Computed tomography (CT) plays a vital role in clinical diagnostics, but normal-dose CT (NDCT) involves high radiation risks. Low-dose CT (LDCT) reduces radiation exposure, but often introduces noise and artifacts that degrade image quality. Recent advances have explored diffusion-based reconstruction techniques to enhance image quality in LDCT acquisitions. But existing approaches predominantly focus on image domain features and fail to effectively integrate the sinogram domain prior. Moreover, the current approaches are limited by the local perception of CNN and the secondary complexity of Transformer, which makes it challenging to achieve optimal structural-textural harmonization. Therefore, we propose a dual-domain Mamba-Wavelet diffusion model (MWDiff) for LDCT reconstruction, which fully utilize the generation capability of the diffusion model and the remote modeling capability of Structured State Space Model (SSM) while maintaining the linear computational complexity. First, MWDiff decomposes images and constructs high/low frequency dual-branch pathways by discrete wavelet transform (DWT). The high-frequency branch enhances texture and edge features while suppressing noise and artifacts. The low-frequency branch optimizes the global structural information. Next, the features from dual-branch are integrated within a Mamba-based denoising module to further eliminate residual noise. Finally, to improve the reconstruction quality from three aspects: pixel accuracy, multi-scale representation and ray information fidelity, we design a multi-domain cooperative loss function to jointly constrain the image, wavelet and sinogram domain. We evaluated MWDiff on three different dose CT datasets, and comprehensive validation shows that MWDiff is superior to competing methods in terms of quantitative indicators and visual quality.

Currently, existing single-domain and single-modality fake news detection methods fall short of meeting the performance requirements of real-world applications. However, existing multi-domain multimodal models still struggle with semantic misalignment caused by cross-modal heterogeneity and poor generalization due to significant domain shifts. To tackle these issues, this paper introduces a novel deep learning framework called SMMDFND. Central to this framework are two innovative fusion mechanisms designed to bridge these gaps. First, we propose the Cross-Modal Synergistic Attention (CMSA) module, which mitigates modal heterogeneity by enabling deep, multi-granular interaction between textual and visual features. Second, to address domain generalization, we introduce the Dynamic Frequency Feature Fusion (DFF) module. Drawing on the concept of frequency in signal processing, DFF treats instance-level multimodal details as "high-frequency" components and categorical domain priors as "low-frequency" components. It then adaptively fuses these streams to balance local evidence with global context. Comprehensive experiments on the widely adopted public benchmark dataset Weibo21 demonstrate that the proposed SMMDFND substantially outperforms existing state-of-the-art methods across multiple evaluation metrics. The source code for this study is publicly available at https://anonymous .4open.science/r/SMMDFND-241C.

Unsupervised feature selection (UFS) can enhance the generalization ability, reduce the computational and storage requirements, and accelerate the training speed by removing irrelevant or redundant features. Subspace learning based UFS has garnered widespread attention, however, many fail to fully leverage the intrinsic statistical information of data when designing a distance metric to measure the distance between the original space and its subspace, leading to the suboptimal solutions. Moreover, some methods frequently overlook feature redundancy or rely heavily on prior knowledge, which may deviate from the true case. To address these limitations, this paper proposes a redundancy-aware UFS approach in the framework of sparse subspace learning, which uses variance-covariance distance as loss function and the l2,1 norm to promote sparsity. Meanwhile, subspace learning minimizes the distance between the original and the target space, and maximizes the projection variance of the data simultaneously. To effectively eliminate redundancy, we fuse multiple relationship metrics of features into an accurate and comprehensive one, which is further used to guide low-redundancy learning. Furthermore, we propose to use Alternating Direction Method of Multiplier (ADMM) to optimize the resulting model, which is often effective and efficient. The extensive experimental results on multiple public datasets demonstrate the effectiveness and superiority.

