Retrieving visual content from electroencephalography (EEG) signals represents a challenging frontier in implicit multimedia analysis, yet it suffers from extreme modal heterogeneity. The high-dimensional, non-stationary noise in neural signals limits conventional point-to-point similarity measures in capturing complex semantic manifolds. To bridge this gap, we propose the Neuro-Image Geometric Contrastive Learning (NIGCL) framework. Departing from reliance solely on simple first-order similarity metrics, NIGCL employs a geometry-aware alignment mechanism rooted in manifold learning. Specifically, we incorporate a Geometric Area Contrastive Loss based on the Gram matrix determinant, which constrains the geometric area of cross-modal feature pairs to enforce intra-class compactness and mitigate the impact of orthogonal perturbations. This global constraint is complemented by a local dot-product objective for fine-grained consistency. Additionally, we propose Geometric Area Ranking (GaR) to replace standard ranking protocols, identifying semantically consistent images via the geometric area in high-dimensional spaces. Experiments on the THINGS-EEG dataset show that NIGCL achieves superior performance, attaining 31.2% Top-1 and 61.6% Top-5 accuracy in 200-way retrieval. Reconstruction evaluations further confirm that our geometrically-aligned representations significantly improve semantic fidelity over traditional methods. This framework offers a novel perspective on aligning highly heterogeneous multimedia data through explicit geometric constraints.

Major Depressive Disorder (MDD) is a common mental disorder, and making an early and accurate diagnosis is crucial for effective treatment. Functional Connectivity Network (FCN) constructed based on functional Magnetic Resonance Imaging (fMRI) have demonstrated the potential to reveal the mechanisms underlying brain abnormalities. Deep learning has been widely employed to extract features from FCN, but existing methods typically operate directly on the network, failing to fully exploit their deep information. Although graph coarsening techniques offer certain advantages in extracting the brain's complex structure, they may also result in the loss of critical information. To address this issue, we propose the Multi-Granularity Brain Networks Fusion (MGBNF) framework. MGBNF models brain networks through multi-granularity analysis and constructs combinatorial modules to enhance feature extraction. Finally, the Constrained Attention Pooling (CAP) mechanism is employed to achieve the effective integration of multi-channel features. In the feature extraction stage, the parameter sharing mechanism is introduced and applied to multiple channels to capture similar connectivity patterns between different channels while reducing the number of parameters. We validate the effectiveness of the MGBNF model on multiple classification tasks and various brain atlases. The results demonstrate that MGBNF outperforms baseline models in terms of classification performance. Ablation experiments further validate its effectiveness. In addition, we conducted a thorough analysis of the variability of different subtypes of MDD by multiple classification tasks, and the results support further clinical applications.

Emotion recognition from multimodal data remains challenging due to the semantic gap and temporal-spatial misalignment between EEG signals and facial expressions. To address this, we propose a cross-modal framework that integrates EEG and facial features via modality-guided semantic representation learning. Temporal features are extracted by stacked MAMBA-based blocks capturing long-range dependencies. A Cross-Modal Scaling and Shifting (CMSS) mechanism uses EEG features to refine and align facial representations, reducing modality discrepancies. The fused features pass through a GRUcontrolled Mixture-of-Experts (MoE-GRU) module, where a learnable gating network dynamically selects specialized Transformer experts. This combination effectively handles misalignment and enables dynamic feature fusion, enhancing recognition accuracy and robustness. Experiments on DEAP and MAHNOBHCI demonstrate state-of-the-art results, validating its real-world applicability in affective computing.

Major depressive disorder (MDD) is a common neuropsychiatric disorder, yet its underlying physiological mechanisms remain unclear, limiting diagnostic advances. Functional connectivity (FC) derived from resting-state functional magnetic resonance imaging (rs-fMRI), when combined with deep learning methods, has shown promise as diagnostic biomarker. Currently, most FC-based diagnostic methods rely on graph structures modeled by FC, and are limited to pairwise interactions between brain regions. Hypergraph representations enable the characterization of higher-order interactions across multiple regions. However, existing hypergraph models ignore the directionality of these interactions, thus limiting their ability to capture complex neural dynamics. To address these limitations, this study proposes a dynamic weight aware directed hypergraph learning method - dwDHGL, for MDD identification and subtype analysis. dwDHGL captures asymmetric causal interactions by modeling temporal lag effects and constructs a directed hypergraph network(DHN). It further utilizes a self-attention mechanism to dynamically learn inter node interactions during message passing and adaptively differentiate node importance. A node weight aware directed hypergraph convolution is designed to aggregate features based on hyperedge directions, incorporating dynamic weights to enhance representation learning. The proposed method is evaluated on the large-scale REST-meta-MDD dataset, achieving an MDD identification accuracy of 73.75 %, and outperforming existing advanced methods in subtype identification. Furthermore, dwDHGL identifies discriminative directed hyperedges, with the inferior frontal gyrus triangular part emerging as key biomarkers, providing new insights into the neural mechanisms of MDD.

