IntroductionThyroid ultrasound is the primary imaging modality for nodule detection, but manual interpretation suffers from subjectivity and inefficiency due to speckle noise, low contrast, and operator dependence. Deep learning-based segmentation methods often overlook anatomical prior information, leading to suboptimal performance on atypical nodules and complex backgrounds.MethodsWe propose RTS-Net, a novel segmentation network that integrates a dual-path attention enhancement mechanism (combining spatial and channel attention) and a cascaded graph convolution decoding architecture to leverage multi-scale feature pyramid fusion. A deep supervision strategy is also employed to accelerate convergence. The model is trained and evaluated on the TN3K, DDTI, and a large-scale clinical dataset.ResultsExtensive experiments demonstrate that RTS-Net achieves superior performance on both in-distribution and cross-dataset settings. On the TN3K dataset, it attains 81.66% F1-score and 71.87% IoU; on the DDTI dataset, it achieves 71.10% F1-score and 60.09% IoU, outperforming state-of-the-art methods including UNet, DeepLabv3+, TransUNet, and recent foundation-model-based approaches. Ablation studies confirm the effectiveness of each proposed component.DiscussionThe proposed dual-path attention and graph convolution modules effectively enhance feature representation and boundary integrity, particularly for small nodules and blurred edges. While RTS-Net shows strong generalization, failure cases reveal challenges in heterogeneous backgrounds and acoustic artifacts, suggesting future integration with foundation models like SAM to further improve robustness.

The Internet of Medical Things (IoMT) facilitates the large-scale sharing of medical images and enhances diagnostic efficiency, yet it also introduces significant security and privacy concerns. However, existing schemes often suffer from limitations such as low batch processing efficiency and insufficient interaction between encryption and authentication. This paper presents a multi-layer security framework for the transmission of bulk medical images, leveraging two-dimensional hyperchaotic maps. To address the limited hyperchaotic behavior exhibited by conventional two-dimensional chaotic maps, we introduce a two-dimensional exponential accumulation chaotic model that effectively balances complex dynamics with computational efficiency. Evaluations based on phase diagrams, Lyapunov exponents, permutation entropy, and sample entropy confirm that the proposed model strengthens chaotic characteristics and improves ergodicity. Building upon this model, we develop a fast ciphertext feedback synchronous scrambling-diffusion algorithm utilizing the maximum effective region, which enhances security while maintaining high computational speed. Additionally, a zero-watermarking algorithm based on dual-domain feature fusion is designed to enable reliable ownership verification without modifying the original medical image content. Experimental results demonstrate that the proposed approach effectively eliminates pixel redundancy, achieves higher encryption and authentication efficiency, offers a larger key space, and exhibits strong resistance against statistical, differential, cropping, and noise attacks.

Functional brain network analysis plays an important role in understanding and diagnosing psychiatric disorders. However, current methods struggle with subject variations, impairing the model’s generalization ability to the test set. To address this issue, we propose the Subject Invariance-aware Inverse Graph Contrastive Learning (SI-IGCL) model, which adopts a two-stage paradigm with self-supervised subject-invariant pre-training followed by supervised fine-tuning for identification. During the pre-training phase, we construct an inverse contrastive objective that reshapes the embedding space by repelling intra-subject and attracting inter-subject embeddings to learn subject-invariant representations, with an auxiliary correction term to avoid early optimization plateaus. Meanwhile, we incorporate a structure-preserving reconstruction constraint to preserve discriminative information. Moreover, a Hierarchical Topology Enhanced Transformer (HTET) module is designed to enable multi-level modeling of subject-invariant functional patterns. During the fine-tuning phase, a supervised classifier is integrated to perform psychiatric disorder classification. Extensive experiments demonstrate that our method outperforms all state-of-the-art methods. The code is available at https://anonymous.4open.science/r/SI-IGCL.