Visual simultaneous localization and mapping (VS-LAM) is fundamental for autonomous perception and navigation. However, existing systems often suffer from degraded localization and reconstruction performance in highly dynamic scenes, weak-texture regions, and illumination variations. To address these challenges, this paper proposes DANI-SLAM, a dynamic-aware neural implicit SLAM framework that integrates learning-based feature representation, dynamic scene understanding, and neural implicit mapping into a unified pipeline for robust pose estimation and reliable scene reconstruction. Specifically, DANI-SLAM employs a dynamic-aware keyframe selection strategy to suppress unreliable observations and stabilize front-end tracking. An uncertainty-aware loop closure mechanism is introduced by incorporating pose covariance into candidate retrieval and re-ranking, improving loop reliability and global consistency. Furthermore, an uncertainty-guided pixel-level neural implicit sub-map fusion strategy is designed to suppress inconsistent observations caused by dynamic objects or occlusions, resulting in more complete and geometrically consistent reconstructions. Extensive experiments on four public datasets, including the TUM RGB-D, Replica, ScanNet, and Bonn RGB-D datasets, demonstrate that DANI-SLAM consistently outperforms representative traditional and learning-based SLAM methods in both localization and reconstruction tasks. In highly dynamic TUM RGB-D sequences, DANI-SLAM achieves up to 48% lower ATE RMSE than baseline methods, while also obtaining superior reconstruction quality in terms of Depth L1, PSNR, SSIM, and LPIPS. These results validate the robustness, generalization, and effectiveness of the proposed system. The source code is available at: https://github.com/DANI-SLAM/DANI-SLAM.

The widespread application of Large Language Models, coupled with the broad reach of social media, has facilitated the generation of convincing misinformation, posing challenges to societal trust. However, existing detection methods predominantly treat veracity assessment (true vs. false) and generation provenance (human vs. machine) as isolated tasks. We argue that this separation is suboptimal because the manifestation of falsehood differs fundamentally by source: human-authored misinformation often relies on emotional manipulation and rhetorical bias, whereas LLM hallucinations typically exhibit logical inconsistencies and semantic drift. To address this limitation, we propose a unified multimodal dual-detection framework designed to optimize these coupled objectives. Rather than relying on explicit provenance detectors, the proposed approach implicitly leverages learned source patterns to dynamically recalibrate the decision boundary for veracity assessment. Specifically, the framework integrates fine-grained cross-modal alignment to capture the subtle semantic drifts characteristic of machine-generated text. It is noted that the generation provenance objective within this study focuses strictly on textual anomalies, relying on authentic visual media rather than addressing artificial image generation or deepfake detection. Experiments on benchmarks containing both human-authored and LLMgenerated news demonstrate that the proposed framework achieves competitive performance compared to state-of-the-art baselines, particularly in distinguishing subtle hallucinations that traditional detectors often overlook. The code involved in this study is publicly available on website https://github.com/SWLee777/ ProVer.

Magnetic resonance imaging (MRI) is regarded as the clinical diagnostic gold standard. However, its lengthy scan times introduce motion artifacts, which can severely compromise diagnostic accuracy. K-space undersampling is a fundamental strategy to address this issue, but undersampling inevitably introduces quality degradation in reconstructed images. To tackle the challenges in accelerated MRI reconstruction, this paper proposes a Multi-Feature Guided Progressive Divide-and-Conquer reconstruction network (MFG-PDAC). It achieves synergistic optimization through three novel modules. The Multi-Frequency Gated Attention (MFGA) module enhances feature propagation, the Edge Enhanced Feature Modulation (EEFM) module reinforces anatomical boundaries, and the Frequency-Aware Data Consistency (FREDC) module optimizes spectral reconstruction. These three modules form a closed-loop mechanism consisting of feature selection, spatial optimization, and frequency-domain correction. The MFGA enables dynamic fusion of multi-frequency features at U-Net skip connections, providing structural priors for gradient modulation. The gradient modulation amplifies edge response in the image domain, improving high-frequency reconstruction quality. The FREDC dynamically weights constraints based on frequency band errors, creating a feedback mechanism for MFGA refinement. Evaluated on the fast MRI knee dataset, MFG-PDAC achieved a peak signal-to-noise ratio of 37.28 dB and structural similarity index measurement of 0.909 under 8× acceleration, outperforming the current mainstream methods. The network particularly can achieve better reconstruction in key diagnostic regions such as bone-soft tissue interfaces and ligament textures. This study provides an accurate and efficient solution for clinical rapid MRI scanning, demonstrating significant potential for clinical translation.

