Brain tumor segmentation in magnetic resonance imaging (MRI) requires a careful trade-off between accuracy and computational efficiency. In this study a hybrid deep learning framework that integrates Swin Transformer and three dimensional (3D) Convolution is proposed to leverage both global and local feature representations. The encoder incorporates 3D Patch Aggregation and hierarchical feature compression using 3D Patch Merging, while feature recalibration is enhanced through Squeeze-and-Excitation (SE) modules. The bottleneck combines Kolmogorov-Arnold Networks (KAN) with Axial Attention for improved nonlinear modeling and long-range dependency capture. Multi-scale skip connections use Feature Compression Module (FCM) and Cross Attention Fusion (CAF) to strengthen feature alignment, and the decoder reconstructs segmentation masks using transposed convolutions with refined outputs via a Convolutional Block Attention Module (CBAM). A Teacher-Student Knowledge Distillation (KD) strategy enables the development of a lightweight Student model that utilizes Conv3D, Efficient Channel Attention (ECA), and a KAN bottleneck to preserve performance with reduced complexity. The proposed models are evaluated on the BraTS 2020, 2021, and 2023 datasets and demonstrate strong generalization on the MSD dataset without fine-tuning. The Teacher model achieves Dice scores up to 92.85% for whole tumor segmentation, while the Student model maintains competitive accuracy with only 2.804 million parameters and 30 GFLOPs. Visual and quantitative results confirm the framework's robustness and effectiveness across multiple benchmarks in brain tumor segmentation.. For reproducibility, the source code is publicly available at: https://github.com/Rahman768/RCH.

This paper focuses on the ideological and political education in artificial intelligence (AI) specialized courses. A fusion model characterized by knowledge height (disciplinary expertise), knowledge temperature (value orientation) and knowledge length (holistic education) is constructed, emphasizing the organic unity of knowledge transmission and value cultivation. Firstly, through cross-team collaboration, a tripartite educational team comprising disciplinary instructors, ideological and political educators, and student counselors is established, alongside a collective lesson-planning mechanism. Secondly, via pedagogical integration, the objective of socialist core values is operationalized and decomposed into course syllabi, through diverse approaches including case studies, interactive discussions, and digital tools. Finally, through assessment innovation, ideological and political goals are embedded into the full-cycle evaluation system spanning in-class performance to final examinations. Long-term practice demonstrates that this model has significantly enhanced students’ learning motivation and comprehensive competence, offering an all-participant, full-process comprehensive ideological and political education system.

In recent years, medical image segmentation tasks have made significant progress, and the models based on convolutional neural networks (CNNs), represented by U-Net, have become the mainstream architectures. However, the inherent local perception property of CNNs leads to difficulties in realizing long-range information interactions. Although the self-attention mechanism efficiently models long-range dependencies, its computational complexity grows quadratically. In addition, existing attention mechanisms still have significant shortcomings in multi-dimensional global-local feature interaction and multi-scale association modeling. This paper proposes the dual heterodimensional context fusion network (DHCF-Net), the core of which is the heterodimensional complementary context fusion (HCCF) and the heterodimensional multi-scale context fusion (HMCF) modules. The HCCF module constructs complementary global-local context representations in the channel and spatial dimensions. The HMCF module is designed with a dual-path synergy mechanism for multi-scale context capture and interaction along channel and spatial dimensions. Experiments on ISIC 2018, BUSI, and Kvasir-SEG datasets demonstrate that DHCF-Net achieves state-of-the-art performance.

The diagnosis of mental disorders using medical imaging remains a significant challenge due to the inherently heterogeneous and multi-scale organization of brain activity. Accurately modeling the complex structural patterns embedded in neuroimaging data requires integrating diverse representations, which capture complementary aspects of brain function and anatomy. However, effectively combining such heterogeneous information while preserving spatial dependencies and structural semantics remains a non-trivial task. To address these limitations, we propose T-IMHF, a novel Transformer-based integrated multi-view heterogeneous feature fusion framework for medical image-based classification of mental disorders. T-IMHF employs a multi-branch architecture that independently processes Rs-fMRI features from diverse structural views (graph-, voxel-, and region-level), projecting them into a shared latent space for joint representation learning. The framework leverages multi-level brain representations by combining the structural topology of graph data with fine-grained voxel-wise patterns and high-level region-based descriptors. Transformer Unity (Transformer-U) module integrates structure-aware attention mechanisms with cross-view alignment strategies to achieve robust and interpretable cross-structural feature fusion. Extensive experiments on two public datasets demonstrate the superior performance of T-IMHF, achieving AUC scores of 89.56% on autism spectrum disorder (ASD) and 90.14% on major depressive disorder (MDD), significantly outperforming state-of-the-art baselines. Furthermore, T-IMHF successfully identifies biologically meaningful biomarkers that align with findings in current neuroscience literature, highlighting its potential as a reliable and interpretable diagnostic tool. These results validate the effectiveness and generalizability of the proposed framework in capturing heterogeneous brain representations for mental disorder diagnosis.