Accurate preoperative assessment of muscle invasion in bladder cancer (BCa) guides therapy selection. However, MRI interpretation varies across readers and lesion morphologies. Therefore, we aimed to overcome the morphology-associated diagnostic bias through a deep learning method. This multicenter study included 1374 patients with BCa. An nnU-Net was fine-tuned to assist in lesion segmentation on T2-weighted images, providing inputs for a 2.5D ConvNeXt-tiny model to assess muscle invasion. The performance of the model was compared between pedunculated and sessile lesions. Furthermore, a head-to-head comparison was conducted among the model, a senior radiologist, and a junior radiologist. The validation Dice coefficient of nnU-net was 0.834. In the validation and three prospective test sets, the ConvNeXt-tiny model achieved areas under the receiver-operating characteristic curve of 0.915–0.925 for identifying muscle invasion in BCa, with accuracies of 84.9–91.0

Generative Adversarial Networks, a popular deep learning method, have achieved excellent performance in both classification and prediction tasks. However, there have been relatively few applications of generative adversarial networks to EEG data. To study the effect of high-order brain functional networks on schizophrenia patients, a high-order graph attention generative adversarial network prediction model is proposed, and the generator of the model utilizes graph attention networks and long short-term memory networks to capture the high-order topological features of persistence images for early diagnosis and prediction of schizophrenia patients. The research results on the five frequency bands of schizophrenia show that the proposed prediction model performs best in the Theta frequency band, with AUC and MAP values reaching 93.5% and 93.0%, respectively, and an average accuracy of 91.5%, both of which are superior to the selected comparison methods. Moreover, the image quality coefficient is used to quantify the realism and clarity of the images generated by the model. the image quality coefficients of schizophrenia patients were significantly correlated with the PANSS total scores in the Gamma and Theta bands, which provided a new idea for generative adversarial networks in the prediction of schizophrenia high-order topological features.

Functional connectome (FC) fingerprinting is crucial for understanding individual cognitive patterns and advancing personalized medicine for neuro/psychiatric disorders by developing individual-specific biomarkers. However, existing FC fingerprinting methods oversimplify the complex and nonlinear nature of FC patterns. As a result, they fail to effectively extract individual-specific information from variability across different brain states, thereby limiting individual identification performance. To address this issue, we propose Hierarchical Attention and Hard Negatives-aware State Graph Contrastive Learning (HAHN-SGCL) model. HAHN-SGCL directly leverages brain states to generate intra- and inter-individual contrasts, effectively extracting individual-specific connectivity patterns for accurate identification across diverse states. Specifically, to fully extract individual-specific information across multiple topological levels of the FC, we designed a Hierarchical Graph Attention Network (HGAT) encoder. HGAT constructs a hierarchical graph with diverse topological perspectives and employs level-specific attention mechanisms to capture distinctive individual features. Additionally, to overcome the severe sample imbalance that hampers effective gradient propagation, we introduce a Hard Negatives-aware Strategy (HNS). HNS focuses on challenging negatives through Hard Negative Mining (HNM) and incorporating a corrective term, effectively avoiding early convergence plateaus. Extensive experiments demonstrate that our HAHN-SGCL model outperforms state-of-the-art methods. It also exhibits strong cross-task transferability, as evidenced by its robust performance in psychiatric disorder classification. The code of HAHN-SGCL is at https://anonymous.4open.science/r/HAHN-SGCL.