Accurate segmentation of nuclei is essential for structured analysis in digital pathology, directly influencing cancer diagnosis, grading, and prognosis. However, obtaining high-quality pixel-level annotations remains labor-intensive and time-consuming. Existing self-supervised learning methods primarily rely on low-level appearance representations to construct proxy tasks, which leads to limited sensitivity to nuclei morphology, suboptimal boundary discrimination, and insufficient modeling of spatial continuity. To mitigate these problems, a Self-Supervised Nuclei Segmentation Network with Saliency Guidance and Contextual Boundary Attention (SCB-Net) is proposed for instance segmentation of H E stained images. First, SCB-Net incorporates the nuclei saliency guided self-supervision mechanism to generate structure-aware pseudo labels, which leverages cytological priors as structural driving signals to extract salient features by enhancing edge gradients and the structural continuity of nuclei. Then, during the segmentation stage, to enhance the structural expressiveness of the model in complex nuclei clusters and ambiguous boundary regions, the Contextual Boundary Attention (CBA) mechanism is designed to strengthen boundary directionality, reinforce edge gradients, and optimize cross-channel information fusion. Furthermore, the Efficient Triple Fusion (ETF) module is employed to concurrently fuse three complementary feature types (semantic, geometric structural, and visual boundary) to achieve cross-scale semantic alignment and geometrically consistent representations. Evaluations on the MoNuSeg, CPM17, and CryoNuSeg datasets demonstrate that SCB-Net attains AJI scores of 0.5609, 0.5102, and 0.4740, respectively. These results are comparable to weakly supervised methods despite requiring no ground-truth labels, and significantly surpass existing self-supervised approaches.

RGB-Thermal semantic segmentation fuses visible texture details and infrared robustness to improve scene understanding in adverse conditions. However, existing RGB-T methods are inadequate in handling RGB feature degradation and illumination-aware fusion in extreme light. To address this issue, we propose an illumination-modality dual-awareness driven adaptive fine-grained fusion network for RGB-T semantic segmentation (IMAFNet). In this method, the illumination-aware dynamic adaptation module (IDAM) is first proposed to generate dynamic adaptive weight masks for improving the ability of light adaptation. After that, the gated attention-based modality-aware module (GAMAM) is proposed to enhance the ability of multi-modal collaborative expression. In which, GAMAM uses the dynamic adaptive weight masks generated by IDAM for modal recalibration. Then, the integrated features from each modality are split into two streams, feeding into the next encoder stage and the Feature Fusion Module (FFM), respectively. Finally, the fused features are decoded to restore spatial resolution and produce the final segmentation. Extensive experiments on two benchmark datasets of MFNet and PST900 demonstrate that the proposed IMAFNet achieves state-of-the-art performance.

Subspace clustering is an advanced technique that identifies clusters embedded within a union of low-dimensional subspaces of the original data space, thereby revealing its intrinsic structure. Spectral clustering-based methods have gained significant attention in computer vision, image processing and pattern recognition due to their promising performance. However, existing approaches, which typically rely on self-representation for representation coefficient learning, often lack robustness and struggle to comprehensively characterize complex data structures. Traditional reconstruction loss based on the Frobenius or l1 norm are susceptible to noise and outliers. Furthermore, many methods underutilize inherent data characteristics for capturing local geometric structures and adapting to intricate data relationships. To address these limitations, this paper proposes a novel subspace clustering approach, named Multi-graph Fusion Guided Robust Adaptive Learning (MFGRAL), which integrates robust adaptive representation and multi-graph fusion within a unified framework. Specifically, a non-convex logarithmic loss function is adopted to enhance robustness against noise and outliers. To better preserve local manifold structures, a multi-graph fusion strategy is developed to guide the adaptive graph learning process. This facilitates the learning of more discriminative low-dimensional embeddings and enhances the capacity to capture complex neighborhood relationships. An effective and efficient optimization algorithm based on Alternating Direction Method of Multipliers (ADMM) is developed to solve the proposed model. Extensive experimental results on several benchmark datasets demonstrate the effectiveness of the proposed MFGRAL and its superiority over state-of-the-art methods.

