Cardiovascular disease remains a leading cause of global mortality, underscoring the critical need for accurate and efficient diagnosis of cardiac conditions. Although the electrocardiogram (ECG) is a widely used diagnostic tool, its utility is often limited by signal artifacts and phenotypic similarities among distinct pathologies. To address these limitations, we propose a novel Hybrid Dynamic Graph Convolutional Network (HDGCN) to detect multi-lead ECG arrhythmia. In our HDGCN, ECG signals are represented as graph structures, where nodes correspond to sampling points and edges encode spatiotemporal relationships. Local spatiotemporal features are extracted using a pre-trained ResNet module to capture subtle morphological variations, while an adaptive graph convolutional module with a learnable adjacency matrix dynamically models deeper inter-lead dependencies. Additionally, wavelet-based denoising is applied during preprocessing to preserve clinically relevant features, and depthwise separable convolutions are incorporated to substantially reduce computational complexity. Experimental results demonstrate that HDGCN achieves an average accuracy of 99.57%. The model exhibits robust performance in detecting complex arrhythmias, notably attaining an improvement exceeding 6% for categories such as atrial premature beats.

Intelligent fault diagnosis is crucial for ensuring the safety and reliability of modern industrial systems. However, the performance of deep learning models often significantly degrades due to the domain shift between training and testing data. Domain Adaptation (DA) methods, particularly bi-classifier adversarial networks, have proven effective in transferring knowledge from a labeled source domain to an unlabeled target domain. However, existing approaches often pay insufficient attention to target sample prediction accuracy, resulting in reduced feature discriminability and generalization. Additionally, due to the absence of labeled target data, most approaches rely on pseudo-labels, which are often noisy and unreliable, especially in the early stages of training. To address these issues, this paper proposes a novel uncertainty-guided denoising bi-classifier adversarial domain adaptation network (UGDBAN) for cross-domain fault diagnosis. Specifically, a feature generator based on Transformer layers is designed to capture long-range dependencies and local features. To mitigate the impact of noisy pseudo-labels, an uncertainty-based denoising pseudo-labeling mechanism is introduced to enhance the discriminability of features by redefining pseudo-labels and dynamically selecting high-confidence samples as clean samples. Building upon this denoised pseudo-label set, a Dirichlet uncertainty estimation-based class prototype alignment strategy is proposed to align domain features at the class level by selecting low-uncertainty samples representative of each class as prototypes. Extensive experiments demonstrate the effectiveness of UGDBAN, and comparative results with mainstream methods highlight its superiority.

Open-Vocabulary Multi-Label Image Classification (OV-MLIC) is an emerging task in computer vision aimed at recognizing unseen categories in real-world scenarios, leveraging Vision and Language Pre-training (VLP) models like CLIP. However, existing methods overlook the impact of category coupling and scale variation on cross-category knowledge transfer, thereby restricting performance on unseen categories. To address this issue, we propose a novel OV-MLIC method called Hierarchical Dual-Granularity Alignment-CLIP (HDG-CLIP), which emphasizes the complementary characteristics of different modalities and introduces a sample-category matching mechanism. Specifically, to address the category coupling issue, we construct semantic category prototypes to enhance cross-category knowledge transfer. Through the interaction between visual embeddings and category prototypes, we decouple category-specific information from mixed visual features and leverage the visual context of samples to learn category-level visual features. For mitigating the scale variation issue, we build a sample-category dual-granularity matching mechanism based on the difference in capture capability of different modalities across scales, thereby improving the object localization accuracy from a multi-dimensional perspective. Extensive experimental results show that HDG-CLIP exhibits state-of-art performance over existing methods on both the NUS-WIDE and the Open-Images datasets. Our code is available at https://github.com/wakihy/HDG-CLIP.

Graph Neural Networks (GNNs) have achieved remarkable success in modeling graph-structured data. However, in real-world scenarios, graph data often arrive continuously, requiring models to learn new tasks without forgetting previously acquired knowledge. Existing Continual Graph Learning (CGL) methods commonly employ knowledge distillation to alleviate catastrophic forgetting, yet most of them concentrate on preserving output logits or local structural information (e.g., first-order neighborhoods). This restricts their ability to capture the global topology and latent geometric relationships within the graph. To address this challenge, we propose a novel method named Continual Graph Learning with Topology-aware Knowledge Distillation (CGL-TKD). CGL-TKD introduces a structure relation matrix as a unified carrier of structural knowledge, explicitly encoding dense pairwise node dependencies to capture global geometric characteristics beyond local neighborhoods. Within a dual-stream replay framework, CGL-TKD aligns the current-task model with the previously stored model snapshot, ensuring consistent structural relations across tasks while avoiding catastrophic drift. Extensive experiments on benchmark datasets demonstrate that CGL-TKD significantly mitigates catastrophic forgetting and consistently outperforms state-of-the-art baselines.