BACKGROUND:While direct brain connection differences are well-documented in Autism Spectrum Disorder (ASD), brain function also critically depends on indirect, higher-order structural connections (hSC). Understanding the influence of these connections on functional relationships is key to unraveling the neural mechanisms of ASD. METHODS:This study included 76 participants with ASD and 64 typically development (TD). We utilized the node2vec embedding method to characterize brain nodes and construct hSC networks, which were then differentiated into direct and indirect connectivity networks. The structural-functional coupling (SC-FC coupling) method was subsequently employed to quantify the relationship between structural and functional connectivity. RESULTS:Our findings demonstrated significant differences in SC-FC coupling within the ASD group. These alterations were particularly evident in rich club connections and within specific modules, including the default mode network and visual network. Furthermore, the coupling strengths of several brain regions-specifically the left dorsolateral superior frontal gyrus, left middle orbital gyrus, left olfactory cortex, and right superior temporal gyrus-were found to be associated with the severity of ASD symptoms. CONCLUSION:These results underscore the importance of considering higher-order network interactions when analyzing structural and functional relationships in neurodevelopmental disorders. This study offers new insights into the neural mechanisms of ASD by highlighting the role of hSC and its coupling with functional connectivity.

Drug-target interaction (DTI) prediction plays a pivotal role in accelerating drug discovery. Nevertheless, existing AI-driven approaches face three critical limitations: existing attention-based methods lack dynamic bidirectional interaction channels, limiting their ability to model asymmetric drug-target communication patterns; conventional architectures struggle to integrate both local binding patterns and global biological con texts; and rigid prediction heads discarding spatial interaction patterns. These issues hamper the accuracy and comprehensive ness of DTI prediction. To address these, we propose CAMA DTI, an end-to-end framework integrating three innovations. First, a cross-domain bidirectional attention module establishes dual-perspective interaction channels that enable co-evolutionary feature refinement through mutual pharmacological feedback. Second, the Mamba-driven fusion block incorporates state-space modeling to dynamically integrate local binding patterns with global biological contexts across extended sequences. Third, we replace conventional classifiers with Kolmogorov-Arnold Net works (KANs) employing adaptive spline transformations that preserve multi-scale interaction signatures while maintaining parametric efficiency. Extensive experimental results demonstrate that CAMA-DTI achieves robust and accurate predictions across diverse datasets, outperforming state-of-the-art methods in both established and novel drug target scenarios. Notably, the frame work maintains consistent performance across datasets of varying scales, and case studies validate its practical utility in real-world drug development pipelines.

Major Depressive Disorder (MDD) is a prevalent and disabling psychiatric disorder. Its clinical heterogeneity undermines diagnostic precision and treatment effectiveness. While DSM-5 defines melancholic and anxious subtypes based on symptoms, their neurophysiological bases remain unclear. Resting-state electroencephalography (EEG) provides a non-invasive approach for investigating neurophysiological heterogeneity in depression. Beta-band oscillations are associated with reward processing, motivation, and cortical arousal, which may differ across subtypes. We analyzed 325 patients with MDD, classified into melancholic (n = 108), anxious (n = 102), and non-melancholic and non-anxious (n = 115) groups. Relative spectral power and phase-locking value (PLV) in beta1 (12–20 Hz) and beta2 (20–30 Hz) bands were compared. Logistic regression and subtype-specific correlation analyses assessed classification and clinical associations. Melancholic depression was associated with widespread reductions in beta2 power and higher suicide risk, whereas anxious depression showed altered beta-band functional connectivity, particularly involving left temporal-region connections. Both subtypes shared reduced prefrontal–temporal connectivity. Exploratory classification analyses showed modest discriminative performance after leakage-controlled cross-validation. Associations between beta1 activity and symptom severity were observed only in the anxious subtype. Melancholic and anxious depression showed partially distinct beta-band EEG patterns. These findings provide preliminary evidence that beta-band EEG features may capture subtype-related neurophysiological heterogeneity in MDD, with subtype-specific associations between neural activity and symptom severity. Further validation is required before these features can be considered clinically applicable. The trial was registered at the Chinese Clinical Trial Registry on 04/23/2022 ( www.chictr.org.cn ChiCTRID ChiCTR2200059053).