Pogostemon cablin (Blanco) Benth (P.cablin), known for its unique aroma and rich chemical components, occupies an important position in the field of cosmetics, food and medicine. Identifying its origin is crucial for quality control and preventing adulteration. Traditional identification methods are time-consuming and labor-intensive, and usually require complex chemical analysis. In this study, a rapid and universal method was proposed to identify P.cablin from three major origins based on hyperspectral image (HSI) and deep learning, named DeepHSI. Furthermore, metabolomics and transcriptomics analyses were performed to validate the feasibility of HSI analysis for origins identification of P.cablin. HSI data collected under three experimental conditions (batches) were used for model training and transfer learning, which demonstrate the generality of DeepHSI. The simplified multi-origins identification model fusion mechanism ensures scalability for practical research applications and provides a paradigm for multi-classification research. These advantages provide a promising solution for rapid and nondestructive origin identification, quality control, and authenticity verification.

Alzheimer's disease (AD) classification based on brain functional networks has become a research hotspot in neuroimaging, showing significant clinical application potential. However, challenges remain in brain functional network sparsification, multimodal fusion, and interpretability. A sparsity reconstruction method based on Euler characteristics for brain functional networks is proposed, with classification performed by integrating the sparsified network structure, age, sex, and network features, and key brain regions are identified through gradient backpropagation. Specifically, a threshold sparsity method based on the Euler feature fitting curve is first adopted, and stable points in the network structure are identified by analyzing the trend changes of the Euler fitting feature curve. Then, a hypergraph convolutional neural network with a weighted fusion layer is constructed to integrate age, sex features, and the topological features of the sparsified brain functional network for accurate classification of different disease stages. Finally, gradient backpropagation-based localization is used to identify disease-related key brain regions. Experimental results demonstrate that the Euler characteristic-based sparsification method can objectively simplify and preserve critical information in brain functional networks, achieving an accuracy of 90.22% in the AD versus normal control classification task. Moreover, the identified key regions, the identified key regions, including the hippocampus and precuneus, show strong consistency with known clinical neuropathological findings, improving the interpretability of classification results.

Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that is widely used for high-resolution imaging of soft tissues and organs. However, the slow speed of MRI imaging, especially in high-resolution or dynamic scans, makes MRI reconstruction an important research topic. Currently, MRI reconstruction methods based on deep learning (DL) have garnered significant attention, and they improve the reconstruction quality by learning complex image features. However, DL-based MR image reconstruction methods exhibit certain limitations. First, the existing reconstruction networks seldom account for the diverse frequency features in the wavelet domain. Second, existing dual-domain reconstruction methods may pay too much attention to the features of a single domain (such as the global information in the image domain or the local details in the wavelet domain), resulting in the loss of either critical global structures or fine details in certain regions of the reconstructed image. In this work, inspired by the lifting scheme in wavelet theory, we propose a novel Fully Dual-Domain Contrastive Learning Network (FDuDoCLNet) based on variational networks (VarNet) for accelerating PI in both the image and wavelet domains. It is composed of several cascaded dual-domain regularization units and data consistency (DC) layers, in which a novel dual-domain contrastive loss is introduced to optimize the reconstruction performance effectively. The proposed FDuDoCLNet was evaluated on the publicly available fastMRI multi-coil knee dataset under a 6× acceleration factor, achieving a PSNR of 34.439 dB and a SSIM of 0.895.