Depression is often accompanied by impairments in reward-punishment feedback processing, yet the underlying complex-system mechanisms of brain functional networks remain unclear. Complex network theory provides a novel paradigm for characterizing the dynamics of integration and segregation in brain networks and for uncovering nonlinear topological alterations in psychiatric disorders. In this study, we constructed brain functional networks of patients with depression and healthy controls based on task-related fMRI data from a gambling reward-punishment paradigm. Using graph-theoretical and complex network approaches, we evaluated key topological measures, including global efficiency, local efficiency, and small-world properties, and further examined the coupling between network topology and behavioral reaction times. Individuals with depression exhibited significantly reduced network integration capacity, manifested by decreased global and local efficiency, which progressively worsened with clinical severity. In contrast, small-world properties in certain regions showed relative enhancement, which may reflect local reorganization or compensatory tendencies within the network rather than a definitive adaptive mechanism. Moreover, the degree of topological disruption was negatively correlated with reaction time, revealing a coupling between cognitive slowing and information transfer efficiency in complex networks. Overall, depression in reward-punishment contexts is characterized by a nonlinear disruption of the integration-segregation balance of brain networks, with impaired global transmission efficiency and adaptive modulation of local modularity and small-world organization. The results not only delineate the multilayered abnormalities of brain networks in depression but also indicate that complex network theory may provide novel topological biomarkers for the diagnosis and prediction of psychiatric disorders.

Major depressive disorder (MDD) is a mental illness with a significantly higher prevalence, especially among the youth population. In recent years, with the continuous innovation of deep learning technologies, the combination of medical imaging and computer-aided diagnosis has opened up new avenues and possibilities for studying a variety of diseases. However, many challenges are still faced in the research for the disease of depression. Among them, how to effectively improve the performance of network models under the condition of limited and small datasets is an urgent and crucial topic. Therefore, this paper proposes an auxiliary diagnosis method for depression, combining a lightweight deep learning network and a multi-modal feature fusion strategy, aiming to achieve more accurate and efficient depression diagnosis. Specifically, a lightweight dual-path architecture called 3D-CBSResNet12 is constructed, which utilizes multiple effective convolutional approaches as its core operations and employs a multi-modal feature fusion strategy to achieve more comprehensive and accurate classification of depression. Experimental results show that 3D-CBSResNet12 achieves significant performance improvement, fully validating its feasibility and effectiveness.

With the rapid advancement of artificial intelligence, deep learning has become a core technology for automated brain tumor segmentation via Magnetic Resonance Imaging (MRI), providing critical morphological and positional information for clinical diagnosis and treatment planning—a typical medical computer vision application. However, scarce and imperfect high-quality annotated 3D MRI datasets severely restrict the training efficiency, generalization ability, and clinical utility of deep learning segmentation models. To address this challenge, this study proposes a novel Tensor Erasing (TE) data augmentation method for 3D multi-modal brain tumor MRI segmentation, a generalizable strategy to enhance model robustness under data constraints. The core of TE is randomly selecting a specific-size 3D region from tensor-formatted MRI data, erasing voxel values by setting them to zero, and using these erased images as augmented training data. This simulates noisy or corrupted clinical images, forcing the model to focus on global tumor structures rather than local features, improving robustness and adaptability. Validated on the BraTS2019 dataset with 3D U-Net, Attention-UNet, and VGG-Net, TE consistently improves model performance. Specifically, the TE-augmented Attention-UNet achieves Dice coefficients of 87.15%, 82.77%, and 78.28% for whole tumor, tumor core, and enhancing tumor, respectively. As a low-cost, easy-to-implement strategy, TE alleviates data scarcity, enhances generalization, and can be extended to other computer vision tasks, laying a foundation for real-world deployment of intelligent vision systems.