Magnetic resonance imaging (MRI) is of great value in clinical diagnosis due to its ability to present tissue structure and functional information of the brain. However, the acquisition of high-resolution MRI images remains challenging due to constraints in scanning time and hardware limitations. In response, multi-contrast super-resolution (SR) reconstruction has emerged as a promising technique for enhancing image quality. The effectiveness of this approach largely depends on the ability to fully leverage the complementary information across different modalities and to achieve accurate structure matching. To address this challenge, we propose a Mamba-enhanced large-window Transformer network (MC-MambaTrans), which effectively improves the reconstruction accuracy through multi-modal deep feature extraction and structure-guided matching. Specifically, MC-MambaTrans employs the large-window Transformer to model cross-modal multiscale global contextual information, and at the same time introduces the Mamba mechanism-driven coarse-to-fine matching strategy to enhance the guidance of structural information from the reference image slice-by-slice. Ultimately, high-quality SR images are recovered by the multi-scale feature fusion and up-sampling module. Experiments on several publicly available multi-contrast brain MRI datasets show that the method in this paper significantly outperforms the existing state-of-theart methods in terms of reconstruction quality, demonstrating its broad application prospects in medical image reconstruction tasks.

To enhance the efficiency of crowd evacuation and inform collaborative design strategies, it is essential to investigate the effects of exit features on human exit choices. This study explores how exit distance, crowd density near exits, and exit location settings influence individual exit selection. We conducted a virtual reality experiment, revealing that all targeted features significantly impact exit choices, with density exerting the most substantial influence, followed by distance and exit location being the least impactful. Additionally, the interaction between distance and location significantly affected exit decisions. By integrating our findings into machine learning models, we demonstrate the potential of these exit features for informing collaborative evacuation strategies and designing systems that support effective decision-making in crowd dynamics. This research contributes to understanding human behavior in evacuation scenarios, emphasizing the importance of collaborative approaches in optimizing crowd management.

To tackle the challenge of effectively representing time-varying information in cross-subject EEG emotion recognition, we introduce Swin Transformer-Based Temporal-Channel Network (Swin-TCNet), a novel multi-scale neural network with parallel temporal pathways to enhance the extraction of dynamic temporal features. EEG signals are simultaneously processed through a Temporal Swin Transformer for 3D feature representation and a dynamic spatiotemporal convolutional layer with multi-head attention for extracting differential entropy-based channel features, which enhances channel learning while preserving temporal information. Swin-TCNet attains state-of-the-art performance, achieving 93.47% and 86.80% accuracy in cross-subject experiments on SEED and SEED-IV datasets, respectively, as substantiated by ablation studies. By leveraging temporal dynamics, this framework enhances the extraction of temporal variations and spatial information, leading to more robust and generalizable cross-subject emotion recognition.

Molecular property prediction is a fundamental yet crucial task. It relies on molecular representation, which involves transforming molecular structures and features into a form that can be processed by computers. Common representation methods can be divided into two perspectives: global and local. However, using molecular representations from a single perspective leads to the problem of models focusing excessively on certain features while neglecting other important information, which limits the model’s generalization ability and accuracy. To address this issue, this paper proposes a Multi-Representation Local-Global Molecular Property Prediction Model (MRLG). This model adopts a multi-branch architecture, deeply integrating SMILES, molecular fingerprints, molecular graphs, and molecular substructure information. First, a Global-Local Fusion (GLF) module is designed, which can integrate multiple representations and generate new, more comprehensive representations. Second, a detailed feature extraction module, Double-Cross Convolution Mould(DCC), is designed for the generated representations. Experiments conducted on various real-world datasets fully validate the effectiveness of the MRLG model. Moreover, results from branch and module ablation experiments further confirm the effectiveness of the proposed method. Overall, our model demonstrates a promising ability to accurately predict molecular properties, offering valuable insights for the design and optimization of novel compounds in various fields of material science and drug development.

Remote sensing image interpretation is a crucial process for Earth observation and geoscience research, and multi-modal fusion techniques are key to improving semantic segmentation accuracy. CNN-based methods, due to the limited receptive field, suffer from inadequate long-range contextual modeling. While Transformer-based architectures alleviate the limitation, existing multi-modal fusion Transformers still face challenges in efficiently integrating multi-scale information. Therefore, this work proposes a multi-scale dual-modal fusion Transformer, namely MDFformer, to integrate RGB and DSM cross-modal interaction information. More specifically, MDFformer is a dual-stream encoder architecture combined with a multi-scale fusion FPN decoder. MDFformer introduce two novel modules: CSFM (Cross-Modality Shuffle Fusion Module) and SDEM (Spatial Detail Enhance Module), along with deep supervision(DS) to alter the gradient flow and enhance multi-scale feature learning. As a result, MDFformer efficiently utilizes both shallow and deep features from diverse modalities across multiple scales. The validation is performed on two publicly available high-resolution datasets, Vaihingen and Potsdam, achieving an overall accuracy (OA) of 92.06% and 91.25%, respectively. Therefore, the model effectively handles RGB and DSM data, making it suitable for remote sensing tasks and efficient earth sensing.