OBJECTIVE:Motor imagery EEG (MI-EEG) decoding remains challenging due to low signal-to-noise ratios and pronounced inter-subject variability. Although end-to-end deep models reduce reliance on manual feature engineering, many existing architectures may introduce temporal leakage through non-causal operations and often rely on fixed spatial topologies that cannot accommodate subject- and trial-specific connectivity patterns. APPROACH:We propose MAGCANet, which integrates five core components: (i) a Multiscale Causal Convolution Module (MCCM) for hierarchical temporal encoding under explicit causal constraints, (ii) a Temporal Convolution Module (TCM) to capture complex temporal dynamics, (iii) an Adaptive Graph Convolution Module (AGCM) for sample-specific topology learning in latent space, (iv) a Multi-Head Self-Attention Module (MHSAM) for global feature aggregation, and (v) a Classification Block for final decision making. Together, these components enforce temporal causality, adapt spatial interactions to individual dynamics, and produce discriminative representations robust to inter-subject variability. RESULTS:On the BCI Competition IV-2a and IV-2b datasets, MAGCANet achieves strong single-subject accuracies of 88.58% and 91.13%, respectively. Under Leave-One-Subject-Out (LOSO) evaluation, the model maintains accuracies of 70.49% and 79.49%, demonstrating competitive and stable cross-subject generalization. MAGCANet is highly lightweight, with only 0.0194M parameters, and achieves low inference latency (2.23 ms). Qualitative analyses, including feature clustering and channel occlusion, further highlight the model's interpretability and its ability to capture relevant EEG patterns. SIGNIFICANCE:MAGCANet provides a robust and interpretable solution for MI-EEG decoding, balancing high precision with computational efficiency, and offering a reliable method for real-time BCI applications.

Background3D medical image segmentation is a cornerstone for quantitative analysis and clinical decision-making in various modalities. However, acquiring high-quality voxel-level annotations is both time-consuming and labor-intensive. Semi-supervised learning (SSL) provides an appealing solution by effectively utilizing limited labeled data along with abundant unlabeled data to enhance segmentation performance under clinical data constraints.MethodsWe propose a foundation model-driven multi-view collaborative learning framework that exploits zero-shot capabilities of SAM-like foundation models to jointly learn from axial, sagittal, and coronal planes. A collaborative fusion module integrates complementary representations across views, enhancing 3D structural understanding and improving the performance with limited annotation cost.ResultsExtensive experiments on two evaluation datasets including MRI brain tumor segmentation and whole-body PET heart segmentation demonstrate that our proposed method consistently outperforms existing SAM-based semi-supervised approaches. The multi-view collaborative design not only refines boundary precision for organ and tumor delineation but also shows strong transferability across imaging modalities.ConclusionThis study presents a foundation model-driven, multi-view collaborative learning paradigm that efficiently advances semi-supervised 3D medical image segmentation, which provides a scalable and clinically meaningful solution that reduces annotation dependency while maintaining high segmentation accuracy across diverse medical imaging modalities.

Functional magnetic resonance imaging (fMRI) allows the observation of brain functional connectivity patterns. Attention-based diagnostic models have been widely applied in fMRI data for brain disease diagnosis. However, the global attention mechanism of the Transformer faces challenges in adaptively identifying and focusing on significant brain regions and connections relevant to disease diagnosis while reducing attention to non-relevant regions and connections in fMRI data, as well as the degradation problem of the attention mechanism, thereby limiting the improvement in diagnostic accuracy. To address these problems, we propose a connection-mask-residual focused attention network (Trifocal Transformer) based on fMRI data for brain disease diagnosis. In the Trifocal Transformer, a Connection Focus Module is developed to simulate brain functional connectivity, thereby enhancing the attention mechanism's ability to focus on significant regions and connections relevant to disease diagnosis. To mitigate the potential negative impact of non-focused regions in the attention map, a learnable Mask Focus Module is designed to adaptively reduce attention to non-relevant regions and connections. To address the degradation of the attention mechanism's focusing ability, we establish Residual Focus Connections between the attention maps, which reinforce the focusing effect across layers and ensure stable attention to significant features. Comprehensive experimental results demonstrate that the Trifocal Transformer achieves superior diagnostic accuracies of 74.1% and 71.2% on ADHD-200 and ABIDE I datasets, respectively. Furthermore, our method reveals potentially disease-related regions of interest (ROIs), providing a new neuroimaging perspective for brain disease diagnosis and treatment.