Multimodal image fusion is a promising technology designed to fuse information from different medical sensors, which offer structured insights for disease diagnosis and treatment. However, existing 2D-centric fusion methods fail to capture 3D spatial continuity, and conventional wavelet-based approaches lack adaptability to diverse lesion regions and suffer from background artifacts. To address this issue, we propose a 3D multiscale wavelet convolutional neural network for multimodal medical image fusion. Specifically, a 3D Discrete Wavelet Transformation (3D DWT) is introduced to decompose input volumes into multi-frequency bands, isolating anatomical structures and lesion details while reducing 3D spatial redundancy. We embed hierarchical multiple frequency band into a Global and Local Feature Calibration (GLFC) module to adaptively enhance single-modal features by fusing global contextual information and local details. Furthermore, a pyramid group-wise multiscale feature interaction is proposed for capturing complementary features across different spatial scales. Finally, a voxel-wise weighted averaging strategy reconstructs the fused image by adaptively assigning contributions to each modality at every spatial position, effectively eliminating artifacts and improving the visual fidelity of the result. Extensive experiments on the BraTS2020 and Hecktor datasets demonstrate that our proposed method outperforms state-of-the-art (SOTA) fusion methods in both subjective visual quality and objective metrics. Moreover, downstream segmentation validation confirms that fused images from our method significantly improve tumor segmentation accuracy. The source code and pre-trained models will be publicly available.

Depression is a common mental illness that not only profoundly infests the psychological state of patients, but also tends to cause damage to the functioning of patients' brain areas. To construct a comprehensive and detailed framework for a supporting diagnostic network that will help physicians make accurate and timely diagnoses when dealing with patients at different stages of depression, a network model based on three-dimensional (3D) weight group MobileNet (3D-WGMobileNet) and transfer learningis proposed. Firstly, fMRI data is preprocessed, and regional homogeneity analysis is used to reduce the dimension of the image. Then, the characteristics of Alzheimer's disease are learned by transfer learning and transferred to the proposed model. Next, the dynamic group convolution was used to construct the expert weight matrix of the convolution kernel, and the sliding window group convolution was used to compress the parameters of the model to improve the expression ability and computing power of the model. By using 5-fold cross-validation, we conducted experiments using data from HCP and REST-meta-MDD. The experiment results show that the proposed model gives a superior performance compared with other state-of-the-art methods, especially on the classification of the healthy group with major depression groups, where the two datasets achieve 88% and 91% accuracy, respectively, which verifies the feasibility and effectiveness of our model.

Due to prolonged exposure to harsh operational environments, wind turbines tend to experience a high incidence of failure. However, effective fault data are usually hard to obtain in real applications. To remediate this complication, this study introduces a few-shot learning-based graph neural network for fault diagnosis of wind turbines with current signals. Initially, one-dimensional current signals from the generator of the wind turbine are transformed into two-dimensional data using Continuous Wavelet Transform (CWT). Subsequently, features are extracted through Convolutional Neural Network (CNN), which are input into Graph Neural Network Based on Few-Shot Learning (GNN-FSL) for fault diagnosis with a small dataset. Experiments were conducted using real wind turbine data and the public CWRU dataset, where ResNet50, EfficientNet-B0, DenseNet121, CNN, and GNN were selected for comparative evaluation. Under optimal parameter settings, the proposed model achieved superior performance compared to the baseline models. The results demonstrate that the proposed method outperforms all previously mentioned approaches.

Segmenting brain tumors is important for effective diagnosis and treatment planning. Conventional 3D segmentation models achieve high accuracy but are computationally intensive, often limiting real-time applicability. In this study, pseudo-3D convolutions, which consist of spatial and depthwise convolutions, are used in place of traditional 3D convolutions. Adaptive Dilated Multi-Fiber (DMF) units dynamically extract multi-scale features and parallel Multi-Fiber (MF) units combine them with weighted sum. Efficient Channel Attention (ECA) and Cross Attention improve feature selection and fusion in decoder and encoder. Structured pruning reduces superfluous parameters and Quantization Aware Training (QAT) increases the speed of inference with the model converted to INT8 precision. Combination of Dice Loss and Boundary Loss enhances the precision of tumor boundaries. The framework has been evaluated on the BraTS 2021 data validation set and achieved high Dice scores of Whole Tumor 91.85%, Tumor Core 88.52%, and Enhancing Tumor 85.55%, with Hausdorff95 values of 2.58 mm, 3.53 mm, and 3.65 mm. Our proposed model requires only 3.57M parameters and 21.26 GFLOPs, achieving an inference time of 0.016 seconds per 3D volume while maintaining precision alongside efficiency to clinical application.

