patterns. Third, a spatial-temporal stepwise search strategy is employed to effectively balance the extraction of spatial and temporal features, and a differentiable search method is utilized to enhance search efficiency. Comprehensive experiments conducted on three public traffic datasets demonstrate that the MPSTSS model can effectively identify optimal network architectures, achieving superior prediction performance and demonstrating robust generalization capabilities.

Driver fatigue represents a nonlinear and nonstationary cognitive process that introduces significant instability into human-machine systems. Conventional binary detection approaches are unable to characterize transitional states such as drowsiness, which limits their ability to capture the full spectrum of driver alertness. To overcome this limitation, this study proposes an entropy-informed deep residual network for fine-grained driver state recognition based on electroencephalogram (EEG) dynamics. In this framework, EEG signals are first preprocessed through filtering and artifact removal to ensure data reliability. Differential entropy features are then extracted using a Gaussian-based method so that the temporal complexity of the EEG can be represented more effectively. A residual neural network with an embedded attention mechanism (ResNet-AM) is further designed, which enables the model to emphasize salient neural features that reflect nonlinear transitions among alert, drowsy, and fatigued states. Experimental validation on the SEED-VIG dataset, which includes EEG recordings from 23 participants, demonstrates that the proposed model achieves an average accuracy of 87.30%, precision of 87.78%, sensitivity of 87.95%, and F1-score of 87.87%. The model attains a peak accuracy of 88.89% and achieves a drowsy-state recognition rate of 91.19%, which indicates its effectiveness in capturing transitional fatigue states. These results indicate that entropy-informed and attention-augmented deep learning can provide an effective means to model nonlinear EEG dynamics and support real-time cognitive monitoring in intelligent transportation systems.

The primary aim of this review paper is to examine current methodologies for diagnosing battery faults in electric vehicles (EVs) to enhance safety and reliability. With the growing adoption of electric vehicles (EVs), ensuring battery performance and preventing failures are critical challenges. Battery malfunctions can lead to significant safety risks, including thermal runaway, reduced efficiency, and vehicle breakdowns. Therefore, this paper evaluates various diagnostic techniques and strategies that can effectively detect, classify, and mitigate battery faults in electric vehicles (EVs), ultimately enhancing their operational safety and longevity. This review synthesizes recent advancements in battery fault diagnosis, covering both model-based and data-driven approaches. The paper highlights that hybrid diagnostic techniques achieve superior fault-detection accuracy compared to standalone model-based or data-driven methods. Key findings include:• ML and DL models demonstrate significant improvements in identifying subtle fault patterns that traditional methods may overlook.• Real-time monitoring using AI-enhanced BMS allows early detection of battery issues, reducing the risk of catastrophic failures.• While model-based techniques offer high interpretability, they require extensive domain expertise and computational power.• The integration of AI and IoT (Internet of Things) in battery diagnostics is a promising trend, enabling predictive maintenance and enhancing the safety of electric vehicles (EVs).Effective battery fault diagnosis is crucial for ensuring the safety and reliability of electric vehicles (EVs). The paper suggests that the future of battery fault detection will rely on advanced hybrid AI-driven techniques combined with real-time sensor-based monitoring. Further research is recommended to optimize computational efficiency and improve diagnostic robustness under varying operational conditions.

Unfavorable driving states (UDS) not only jeopardize the safety of the driver but also pose a serious threat to other road users. Therefore, our goal is to develop a novel EEG-based framework to accurately identify UDS in unseen subjects. By utilizing the improved synchronization likelihood (ISL) method, which is based on the L1 norm of the Minkowski distance, we construct brain functional network connectivity matrices and brain network topologies across different frequency bands to capture subtle coupling variations between EEG signals from different channels. We customize various deep learning models and compare their ability to extract discriminative high-level features from the brain functional connectivity matrices. All proposed models are evaluated using the leave-one-subject-out cross-validation strategy. Experimental results demonstrate that the combination of the ISL-based brain functional connectivity matrices, constructed by concatenating the beta and gamma frequency bands, with customized depthwise separable convolutional neural network (CDSCNN), significantly outperforms the single frequency band feature input into CDSCNN for UDS detection in unseen drivers, as confirmed by a one-way analysis of variance followed by post-hoc multiple comparisons with Bonferroni test. The highest average performance is achieved with an accuracy of 95.69%, precision of 95.84%, sensitivity of 95.63%, specificity of 95.75%, and F1 score of 95.70%, outperforming traditional machine learning methods and demonstrating the potential of the proposed CDSCNN framework for EEG-based UDS identification.