Background: Epilepsy poses ongoing physical and mental threats and causes substantial economic burdens. Better seizure forecasting enables faster medical responses, improving patients’ quality of life and lowering healthcare costs. Research mainly focuses on early forecasting within a short preictal window, often too brief for effective drug administration. A major challenge is that a longer preictal phase may resemble the interictal state, making differentiation difficult. New methods: We propose a causal attention network (CANet) with a longer interictal and preictal of 1 h and 2 h respectively as the research object. In the feature extraction, a dilated causal convolution network is employed to extract local features. Causal attention is innovatively incorporated into epilepsy prediction to capture global correlation features. The complementary integration of these two methods enhances feature extraction and enables a more precise distinction between interictal and preictal periods. A double-layer dynamic window algorithm is developed for seizure prediction. Results: We evaluate the performance on Freiburg and CHB-MIT datasets. On the Freiburg dataset, the sensitivity(Sen) of the 1/2-hour preictal intervals was 100.00%/96.67%, with a false alarm rate per hour (FAR) of 0.0077/h/0.0472/h, and the average prediction time (APT) was 97.59 min. On the CHB-MIT dataset, we achieved Sen of 97.06%/92.31%, FAR of 0.0251/h/0.0666/h, and APT of 94.85 min, under the same conditions. Comparison with existing methods and conclusion: Our approach outperforms most of the previous methods, and the intracranial EEG (Freiburg) can more effectively distinguish interictal and preictal periods than scalp EEG (CHB-MIT).

Single-cell RNA sequencing (scRNA-seq) provides an unprecedented opportunity to dissect cellular heterogeneity and investigate disease mechanisms at single-cell resolution. However, accurate cell-type annotation remains challenging due to strong batch effects across studies, high-dimensional and sparse gene expression profiles, and the need to reliably detect rare cell populations. To address these issues, we propose scNSATrans, a Transformer-based deep learning model using Native Sparse Attention (NSA) and Kolmogorov-Arnold network (KAN). Firstly, fine-grained gene embedding is used for efficient gene representation of scRNA-seq data, and then a Transformer based on the native sparse attention mechanism is used to capture complex gene interactions. NSA combines compressed, selected, and sliding-window attention branches to simultaneously capture global co-expression patterns, rare-cell-specific marker signals, and local gene dependencies, while maintaining computational efficiency. Finally, the KAN network is used as a cell classifier for accurate cell type annotation. KAN employs learnable activation functions to model complex nonlinear mappings between transcriptomic signatures and cell phenotypes, enhancing classification flexibility. Through comprehensive evaluation on 11 publicly available datasets, we have demonstrated that scNSATrans exhibits superior performance in rare cell identification, intra- and cross-dataset annotation.

To mitigate the impact of negative transfer when extracting valid information from multiple censored source datasets, we propose two transfer-learning-based methods for estimating high-dimensional Cox proportional hazards models. We first develop a fused-regularized transfer learning method tailored to high-dimensional Cox models, enabling efficient information transfer from source domains to the target estimation task. By quantifying and dominating discrepancies between the target and source datasets, we effectively mitigate the risk of negative transfer. To further circumvent negative transfer impact, we make several improvements and propose a post-detection transfer learning method for high-dimensional Cox model estimation. Prior to information aggregation, transferable source datasets are detected via a dedicated detection procedure to circumvent negative transfer. We then integrate these transferable sources to enhance the estimation efficiency of target parameters. We also established the consistency of transferability detection procedure in post-detection transfer learning method, and derive the estimation error bounds for both transfer learning estimators. Simulation results demonstrate that the two proposed methods exhibit superior performance compared to conventional non-transfer-learning approaches. Their practical utility is further validated by analyzing clinical data from The Cancer Genome Atlas (TCGA) study, confirming their effectiveness in real-world clinical research scenarios.

Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder, and accurate diagnosis is critical for ensuring timely intervention. The integration of deep learning and fMRI can effectively explore the abnormal spatiotemporal features of ADHD. However, existing deep learning models are unable to fully capture the temporal dependence and spatial consistency of fMRI data, primarily due to insufficient modeling of multi-scale temporal dependencies and the lack of explicit interaction between static and dynamic brain networks, resulting in poor diagnostic performance for ADHD. To comprehensively and efficiently extract the spatiotemporal features of fMRI signals, we propose an Attention-Enhanced Spatiotemporal Feature Extraction Network (AE-STEN), which comprises a Temporal Cross-scale Convolutional Attention Module (TCAM), a Spatial Collaborative Attention-Guided Graph Representation Module (SCGRM), and a Spatial-Temporal KAN Network (STKAN). TCAM is designed to capture short- and long-term dependencies in fMRI time series by jointly modeling local transient fluctuations and global temporal dependencies. SCGRM effectively extracts consistent spatial features from both dynamic and static fMRI data by explicitly modeling their collaborative interaction rather than treating them independently. STKAN integrates the extracted spatiotemporal features for final classification. Experiments on the ADHD-200 dataset, involving 747 subjects across seven sites, demonstrate that AE-STEN achieves a classification accuracy of up to 76.06% ± 0.65%. Moreover, AE-STEN identifies brain regions associated with ADHD consistent with clinical findings, indicating strong interpretability and highlighting the model's potential for clinical application.

Electroencephalogram (EEG) signals have unique individual characteristics and have broad application prospects in identity authentication. At present, person identification (PI) based on EEG using the temporal-spatial-spectral feature extraction framework has achieved remarkable success. However, the existing methods suffer from coupled cross-domain feature parameters and insufficient feature fusion during feature extraction, which limits the recognition ability. Moreover, fixed-scale feature extractors can hardly exploit the subject-specific multi-scale information. To address these challenges, we propose CMDFM: a complete multi-domain decoupled fusion model for EEG-based PI. Firstly, we design an independent temporal-spatial-spectral attention mechanism to eliminate cross-domain parameter coupling. Secondly, a full-domain fusion mechanism is designed to comprehensively integrate the features of the temporal domain, spatial domain and spectral domain. Finally, an adaptive multi-scale CNN is designed to adjust the contribution of the multi-scale convolution kernel, thereby making full use of individual-specific multi-scale information. We use four datasets to verify our method. The experimental results show that our method is superior to all the state-of-the-art methods. The code of CMDFM is at https://github.com/2538441690/CMDFM.

Reliable individual identification via functional connectivity (FC) enables accurate prediction of cognitive and behavioral traits, and facilitates the advancement of personalized medicine. Compared to statistical and deep learning methods, component decomposition demonstrates great potential as it explicitly models common components thereby captures individual-specific components through residuals. However, current decomposition methods are unable to directly extract individual-specific features, which introduces errors and limits the performance of individual identification. Facing this bottleneck, we find that low-rank and sparse (LS) decomposition algorithms can distinctly separate commonalities and individual-specific variations through low-rank and sparse components. To design an LS decomposition algorithm for individual identification, we need to tackle two challenges. On the one hand, current algorithms rely on a pre-set sparse threshold, making it impossible to determine whether the sparse threshold is optimal. On the other hand, the existing LS decomposition algorithm has a slow convergence rate. In our work, we propose a Low-Rank and Sparse Decomposition Model with Adaptive Regularization and Residual Feedback (ARRF-LS). The adaptive regularization strategy based on sparse feedback significantly improves the recognition accuracy by self-calibrating the sparse threshold to determine the optimal threshold. The residual balancing mechanism that adjusts the step size through iterative feedback greatly accelerates the convergence rate in LS decomposition. Extensive experiments demonstrate that our ARRF-LS framework outperforms all state-of-the-art methods and exhibits strong performance in cross-task individual identification.