With the accelerating pace of urbanization, urban rail transit systems are playing an increasingly vital role for alleviating traffic congestion, and their operational reliability and safety have been paid significant attention. This study addresses the risk of cascading failures faced by urban rail transit (URT) networks under natural disasters and unexpected events. A resilience assessment model— URT-CML—is proposed, based on the Coupled Map Lattice (CML) theory. This model innovatively integrates station heterogeneity, dynamic load redistribution mechanisms, and multi-modal redundant connectivity to construct a dynamic propagation model of cascading failures in URT networks. Monte Carlo simulations are employed to model failure pathways under various attack strategies, and system resilience is evaluated from the perspectives of robustness and service coverage. Using the Beijing metro network as a case study, results demonstrate that the URT-CML model improves the prediction accuracy of failure propagation speed by 22.7% compared to conventional models, offering new theoretical foundations and decision-making support for resilience-oriented design and emergency management in urban rail transit systems.

Due to the harsh working conditions of wind turbines, it is difficult to obtain sufficient fault data from the gearbox. To address this issue, this study proposes a graph neural network model based on few-shot learning (GNN-FSL) for the fault diagnosis on planetary gears of wind turbine gearbox with acoustic signals. The short-time Fourier transform (STFT) is chosen for preprocessing of acoustic signals as input data, which converts the raw data into two-dimensional data. Then, convolutional neural networks (CNN) are used to extract data features. Finally, the extracted features are input for the graph neural network for fault classification. The experimental results show that the model performs well in fault classification on small sample datasets, providing an effective method for fault diagnosis of planetary gears in wind turbines. By comparing EfficientNet-b0, ResNet-50, Densenet121, Mobilenet_v2 and DiffKendall models, it is verified that the proposed method is effective in planetary gear fault classification, and its performance is superior to all these networks.

Brain tumor segmentation in MRI images is crucial for clinical diagnosis and treatment planning but those scans are usually affected by imaging artifacts which decrease the quality of data and hamper segmentation performance. To address these challenges, this study proposed a unique framework that seamlessly combines artifact correction with segmentation of tumors. The framework features a data preparation module which is able to prepare realistic artifact-contaminated and artifact-free MRI image pairs that have been used for training. It also includes a diffuse model which acts on MRI images and removes the artifacts thus giving high-quality inputs for segmentation. In ad-dition, a modified 3D Convolutional Neural Network (CNN) architecture which integrates attention blocks and squeeze-and-excitation (SE) layers is used to segment the tumor sub-regions, including the enhancing tumor (ET), tumor core (TC), and whole tumor (WT). The framework was evaluated with artifact-corrupted data and clean data and achieved better results regarding the generation of artifact-free data and stable segmentation than the other baseline methods. This method emphasizes the magnitude of imaging artifacts on MRI-based segmentation and facilitates improvement in the clinical workflows. The code is available at https://github.com/Rahman3175/MMR

The diagnosis and analysis of major depressive disorder (MDD) faces some intractable challenges such as dataset limitations and clinical variability. Resting-state functional magnetic resonance imaging (Rs-fMRI) can reflect the fluctuation data of brain activity in a resting state, which can find the interrelationships, functional connections, and network characteristics among brain regions of the patients. In this paper, a brain functional connectivity matrix is constructed using Pearson correlation based on the characteristics of multi-site Rs-fMRI data and brain atlas, and an adaptive propagation operator graph convolutional network (APO-GCN) model is designed. The APO-GCN model can automatically adjust the propagation operator in each hidden layer according to the data features to control the expressive power of the model. By adaptively learning effective information in the graph, this model significantly improves its ability to capture complex graph structural patterns. The experimental results on Rs-fMRI data from 1601 participants (830 MDD and 771 HC) and 16 sites of REST-meta-MDD project show that the APO-GCN achieved a classification accuracy of 91.8%, outperforming those of the state-of-the-art classifier methods. The classification process is driven by multiple significant brain regions, and our method further reveals functional connectivity abnormalities between these brain regions, which are important biomarkers of classification. It is worth noting that the brain regions identified by the classifier and the networks involved are consistent with existing research results, which suggest that the pathogenesis of depression may be related to dysfunction of multiple brain networks.