Integrating multicenter resting-state functional magnetic resonance imaging (fMRI) datasets is essential for computer-aided diagnosis of brain diseases. However, differences in scanners and acquisition protocols lead to heterogeneity of data from different centers. To address this problem, we introduce a novel Centralized Limited Data Sharing (CLDS) framework for multicenter fMRI-based classification, which contains a centralized server and multiple local models, and propose three mechanisms to improve the classification performance. First, aiming at coordinating the global data without favoring any center, an adversarial network is introduced in the centralized server with limited noise-added data uploaded from each center. Second, a Gradient Adaptive Strategy based on the Tasks Association (GASTA) is proposed to restrict the gradient to a desirable direction. Third, a Correction Factor based on the Cosine Similarity (CFCS) is proposed to reduce the impact of non-independent and identically distributed (non-IID) of data on models. The experimental results show that CLDS exhibits superior classification performance to other related methods on the ADHD-200 dataset, with an accuracy of 70.4%. We also extend CLDS to the ABIDE-I dataset with an accuracy of 73.4%, demonstrating that CLDS has the potential for generalizability of other brain diseases and more centers.

To decode Motor Imagery EEG signals (MI-EEG), most studies have increasingly complicated network models and parameters without fully considering EEG characteristics, thereby limiting advancements in Brain-Computer Interface (BCI) systems and classification performance. To address these issues, we propose a lightweight end-to-end tri-domain feature fusion network named LTDFNet. Firstly, we introduce an Attention-based Spatio-temporal Convolution module (ABST) to extract low-dimensional spatio-temporal features from EEG. This module employs a lightweight Squeeze-and-Excitation (SE) attention mechanism to enhance the model's perception of crucial information. Secondly, Temporal Domain Convolutional (TDC) and Frequency Domain Convolution (SDC) modules utilize Temporal Convolutional Networks (TCN) and Fast Fourier Transform (FFT) to respectively learn high-dimensional temporal and frequency domain information. Finally, the Feature Fusion (FF) module integrates low-dimensional spatio-temporal features and high-dimensional temporal-frequency features effectively through learnable parameters. LTDFNet is trained with joint constraints of Softmax loss and Center loss functions to achieve optimal inter-class separation and intra-class compactness, thereby enhancing overall model performance. This study conducts extensive experimental validation on BCI Competition IV datasets 2a (BCI 2a) and 2b (BCI 2b). LTDFNet achieves classification accuracies of 76.89% (kappa score: 0.692) and 85.22% (kappa score: 0.704) on the BCI 2a and BCI 2b datasets, respectively. Compared to other high-performance decoding methods, LTDFNet utilizes only 20,576 parameters, balancing network scale and decoding performance requirements.

Functional connectivity (FC) derived from fMRI reflects the interactions between brain regions of interest (ROIs). It has become one of the important features of individualized prediction. However, in studies using FC as input, some studies have extracted global features directly from the entire brain FC, lacking focus on local critical information. On the other hand, some studies have extracted local critical features from selected ROIs or connections, but lack access to global contextual information. In this paper, we propose a novel method, namely, Global-Local Brain Network (GLBN), focusing on both global contextual information and critical local information. We validate our proposed method on the largescale public dataset, the Cambridge Centre for Ageing and Neuroscience (Cam-CAN). For the prediction of age and fluid intelligence, GLBN achieves the mean absolute errors of 6.084, and 4.160, with Pearson’s correlations of 0.908, and 0.641, respectively. Our method demonstrates superior prediction accuracy compared to existing studies. Additionally, we visualize the brain ROIs that play crucial roles in the prediction tasks, affirming the biological interpretability of GLBN.

The prediction of drug-target interaction (DTI) is a momentous task in modern medicine. Using efficient computational methods to predict the relationship between drugs and targets can effectively shorten the drug development cycle, reduce the blindness of new drug development, and provide theoretical guidance for biochemical experiments. It also plays a crucial role in drug repurposing, optimizing drug treatment plans, and drug resistance. Due to the vast number of molecules and proteins in the real world, individually screening them is impractical and costly. Therefore, utilizing deep learning to learn the features of proteins and molecules and enable the prediction of drug-target interaction (DTI) has gradually become mainstream. However, most studies extract features using the entire molecular structure and target, which introduces a significant amount of noise and overlooks the complex interactions within crucial components. To address these issues, we propose an end-to-end prediction model called BN-DTI for drug-target interaction. In this model, we extract numerous fragments from the amino acid sequences of drugs and proteins. Then, we employ CNN for feature extraction and utilize a multi-head cross-attention mechanism to learn the interactions between drugs and targets. Finally, the input is fed into the prediction module for prediction. Our model achieved significant results on the Human dataset, C.elegans dataset, and Davis dataset, with AUC values of 0.988, 0.992, and 0.911, respectively. The results show that BN-DTI requires lower training costs and provides new ideas for drug discovery and other fields.