Background: Autism Spectrum Disorder (ASD) affects approximately 1% of the global child population, yet current gold-standard diagnostic methods remain time-intensive and expertisedependent. Electroencephalography (EEG) offers an objective and scalable approach for neurophysiological measurement, facilitating early detection. Methods: This study evaluated three neural sequence architectures -Long Short-Term Memory (LSTM), Transformer, and Mamba (Selective State Space Model) -for ASD classification using 47-channel, 150-second resting-state EEG recordings from 56 adults (28 with ASD, 28 controls) from the University of Sheffield dataset. Data were preprocessed using MNE-Python with bandpass filtering (0.50-50 Hz), Independent Component Analysis (ICA) artifact removal, and zscore normalization. Models were trained on epochs of varying durations (1 s, 2.50 s, 5 s) using stratified 5-fold cross-validation, with performance evaluated on a held-out test set (15%). Mixture-of-Experts (MoE) ensembles were constructed using performance-based weighted averaging. Regional classification and spectral analyses identified anatomical and frequency-specific biomarkers. Results: The Mamba model achieved 98.18% accuracy with only 2972 parameters and a training time of 0.09 min at 2.50-second epochs. LSTM (144,578 parameters) reached 95.25% accuracy, while Transformer (38,946 parameters) attained 94.41%. The optimal Mamba+LSTM ensemble achieved 98.46% accuracy (Cohen's kappa=0.97, ROC-AUC=99.84%) with only 11 misclassifications from 716 test samples. Regional analysis revealed frontal lobe dominance (76.81% accuracy, 25 channels) with theta-band (4-8 Hz) biomarkers. Spectral analysis confirmed characteristic ASD patterns: elevated delta/theta power, suppressed alpha rhythm, and increased beta/gamma activity. Single-channel analysis identified C5 (left central, 58.80% accuracy) as the most discriminative electrode. Conclusions: Neural sequence models, particularly the parameter-efficient Mamba architecture and the Mamba+LSTM ensemble, demonstrate exceptional performance for EEG-based ASD classification, offering a clinically scalable and objective diagnostic tool. The frontal-central electrode configuration and theta-band biomarkers provide neurophysiologically interpretable features suitable for portable EEG systems and early screening applications.

Predicting prostate cancer (PCa) risk from magnetic resonance imaging (MRI) requires identi fying representative causal features that are critical for accurate diagnosis. However, existing methods often overlook the rich causal dependencies among radiomics or deep learning-based features, limiting their robustness under data heterogeneity. To address this issue, we propose an end-to-end multi-task deep causal network that integrates causal structure learning to model local causal relationships between image-derived features and multiple PCa risk categories. The proposed framework explicitly captures risk-specific causal features through graph-regularized causal discovery method, enabling the selection of an optimal subset of causal features. Two complementary tasks are jointly learned: an overlapping component for tumor region detection and a classification component for aggregating representative causal features for risk prediction. Theoretically, the proposed causal network mitigates distribution shift effects induced by hetero geneous MRI data. Extensive experiments on three public datasets demonstrate that our method consistently outperforms state-of-the-art approaches.