BACKGROUND:Near-infrared spectroscopy (NIR) is widely used in online monitoring and quality control, due to its fast and non-destructive characteristics. Successful NIR analysis often requires accurate calibration models, relating spectral data to sample properties of interest. However, the same sample has different spectral responses in different instruments, preventing quantitative model established on high-quality instruments being directly applied across instruments. Calibration transfer address the above problem and replace the time-consuming and labor-intensive recalibration process. Through calibration transfer, spectra obtained from different instruments can be predicted using established calibration models with ideal performance. RESULTS:This study proposes a systematic calibration transfer method, combing principal component analysis and extreme learning machine, followed by an ELM based quantitative calibration model (PCELM-ELM). Three NIR benchmarking datasets of corn, tobacco and pharmaceutical tablets were used to test the reliability of proposed method. Comparing with competitive methods, PCELM-ELM method demonstrated state-of-the-art transfer capabilities and quantitative capabilities, and it achieved generally smaller root mean square errors of prediction (RMSEPs) than that of other methods. The introduction of partial least squares principal components significantly improves the performance of calibration transfer. The visualized variable importance and the weight of principal components in partial least squares explained the good transfer capabilities of PCELM. Thousands of modeling results with random parameters also demonstrate the robustness of the PCELM method. SIGNIFICANCE:The comprehensive results guarantee that PCELM-ELM is an accurate and practical method to transfer the NIR spectra of the slave instrument toward a well-established and maintained calibration model without costly and time-consuming recalibration. Results across diverse datasets confirm PCELM-ELM is a promising calibration transfer and quantitative method in NIR application.

Unsupervised Feature Selection (UFS), which identifies the optimal-related feature subset from the original feature set to lower the dimensionality of data without label information, has had a high profile in recent years. Given the absence of label information, the existing UFS approaches usually utilize graph and manifold learning techniques to retain the intrinsic structure of the data. The inclusion of irrelevant and redundant features and noise, would inevitably lower the quality of the structure. For this purpose, in this paper, we come up with Robust Sparse Orthogonal Basis Clustering (RSOBC), a novel method for UFS that integrates feature selection process with clustering task into a unified framework. Instead of explicitly utilizing the pre-computed local information, such a strategy focuses on exploring the inherent clustering structures of data. RSOBC leverages the log-based function as the loss to lessen the effect of noise and outliers, thereby enhancing its robustness. To select the more useful and discriminative features, the l2,1 norm is employed as the sparse regularization to encourage sparsity of the projection matrix. Meanwhile, we adopt the low redundancy regularization to make the weights of the correlated features small. In this way, the correlated features cannot be selected simultaneously. Consequently, the projection matrix, centroid matrix and cluster label matrix are learned simultaneously, such that the intrinsic structure is constructed in amore accurate way. The resulting optimization can be readily tackled by multi-block Alternating Direction Method of Multipliers (ADMM) based algorithm. Comprehensive experiments have been carried out on nine diverse real-world datasets. The results demonstrate that RSOBC surpasses many state-of-the-art UFS approaches, which indicates its effectiveness and superiority.

As a pivotal strategy to deal with complicated and high-dimensional data, subspace clustering is to find a set of subspaces of a high-dimensional space and then partition each data point in dataset into the corresponding subspace. This field has witnessed remarkable progress over recent decades, with substantial theoretical advancements and successful applications spanning image processing, genomic analysis and text analysis. However, existing surveys predominantly focus on conventional shallow-structured methods, with few up-to-date reviews on deep-structured methods, i.e., deep neural network-based approaches. In fact, recent years has witnessed the overwhelming success of deep neural network in various fields, including computer vision, natural language processing, subspace clustering. To address this gap, this paper presents a comprehensive review on subspace clustering methods, including conventional shallow-structured and deep neural network based approaches, which systematically analyzes over 150 papers published in peer-reviewed journals and conferences, highlighting the latest research achievements, methods, algorithms and applications. Specifically, we first briefly introduce the basic principles and evolution of subspace clustering. Subsequently, we present an overview of research on subspace clustering, dividing the existing works into two categories: shallow subspace clustering and deep subspace clustering, based on the model architecture. Within each category, we introduce a refined taxonomy distinguishing linear and nonlinear approaches based on data characteristics and subspace structural assumptions. Finally, we discuss the challenges currently faced and future research direction for development in the field of subspace clustering.