Early fault detection (EFD) presents significant challenges due to the subtle nature of abnormal signals during this phase. Traditional EFD methods primarily focus on identifying the onset time of failures. In this paper, we propose a novel EFD approach that simultaneously achieves early fault detection and health stage classification. We begin by developing a threshold-free fault detector utilizing adversarial learning, enabling the network to perform supervised classification and unsupervised fault detection concurrently. To further enhance the stability of the adversarial network, we introduce a straightforward ensemble learning technique. A pivotal aspect of this study is the conceptualization of the EFD problem as fault detection with continuously emerging new classes (FDENC), wherein different stages of a fault are treated as new classes. To support this framework, we investigate a self-model update mechanism based on the designed network. Unlike traditional EFD paradigms, our proposed method not only identifies the onset of early failures but also facilitates health stage classification, thereby indicating the progression of the fault. We evaluate the effectiveness of our approach through two run-to-failure experiments designed to induce sub-surface cracks. The experimental results demonstrate the method's sensitivity to the emergence and progression of micro-cracks in bearings. Additionally, we employed two benchmark datasets to assess the method's performance, further highlighting its superiority in both fault detection and health stage classification.

In recent years, target tracking algorithms based on deep learning have made a lot of progress in the field of target tracking. This paper mainly introduces the research of traffic incident detection algorithm. Firstly, the overall traffic incident detection algorithm flow is introduced and analyzed in detail, and then the model framework and algorithm flow of ByteTrack algorithm are introduced. Then, the improved ByteTrack algorithm is described compared with the previous version, including the improvement of the backbone network and the optimization of the network module. Next, the algorithm flow chart of the traffic incident detection algorithm is introduced, and the general process of the whole traffic incident detection is summarized through the traffic incident detection flow chart.

Brain tumors refer to abnormal cell proliferation formed in brain tissue, which can cause neurological dysfunction and cognitive impairment, posing a serious threat to human health. Therefore, it becomes a very challenging work to full-automaticly segment brain tumors using computers because of the mutual infiltration and fuzzy boundary between the focus areas and the normal brain tissue. To address the above issues, a segmentation method which integrates edge features is proposed in this paper. The overall segmentation architecture follows the encoder decoder structure, extracting rich features from the encoder. The first two layers of features are input to the edge attention module, and to extract tumor edge features which are fully fused with the features of the decoder segment. At the same time, an adaptive weighted mixed loss function is introduced to train the network by adaptively adjusting the weights of different loss parts in the training process. Relevant experiments were carried out using the public brain tumor data set. The Dice mean values of the proposed segmentation model in the whole tumor area (WT), the core tumor area (TC), and the enhancing tumor area (ET) reach 91.10%, 87.16%, and 88.86%, respectively, and the mean values of Hausdorff distance are 3.92, 5.12, and 1.92 mm, respectively. The experimental results showed that the proposed method can significantly improve segmentation accuracy, especially the segmentation effect of the edge part.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI

Condition monitoring (CM) process can be viewed as the problem of streaming classification with emerging new classes (SENC). There are four fundamental challenges we are faced with: (1) timely detection of emerging new classes; (2) model update to adapt to new classes; (3) pattern recognition of these already known classes with high accuracy and (4) distinguishing between different new classes. Although intelligent surrogate models like deep learning have achieved remarkable success in the field of CM, SENC still remains a thorny challenge that has seldom been studied. This paper presents an approach to address the challenge of SENC using an ensembled multi-classification generative adversarial network (EMC-GAN). The proposed method includes a novel deep network architecture called MC-GAN that integrates the tasks of novelty detection and multi-classification into a single framework. To address issues of model stability, an efficient history-state ensemble (HSE) method that does not require additional training costs to generate multiple base models is introduced. Experimental validation is conducted on four simulated SENC tasks using benchmark data, and the results have shown the effectiveness of the proposed approach.