Age-related macular degeneration (AMD) is a leading cause of visual impairment in middle-aged and elderly populations. Existing diagnostic methods are often inefficient and subjective. To achieve accurate and efficient diagnosis of AMD lesions, this paper designs BELD-Net, an efficient segmentation network based on a large visual model. The model employs DINOv2 as its backbone to capture long-range and cross-scale features of AMD lesions through large-scale self-supervised pre-training. Combined with the query-based segmentation mechanism of Mask2Former, it significantly improves generalization in few-shot scenarios. BELD-Net incorporates LoRA modules into the Transformer layers of DINOv2 for low-rank adaptation, reducing computational cost and enhancing the model’s adaptability to fundus images while avoiding overfitting. Additionally, a Multi-scale Boundary Enhancement Module (MBEM) is designed using heterogeneous convolutions and an adaptive squeeze-and-excitation (SE) mechanism to address blurred and variable-sized lesion boundaries, thereby improving sensitivity to low-contrast edges. Experimental results demonstrate that BELD-Net achieves outstanding performance in automatic multi-lesion segmentation of AMD, with an mDice value of 59.61

OBJECTIVE:To design and implement a deep learning-based multimodal data fusion classification model that integrates laryngoscope images and voice signals to improve the diagnostic accuracy of laryngeal diseases, enabling rapid and precise identification for clinical support. RESULTS:The model demonstrated high classification accuracy and robustness, achieving an overall accuracy of 87.92 % on the independent test set. Precision, recall, specificity, and F1-score were 0.879, 0.887, 0.966, and 0.883, respectively. The model outperformed single-modal approaches and existing multimodal frameworks. CONCLUSION:The proposed HSNet effectively integrates hierarchical features from laryngoscope images and voice modalities, enabling accurate classification of six laryngeal diseases. This method holds significant potential for clinical applications.

With the continuous improvement of processes by textile manufacturers, the requirements for pilling of fabric materials are also increasing. At present, in the textile industry, the rating method for fabric pilling is to compare the sample with the standard sample in a specific environment to determine the grade of fabric pilling. However, this method is greatly influenced by subjective factors, and when facing a large number of samples, it consumes a large amount of manpower and material resources. Therefore, an objective, stable, and highly accurate rating method is needed. This paper proposes an end-to-end objective rating method based on a convolutional neural network model called DENet. The network employs a two-branch architecture that integrates an attention mechanism with multiscale convolution and connects large convolutional kernels across layers. To align with the subjective grading criteria for fabric pilling, this study simulates the pilling process of eight types of fabrics using two methods: the pilling box method and the circular trajectory pilling method. Consequently, eight fabric pilling image datasets were generated, covering knitted fabrics, woven fabrics, and nonwoven fabrics, including two patterned fabric types. The experiment results show that the accuracy of the rating method proposed in this paper is 97.95% for combined training on eight datasets, and the average accuracy is 98.07% for separate training on eight datasets, and the accuracy and generalization of the designed network model is high. The proposed system provides an efficient, objective, and stable solution for fabric pilling evaluation.

Multisource domain adaptation aims to transfer knowledge from multiple source domains to improve the generalization performance of fault diagnosis. However, most existing approaches treat all source domains equally, which may cause negative transfer by ignoring interdomain discrepancies, when some source domains are less relevant to the target domain. To address this, this article proposes a multisource importance-based hierarchical adaptation network (MSIHAN) for cross-domain fault diagnosis. MSIHAN employs a domain importance recognition module to estimate the probability that each target sample belongs to each source domain. Based on the estimated importance, weighted moment matching and weighted class-aware distribution alignment are presented to achieve marginal and conditional distribution alignment by prioritizing source domains with higher importance. In addition, the weighted classification loss is introduced to constrain the feature generator to generate aligned and distinguishable representations. Extensive experiments demonstrate the effectiveness of MSIHAN, with comparative results highlighting its superiority over mainstream methods.