Compared to driver-specific and vehicle-specific characteristics, physiological signal features exhibit superior efficacy in detecting driver fatigue, particularly electroencephalogram (EEG) signals, which are less influenced by subjective human factors and directly reflect brain neural activity. Currently, many detection methods primarily focus on the analysis of EEG signals but fail to consider the interdependencies between signal acquisition channels. To improve the accuracy of driver state detection, a Graph Attention Convolutional Neural Network (GAT-CNN) is proposed in this study. The method incorporates channel relationships and performs an end-to-end learning process, which eliminates the need for manual feature extraction. The preprocessed EEG signals and the adjacency matrix based on mutual information serving as the input to the GAT-CNN. Finally, the EEG signals are classified into two states namely alert and fatigued by fully connected layers and a softmax classification layer. The performance of the method is validated on the SEED-VIG dataset with an average accuracy of 90.14 %, a peak accuracy of 99.23 % and an average F1 score of 91.54 %. Additionally, the Brier Score, used as an evaluation metric, yields an average value of 0.0841, which indicates high predictive accuracy and strong generalization ability. Compared to existing state-of-the-art methods, the GAT-CNN demonstrates superior performance.

Radiology reports contain rich diagnostic information beyond simple binary labels, including probability assessments, anatomical locations, and severity grades. Recent large language model annotation pipelines, such as MAPLEZ, automatically extract multiple annotation types, but current approaches integrate them via additive multi-task loss formulations that suffer from gradient interference and ignore the heterogeneous reliability of each annotation source. We address this gap by reframing multi-type clinical annotations as a multi-source information fusion problem and propose UFCA, a hierarchical fusion framework operating at three levels. At Level 1, multi-head cross-attention with uncertainty-aware gating fuses categorical labels, probability scores, location descriptions, and severity grades into a unified annotation embedding. At Level 2, spatial attention guided by anatomical priors and FiLM conditioning integrates this embedding with visual features. At Level 3, a pathology-adaptive weighting layer combines vision-only, annotation-conditioned, and annotation-only prediction branches. We evaluated UFCA on the NIH ChestX-ray14 dataset (112,120 images, 13 pathologies) using MAPLEZ annotations, with external validation on the Kermany pediatric pneumonia dataset and the IQ-OTH/NCCD lung CT dataset. UFCA achieved a macro-averaged AUROC of 0.838, an improvement of 2.3 %age points over MAPLEZ-MTL (p < 0.001, DeLong's test, Bonferroni-corrected). The largest gains are seen in rare pathologies, including fractures (+4.2 pp) and pneumonia (+3.9 pp). Severity annotations, previously reported as harmful under additive loss, contribute + 0.7 pp under the proposed fusion. The framework adds 11.7% parameters and 23% training time over the multi-task baseline.

Although recent works in fully supervised learning has achieved significant success in medical image segmentation, obtaining high-quality pixel-wise expert annotations remains a challenge in medical imaging. Therefore, semi-supervised learning (SemiSL) has recently attract increasing attention to medical image segmentation. Contrastive learning (CL) framework has demonstrated strong inter-class separability in the field of image segmentation. However, its application in medical imaging is impacted by the long-tailed distribution of data, leading to imbalanced learning. To this end, We propose a novel Own-background Contrastive Learning (OBCL) framework for semi-supervised medical segmentation. Unlike other CL frameworks, OBCL can effectively incorporate background pixels into CL and avoid being constrained by the long-tailed distribution of the data. We leverage a student-teacher model to generate pseudo-labels, guiding the creation of foreground- background feature pairs, then use of discriminative class information learned in CL to produce accurate multi-class segmentation. Specifically, we decompose a multi-class segmentation task into multiple binary segmentation tasks, each focusing on segmenting a specific foreground class and the background in the images. Additionally, we integrate a small fraction of foreground information into the background features to improve inter-class separability, called Own-background. We achieve significant improvements on both cardiac MRI and colonoscopy polyp segmentation tasks compared with state-of-the-art methods. Even with only a 5% labeling rate for model training, we achieved 88.30% Dice on the ACDC dataset.