A Cyber-Physical System (CPS) incorporates communication dynamics and software into phsical processes, providing abstractions, modeling, design, and analytical techniques for the system. Based on spatial temporal graph neural networks (STGNNs), anomaly detection technology has been presented to detect anomaly data in smart grids with good performance. However, since topological changes of power networks in smart grids often already predict the occurrence of anomalies, traditional models based on STGNNs to portray network evolution cannot be directly utilized in smart grids. Our research proposed a smart grid anomaly detection method on the grounds of STGNNs, which used evolution in the information of several attributes that affected the power network to represent the evolution of the power network, subsequently used STGNNs to obtain the time-space dependencies of nodes in several information networks, and used a cross-domain method to help the anomaly detection of the power network through anomaly information of other related networks. Laboratory findings reveal that the abnormal data detection rate of our scheme reaches 90% in the initial stage of data transmission and outperforms other comparative methods, and as time goes by, the detection rate becomes higher and higher.

Background Alzheimer's disease (AD) is a neurodegenerative disorder. There are no drugs and methods for the treatment of AD, but early intervention can delay the deterioration of the disease. Therefore, the early diagnosis of AD and mild cognitive impairment (MCI) is significant. Structural magnetic resonance imaging (sMRI) is widely used to present structural changes in the subject's brain tissue. The relatively mild structural changes in the brain with MCI have led to ongoing challenges in the task of conversion prediction in MCI. Moreover, many multimodal AD diagnostic models proposed in recent years ignore the potential relationship between multimodal information. Objective To solve these problems, we propose a multimodal fine-grained classification model based on deep metric learning for AD diagnosis (DML-MFCM), which can fully exploit the fine-grained feature information of sMRI and learn the potential relationships between multimodal feature information. Methods First, we propose a fine-grained feature extraction module that can effectively capture the fine-grained feature information of the lesion area. Then, we introduce a multimodal cross-attention module to learn the potential relationships between multimodal data. In addition, we design a hybrid loss function based on deep metric learning. It can guide the model to learn the feature representation method between samples, which improves the model's performance in disease diagnosis. Results We have extensively evaluated the proposed models on the ADNI and AIBL datasets. The ACC of AD vs. NC, MCI vs. NC, and sMCI vs. pMCI tasks in the ADNI dataset are 98.75%, 95.88%, and 88.00%, respectively. The ACC on the AD vs. NC and MCI vs. NC tasks in the AIBL dataset are 94.33% and 91.67%. Conclusions The results demonstrate that our method has excellent performance in AD diagnosis.