Heterogeneous Graph Neural Networks (HGNNs), have demonstrated excellent capabilities in processing heterogeneous information networks. Self-supervised learning on heterogeneous graphs, especially contrastive self-supervised strategy, shows great potential when there are no labels. However, this approach requires the use of carefully designed graph augmentation strategies and the selection of positive and negative samples. Determining the exact level of similarity between sample pairs is non-trivial.To solve this problem, we propose a novel self-supervised Heterogeneous graph neural network with Optimal Transport (HGOT) method which is designed to facilitate self-supervised learning for heterogeneous graphs without graph augmentation strategies. Different from traditional contrastive self-supervised learning, HGOT employs the optimal transport mechanism to relieve the laborious sampling process of positive and negative samples. Specifically, we design an aggregating view (central view) to integrate the semantic information contained in the views represented by different meta-paths (branch views). Then, we introduce an optimal transport plan to identify the transport relationship between the semantics contained in the branch view and the central view. This allows the optimal transport plan between graphs to align with the representations, forcing the encoder to learn node representations that are more similar to the graph space and of higher quality. Extensive experiments on four real-world datasets demonstrate that our proposed HGOT model can achieve state-of-the-art performance on various downstream tasks. In particular, in the node classification task, HGOT achieves an average of more than 6% improvement in accuracy compared with state-of-the-art methods.

Cancer is generally thought to be caused by the accumulation of mutations in driver genes. The identification of cancer driver genes is crucial for cancer research, diagnosis and treatment. Despite existing methods, challenges remain in comprehensively learning of the attributes and intricate interactions of genetic data. We propose a novel Multi-information Fusion Graph Convolutional Network (MF-GCN) for cancer driver gene identification, based on multi-omics pan-cancer data and Gene Regulatory Network (GRN) data. Directed topological and attribute graph networks learn gene interactions and self-attribute information, while a common graph network captures consistency between topology and attributes. An attention mechanism adaptively fuses these information with importance weights to identify cancer driver genes. Experimental results showed that MF-GCN can effectively identify cancer driver genes across three GRN datasets, with AUROC and AUPRC improvements of 2.66% and 2.69%, respectively, compared with the state-of-the-art approaches.

Diabetic retinopathy (DR), the leading cause of vision loss among diabetic adults worldwide, underscores the importance of early detection and timely treatment using fundus images to prevent vision loss. However, existing deep learning methods struggle to capture the correlation and contextual information of subtle lesion features with the current scale of dataset. To this end, we propose a novel Multi-scale Spatial-aware Transformer Network (MSTNet) for DR classification. MSTNet encodes information from image patches at varying scales as input features, constructing a dual-pathway backbone network comprised of two Transformer encoders of different sizes to extract both local details and global context from images. To fully leverage structural prior knowledge, we introduce a Spatial-aware Module (SAM) to capture spatial local information within the images. Furthermore, considering the differences between medical and natural images, specifically that regions of interest in medical images often lack distinct subjectivity and continuity, we employ a Multiple Instance Learning (MIL) strategy to aggregate features from diverse regions, thereby enhancing correlation to subtle lesion areas. Ultimately, a cross-fusion classifier integrates dual-pathway features to produce the final classification result. We evaluate MSTNet on four public DR datasets, including APTOS2019, RFMiD2020, Messidor, and IDRiD. Extensive experiments demonstrate that MSTNet exhibits superior diagnostic and grading accuracy, achieving improvements of up to 2.0% in terms of ACC and 1.2% in terms of F1 score, highlighting its effectiveness in accurately assessing fundus images.

Network traffic classification refers to the identification of collected network traffic data of various applications, which is widely used in research fields such as network resource allocation, traffic scheduling and intrusion detection systems. With the widespread application of encryption technology in the network, encrypted traffic classification has become a hot research topic. At present, most existing methods only focus on the accuracy of network traffic classification. Yet, few work studies the reliability of the classification model, which plays an important role in network regulation and network security. In this paper, we propose a novel traffic classification method based on trustworthy deep learning model, which can effectively improve the reliability of encrypted traffic classification models by correcting the confidence of model output. Specifically, we firstly perform data preprocessing on the original network traffic, and then adopt a ConvNet for feature learning and a ClassifyNet for traffic classification in the initial stage. At the same time, we utilize a trustworthy confidence criterion to design a ConfidNet trained according to the probability of the true class. The ConfidNet can provide a reliable confidence measure for the prediction of the classification model. Finally, we demonstrate the effectiveness of our framework through comprehensive experiments on two benchmark datasets ISCX VPN-nonVPN and USTC-TFC2016, and show that our method can improve the reliability of the classification model and has a good ability to identify misclassified samples compared with state-of-the-art methods.