An enormous worldwide health problem, cervical cancer is defined by the uncontrolled proliferation of cervix cells that may spread to other parts of the body. Timely discovery has the ability to cure the condition, hence effective treatment of cervical cancer needs early detection. One of the most important screening tools for early detection is the inexpensive Pap smear. Clinical image processing is improved by computer-aided diagnostic (CAD) approaches, which speeds up and improves cancer diagnosis. Variation in image look, morphology, and size, as well as problems with data availability and quality, pose obstacles to deep learning approaches for cervical cancer classification. We use deep learning in two ways: first, by extracting features from pre-trained models using a variety of machine learning techniques for image classification; and second, by applying transfer learning to cervical cancer images using pre-trained models. An innovative approach is introduced here that improves classification by combining ResNet50 and VGG19 architectures. This all-encompassing plan aims to make cervical cancer screening more efficient and accessible over the world while simultaneously improving the accuracy of diagnoses. The proposed method achieved accuracy of 94.53% and F1 score of 93% using transfer learning and pretrained model (ResNet50 and VGG19), thus achieved superior classification performance by outperforming existing methods.

In this study, a publicly available radiological image dataset is used for stroke classification and segmentation tasks. The dataset includes brain Non-Contrast Computed Tomography (NCCT) images of patients with different ages, genders and clinical conditions. In our study, a transformer-based model is preferred, the most effective 40 radiomic features detected from NCCT images are extracted, and the effects of transformer architecture on the classification of ischemic stroke lethality are observed with loss functions and Simple Attention, Multi Head Attention mechanisms. Our RadiomicsTransformer model, with radiomics selection, focal loss, 5-fold cross-validation and simple attention applications, provides 0,6889 accuracy, 0,693 precision, 0,6778 sensitivity, 0,6854 f1-score and 0,7167 ROC AUG values. The results demonstrate that, on imbalanced and limited datasets, the transformer-based DL model outperforms classical ML models, and that our transformer architecture holds clinically testable potential for stroke diagnosis.

This study investigated the classification of bearing faults in electric machines using machine learning and optimization techniques. A Random Forest (RF) classifier was first applied to the dataset, followed by the integration of various metaheuristic algorithms, including Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Ant Colony Optimization (ACO), Artificial Bee Colony (ABC), and the Jellyfish Algorithm (JFA). These algorithms were employed to optimize the hyperparameters of the RF model and improve classification performance. The models were evaluated based on accuracy, F1 score, precision, and recall. Experimental results demonstrated that the RF-ACO hybrid model reached the highest performance across all metrics, with an accuracy of 93.53%, outperforming the other approaches. The obtained results demonstrated that optimizing algorithms significantly enhances the fault detection capability of the Random Forest classifier. This study provides valuable insights into the application of hybrid machine-learning models for effective fault diagnosis in electric motor systems.

Multiple Sclerosis (MS) is a chronic condition caused by the immune system attacking the central nervous system. This process results in demyelination and axonal damage in the brain and spinal cord, leading to the formation of MS lesions. Magnetic Resonance Imaging (MRI) is widely used for diagnosing and monitoring the disease. However, manually labelling lesions is a time-consuming process and subject to variability among experts. In this study, a deep learning-based 3D U-Net model was employed for MS lesion segmentation, and the influence of differing expert annotations on model performance was examined. The ISBI 2015 dataset, annotated by two independent experts, was used for this purpose. The model's training outcomes were evaluated using the Dice Coefficient (F1 Score) and IoU Score (Jaccard Index). The findings underscore the discrepancies between expert opinions. The results of this study indicate that variations in expert annotations impact the model's training performance and reduce its generalizability.

This study presents a novel technique for the segmentation and classification of regions in breast tumour images by integrating a convergence-based density model, coupled with texture feature-based clustering. The segmentation process starts with an active contour that estimates probability densities for foreground, background, and tumour regions. A key advantage of the proposed method is its independence from an annotated training set. Thus, it reduces sensitivity to dataset variability by using intensity-driven convergence. To address overlapping structures in mammographic images, an edge-path contour-splitting methodology is employed for accurate boundary separation. Finally, a probabilistic neural network (PNN) is used to classify the tumour regions based on texture features. Both qualitative and quantitative results are presented to demonstrate the effectiveness of the proposed method. The proposed method achieves an accuracy of 92