BACKGROUND:Cardiac magnetic resonance imaging (CMR) provides critical pathological information, such as scars and edema, which are vital for diagnosing myocardial infarction (MI). However, due to the limited pathological information in single-sequence CMR images and the small size of pathological regions, automatic segmentation of myocardial pathology remains a significant challenge. PURPOSE:In the paper, we propose a novel two-stage anatomical-pathological segmentation framework combining Kolmogorov-Arnold Networks (KAN) and Mamba, aiming to effectively segment myocardial pathology in multi-sequence CMR images. METHODS:First, in the coarse segmentation stage, we employed a multiline parallel MambaUnet as the anatomical structure segmentation network to obtain shape prior information. This approach effectively addresses the class imbalance issue and aids in subsequent pathological segmentation. In the fine segmentation stage, we introduced a novel U-shaped segmentation network, KANMambaNet, which features a Dual-Stream Fusion Mamba module. This module enhances the network's ability to capture long-range dependencies while improving its capability to distinguish different pathological features in small regions. Additionally, we developed a Kolmogorov-Arnold Network-based multilayer perceptron (KAN MLP) module that utilizes learnable activation functions instead of fixed nonlinear functions. This design enhances the network's flexibility in handling various pathological features, enabling more accurate differentiation of the pathological characteristics at the boundary between edema and scar regions. Our method achieves competitive segmentation performance compared to state-of-the-art models, particularly in terms of the Dice coefficient. RESULTS:We validated our model's performance on the MyoPS2020 dataset, achieving a Dice score of 0.8041  ± $\pm$  0.0751 for myocardial edema and 0.9051  ± $\pm$  0.0240 for myocardial scar. Compared to the baseline model MambaUnet, our edema segmentation performance improved by 0.1420, and scar segmentation performance improved by 0.1081. CONCLUSIONS:We developed an innovative two-stage anatomical-pathological segmentation framework that integrates KAN and Mamba, effectively segmenting myocardial pathology in multi-sequence CMR images. The experimental results demonstrate that our proposed method achieves superior segmentation performance compared to other state-of-the-art methods.

Medical image registration is a fundamental and indispensable element in medical image analysis, which can establish spatial consistency among corresponding anatomical structures across various medical images. Since images with different modalities exhibit different features, it remains a challenge to find their exact correspondence. Most of the current methods based on image-to-image translation cannot fully leverage the available information, which will affect the subsequent registration performance. To solve the problem, we develop an unsupervised multimodal image registration method named DTR-GAN. Firstly, we design a multimodal registration framework via a bidirectional translation network to transform the multimodal image registration into a unimodal registration, which can effectively use the complementary information of different modalities. Then, to enhance the quality of the transformed images in the translation network, we design a multiscale encoder–decoder network that effectively captures both local and global features in images. Finally, we propose a mixed similarity loss to encourage the warped image to be closer to the target image in deep features. We extensively evaluate methods for MRI-CT image registration tasks of the abdominal cavity with advanced unsupervised multimodal image registration approaches. The results indicate that DTR-GAN obtains a competitive performance compared to other methods in MRI-CT registration. Compared with DFR, DTR-GAN has not only obtained performance improvements of 2.35% and 2.08% in the dice similarity coefficient (DSC) of MRI-CT registration and CT-MRI registration on the Learn2Reg dataset but has also decreased the average symmetric surface distance (ASD) by 0.33 mm and 0.12 mm on the Learn2Reg dataset.

Objective. Convolutional neural network (CNN) is developing rapidly in the field of medical image registration, and the proposed U-Net further improves the precision of registration. However, this method may discard certain important information in the process of encoding and decoding steps, consequently leading to a decline in accuracy. To solve this problem, a multi-channel semantic-aware and residual attention mechanism network (MSRA-Net) is proposed in this paper. Approach. Our proposed network achieves efficient information aggregation by cleverly extracting the features of different channels. Firstly, a context-aware module (CAM) is designed to extract valuable contextual information. And the depth-wise separable convolution is employed in the CAM to alleviate the computational burden. Then, a new multi-channel semantic-aware module (MCSAM) is designed for more comprehensive fusion of up-sampling features. Additionally, the residual attention module is introduced in the up-sampling process to extract more semantic information and minimize information loss. Main results. This study utilizes Dice score, average symmetric surface distance and negative Jacobian determinant evaluation metrics to evaluate the influence of registration. The experimental results demonstrate that our proposed MSRA-Net has the highest accuracy compared to several state-of-the-art methods. Moreover, our network has demonstrated the highest Dice score across multiple datasets, thereby indicating that the superior generalization capabilities of our model. Significance. The proposed MSRA-Net offers a novel approach to improve medical image registration accuracy, with implications for various clinical applications. Our implementation is available at https://github.com/shy922/MSRA-Net.