Graph neural networks (GNNs) have demonstrated remarkable success for semisupervised node classification. However, these GNNs are still limited to the conventionally semisupervised framework and cannot fully leverage the potential value of large numbers of unlabeled samples. The pseudolabeling method in semisupervised learning (SSL) is widely recognized because it can clearly leverage unlabeled samples. Nevertheless, the existing pseudolabeling methods usually utilize a fixed threshold for all classes and only use a portion of unlabeled samples (ones with high prediction confidence), which leads to class imbalance and low data utilization. To solve these problems, we propose GNNs with adaptive confidence discrimination (ACDGNN) to fully utilize unlabeled samples for facilitating semisupervised node classification. Specifically, an adaptive confidence discrimination module is designed to divide all unlabeled nodes into two subsets by comparing their confidence scores with the adaptive confidence threshold at each training epoch. Then, different constraint strategies for two subset nodes are employed. Unlabeled nodes with high confidence are used to iteratively expand the label set, while ones with low confidence learn discriminative features by applying contrastive learning. Validated by extensive experiments, the proposed ACDGNN delivers significant accuracy gains over the previous SOTAs: an average improvement of 2.0% on all datasets and 5.7% on the Flickr dataset in particular.

Graph contrastive learning (GCL), as a typical self-supervised learning paradigm, has been able to achieve promising performance without labels and gradually attracts much attention. Graph-level method aims to learn representations of each graph by contrasting two augmented graphs. Previous studies usually simply apply contrastive learning to keep the embeddings of augmented views from the same anchor graph (positive pairs) close to each other, as well as separate the embeddings of augmented views from different anchor graphs (negative pairs). However, it is well-known that the structure of graph is always complex and multiscale, which gives rise to a fundamental question: after graph augmentation, will the previous assumption still hold in reality? Through experimental analytics, we find that the semantic information of two augmented graphs from the same anchor graph may be not consistent, and whether two augmented graphs are positive or negative sample pairs is highly correlated with the multiscale structure of the graph. Based on this observation, we then propose a multiscale subgraph contrastive learning method, named MSSGCL, which can characterize the fine-grained semantic information. Specifically, we generate global and local views at different scales based on subgraph sampling and construct multiple contrastive relationships according to their semantic associations to provide richer self-supervised information. Furthermore, to further improve the generalization performance of the model, we propose an extended model called MSSGCL++. It adopts an asymmetric structure to avoid pushing semantically similar negative samples far away. We further introduce adversarial training to perturb the augmented view and thus construct a more difficult self-supervised training task. Finally, a min-max saddle point problem is optimized and the “free” strategy is used to speed up the training process. Extensive experiments and parametric analysis on 16 real-world graph classification datasets confirm the effectiveness of our proposed approach. Compared with state of the art (SOTA) method, our method achieves improvements of 2% and 1.6% in unsupervised and transfer learning settings, respectively.

Fundus disease is a complex and universal disease involving a variety of pathologies. Its early diagnosis using fundus images can effectively prevent further diseases and provide targeted treatment plans for patients. Recent deep learning models for classification of this disease are gradually emerging as a critical research field, which is attracting widespread attention. However, in practice, most of the existing methods only focus on local visual cues of a single image, and ignore the underlying explicit interaction similarity between subjects and correlation information among pathologies in fundus diseases. In this paper, we propose a novel label-aware dual graph neural networks for multi-label fundus image classification that consists of population-based graph representation learning and pathology-based graph representation learning modules. Specifically, we first construct a population-based graph by integrating image features and non-image information to learn patient's representations by incorporating associations between subjects. Then, we represent pathologies as a sparse graph where its nodes are associated with pathology-based feature vectors and the edges correspond to probability of the co-occurrence of labels to generate a set of classifier scores by the propagation of multi-layer graph information. Finally, our model can adaptively recalibrate multi-label outputs. Detailed experiments and analysis of our results show the effectiveness of our method compared with state-of-the-art multi-label fundus image classification methods.