PurposeThis study aimed to develop and validate a radiomics nomogram that integrates multi-parametric MRI and clinical factors for the preoperative prediction of parametrial invasion (PMI) in early-stage cervical cancer (ECC).Materials and methodsA total of 363 patients with ECC (FIGO stages IB-IIA) were divided into training, internal validation, and external validation cohorts. All patients underwent T2WI, DWI, and T1c scans before radical hysterectomy. Radiomics features were extracted from T2WI, DWI, and T1c images, and selected using the max-relevance and min-redundancy (mRMR) method and the least absolute shrinkage and selection operator (LASSO). Radiomics signatures were then derived from these selected features. An MRI model was built using the radiomics signatures to evaluate their performance in distinguishing patients with PMI. A radiomics nomogram was constructed based on the optimal radiomics signature, pre-procedure hematocrit levels, and CA-125 levels. The discrimination performance of the nomogram was subsequently evaluated.ResultsFor the MRI model, the radiomics signatures yielded AUCs of 0.834 (95% CI: 0.7275-0.9399) and 0.800 (95% CI: 0.6902-0.9105) in the internal and external validation cohorts, respectively. The radiomics nomogram, which integrated the radiomics signatures from T2WI, DWI, and T1c, along with hematocrit and CA-125 levels, showed excellent discrimination between PMI and non-PMI groups. The nomogram achieved an AUC of 0.827 (95% CI: 0.7116-0.9430) in the internal validation cohort and 0.806 (95% CI: 0.6997-0.9114) in the external validation cohort. The specificity and sensitivity were 0.866 and 0.762, respectively, in the internal validation cohort, and 0.875 and 0.583 in the external validation cohort.ConclusionsThe developed radiomics nomogram provides a non-invasive and reliable tool for preoperative PMI prediction in ECC. By facilitating more accurate risk stratification, it has the potential to inform personalized therapeutic planning.

Constructing dynamic virtual brain models is essential for understanding brain functions and pathological mechanisms, crucial in computational neuroscience. Current modeling methods can be grouped into two paradigms: deep learning models for accurate simulation, and neural dynamics models emphasizing physiological interpretability. However, these methods entail a fundamental tradeoff between accuracy and interpretability. To address this challenge, we introduce the neurodynamics-informed brain simulator (NDIB-Sim), a multimodal bidirectional physics-informed neural network (PINN) model. NDIB-Sim is a unified framework integrating a data-driven module constrained by multimodal data and a multiscale neural dynamics mechanism module, jointly optimized under a composite loss function. It contains two data loss and two physical constraint terms. This design ensures that the generated brain signals adhere to fundamental neurophysiological principles while achieving high fidelity to empirical data. We also designed a dynamic weighting strategy to adaptively balance these objectives during optimization. This framework simultaneously addresses the forward problem of predicting long-term brain activity and the inverse problem of estimating individual-specific neurophysiological parameters. Extensive experiments demonstrate that NDIB-Sim can achieve high-fidelity long-term brain activity prediction from short-term observations, with an average functional connectivity similarity above 0.97. The inferred effective connectivity (EC) shows excellent reliability and strong alignment with underlying structural and functional architecture. When applied to Alzheimer’s disease (AD) classification, these subject-specific parameters achieve high accuracy in distinguishing cognitively normal (CN) individuals from AD patients. This work presents a powerful computational framework that effectively reconciles mechanistic interpretability with data-driven performance, offering a novel approach for exploring brain dynamics and identifying potential disease biomarkers.