Glaucoma is a common ocular condition that can cause significant physiological distress and may lead to vision loss in advanced stages. In clinical practice, segmenting the optic cup and disc regions in fundus images provides key physiological indicators, such as the cup-to-disc ratio and optic nerve fiber layer thickness. These indicators are crucial for clinicians to make accurate diagnoses and informed treatment decisions. Therefore, this research is of significant clinical and societal relevance for the early detection and grading of glaucoma. This thesis introduces a novel network architecture, SMS-U-Net (SE Multi-resolution module Self-attention U-Net), based on the U-Net framework, for joint segmentation of the optic cup and disc in fundus images. The proposed approach scales features from input images and uses the SE (Squeeze-and-Excitation) module to adaptively extract features from each feature map, minimizing information loss. To preserve image details and capture features from multiple receptive fields before downsampling, the conventional convolutional sequence in the encoding path is replaced by a multi-resolution module employing 3 × 3, 5 × 5, and 7 × 7 convolutional kernels. Furthermore, a self-attention mechanism is integrated into the decoder to highlight critical features for segmentation. Experimental results show that the proposed SMS-U-Net method achieves an optic disc Dice coefficient of 0.985 and an IoU of 0.968, as well as an optic cup Dice coefficient of 0.919 and an IoU of 0.851, demonstrating its effectiveness.

BACKGROUND Gastrointestinal stromal tumors (GIST) are prevalent neoplasm originating from the gastrointestinal mesenchyme. Approximately 50% of GIST patients experience tumor recurrence within 5 years. Thus, there is a pressing need to accurately evaluate risk stratification preoperatively. AIM To assess the application of a deep learning model (DLM) combined with computed tomography features for predicting risk stratification of GISTs. METHODS Preoperative contrast-enhanced computed tomography (CECT) images of 551 GIST patients were retrospectively analyzed. All image features were independently analyzed by two radiologists. Quantitative parameters were statistically analyzed to identify significant predictors of high-risk malignancy. Patients were randomly assigned to the training (n = 386) and validation cohorts (n = 165). A DLM and a combined DLM were established for predicting the GIST risk stratification using convolutional neural network and subsequently evaluated in the validation cohort. RESULTS Among the analyzed CECT image features, tumor size, ulceration, and enlarged feeding vessels were identified as significant risk predictors (P < 0.05). In DLM, the overall area under the receiver operating characteristic curve (AUROC) was 0.88, with the accuracy (ACC) and AUROCs for each stratification being 87% and 0.96 for low-risk, 79% and 0.74 for intermediate-risk, and 84% and 0.90 for high-risk, respectively. The overall ACC and AUROC were 84% and 0.94 in the combined model. The ACC and AUROCs for each risk stratification were 92% and 0.97 for low-risk, 87% and 0.83 for intermediate-risk, and 90% and 0.96 for high-risk, respectively. Differences in AUROCs for each risk stratification between the two models were significant (P < 0.05). CONCLUSION A combined DLM with satisfactory performance for preoperatively predicting GIST stratifications was developed using routine computed tomography data, demonstrating superiority compared to DLM.

Gastrointestinal stromal tumor (GIST) is the most common mesenchymal tumor of the gastrointestinal tract. GIST risk classification based on CT images is of great clinical significance. However, there are still some issues with the current methods in predicting the risk level of gastrointestinal stromal tumors. Existing 2D CNN methods using only a single CT image cannot fully characterize tumors; 3D CNN methods require a large amount of labeled samples to achieve better performance and have significant computational requirements. To this end, we propose a novel Inter-Slice Attention-based Vision Transformer (ISA-ViT) method that automatically integrates the sequential information of CT images to learn the three-dimensional features of tumors for GIST risk classification. Specifically, we incorporate inter-slice attention mechanism into the Vision Transformer to learn the sequential information among CT slices, so that three-dimensional features can be learnt by our model. The extracted CT image information with clinical information of patients are then fused for GIST risk classification. The experimental results show that our model reaches 84% in terms of classification Accuracy (ACC) and 0.94 in terms of Area Under Curve (AUC). Compared with other baseline models, the proposed model achieves state-of-the-art performance results.

Gastrointestinal stromal tumor (GIST) is a common mesenchymal tumor in the gastrointestinal system. Its treatment outcomes and prognosis are closely related to mutations in the KIT11 gene locus. Therefore, accurately predicting the mutation status of the KIT11 gene locus is crucial for developing personalized treatment plans and improving patient prognosis. We propose a multimodal prediction model with co-attention fusion. This method accurately extracts features from clinical and CT image data using a two-stream interaction module to facilitate information interaction and fusion between modalities, which can effectively predict the KIT11 gene mutation status. Experimental results show that the proposed multimodal model with co-attention fusion achieves an accuracy of 0.840 and an AUC value of 0.862 in predicting KIT11 gene mutations in gastrointestinal stromal tumors.