Mild cognitive impairment (MCI) is an early stage of Alzheimer's disease (AD). It is classified into two subtypes based on the cognitive domains affected. The first subtype, stable MCI (sMCI), primarily affects memory while the second, progressive MCI (pMCI), results in a decline in other cognitive functions. Computer-aided diagnosis systems use transfer learning to differentiate sMCI and pMCI by transferring knowledge from a source domain (e.g., normal controls or AD patients) to a target domain (sMCI/pMCI). Current methods identify AD-related regions in the source domain or manually generate attention maps post-training, then apply these to the target domain. While improving accuracy, these approaches depend on prior manual intervention and may not fully adapt to target domain characteristics. In this study, we propose the knowledge distillation with attention maps (KDAM) method to effectively differentiate between sMCI and pMCI from magnetic resonance imaging (MRI) data. This technique transfers two types of information from teacher models: (1) logits, which minimize prediction discrepancies in MRI data, and (2) feature representations. Furthermore, the integration of the attention generation module employs a dual attention mechanism. By incorporating self- and cross-attention, it captures both inter- and intra-domain interactions. Using feature representations from both source and target domains, attention maps are generated for individual MRI images, highlighting discriminative regions associated with AD, thereby enhancing differentiation between sMCI and pMCI. Our experimental findings on the ADNI dataset demonstrate that the proposed KDAM method consistently outperforms both conventional and state-of-the-art techniques in qualitative and quantitative evaluations.

Accurate gastric tumor segmentation and lymph node metastasis (LNM) prediction from computed tomography (CT) scans can provide sufficient and useful image information to guide the diagnosis and treatment of gastric cancer. Domain shift, arising from equipment variations and hospital population diversity in multi-source data, significantly challenges model generalization to unseen domains. In this paper, we study the problem of multi-source domain generalization in gastric tumor segmentation and LNM classification tasks. To tackle this challenge, we propose a novel multi-source multi-task meta-learning (M3L) framework with task-oriented distribution alignment to train a generalizable multi-task model for unseen domains. Specifically, each task prediction branch acts as a meta-learner and is optimized by a meta-learning strategy for task-specific generalization.To further mitigate distribution differences between domains, we introduce two task-oriented distribution alignment losses to directly regularize the two meta-learners, i.e., a segmentation-oriented distribution alignment loss for aligning segmentation-related features and a classification-oriented distribution alignment loss for aligning classification-related features. They are both designed as part of the meta-train and meta-test objectives, which facilitate high-order derivatives of parameters during the meta-optimization process to further enhance task-specific domain-invariant features. The extensive experiments are conducted on four CT datasets collected from four medical centers, and comprehensive ablation studies and comparisons with the state-of-the-art methods show the effectiveness of our method for multi-task generalization. Our code is publicly available at https://github.com/infinite-tao/M3L.

Breast tumor pathological imaging constitutes a cornerstone in the diagnosis and clinical management of breast cancer. However, inherent limitations in imaging technology result in significant resolution disparities across varying magnification levels, potentially compromising diagnostic precision. To address this challenge, we introduce a novel computational framework that integrates state space modeling with diffusion processes for pathological image super-resolution. Our approach leverages multi-plane feature alignment via Scale-Invariant Feature Transform to incorporate real optically zoomed images, moving beyond conventional degradation-based super-resolution training and enabling our Diff-Magnifier model to learn authentic optical zoom characteristics. We incorporate a state space module into the diffusion encoder to exploit its capacity for capturing long-range dependencies, thereby enhancing the feature representation capability of the diffusion model for pathological images. Furthermore, we utilize the temporal modeling characteristics of the state space model to formulate the denoising process of the diffusion model. A joint training strategy is employed to optimize the combined architecture through loss minimization, while also reformulating the inference process to achieve accelerated sampling. Experimental validation demonstrates that our super-resolved images substantially enhance the accuracy of downstream automated pathological classification tasks. This work provides a significant advancement toward improving computational pathology workflows for breast cancer diagnosis. Our code is available at: https://github.com/magnetliu/diff-magnifier.

In the context of widespread data transmission, ensuring the digital integrity and intellectual property rights of images has become increasingly challenging. Existing color image watermarking methods continue to struggle with the trade-off between robustness and imperceptibility, because the embedding method with fixed position ignores block-level energy distribution. Uniform RGB quantization overlooks differences in channel sensitivity, and the quantization parameters are often selected empirically. To address these gaps, this paper proposes a blind color image watermarking method based on Hessenberg decomposition and variable quantization modulation. The Hessenberg decomposition is used to identify energy-dense coefficients for watermark embedding, while a hierarchical quantization strategy is designed based on RGB channel correlations. An enhanced offset determined by energy concentration and a dual-layer embedding architecture with primary-secondary collaboration are further constructed to improve extraction stability. In addition, an exponential-trigonometric optimization algorithm is introduced to adaptively search for the optimal quantization step. Experimental results show that the proposed method maintains PSNR values above 40 dB and average NC values above 0.91 under various attacks. Compared with existing methods, the proposed scheme improves PSNR by approximately 2.8 % and SSIM by about 0.5 % on average, with embedding and extraction times of 0.3966 s and 0.0758 s, respectively.

Parkinson's disease (PD) is a neurodegenerative disease that is prevalent in the elderly population. At present, the diagnostic method for PD mainly relies on the clinical experience of doctors, which is limited by its locality and subjectivity. Clinical multimodal data can provide different perspectives to understand patients and can ef fectively improve the accuracy of diagnosis. However, noise interference, high-dimensional feature complexity and category imbalance in multimodal data challenge PD diagnosis. Therefore, to solve the above problems, we propose a multimodal flexible subspace learning framework for PD diagnosis and clinical score regression, which learns the key features from multimodal data. Specifically, first, we provide diversified information by fusing PD neuroimaging data(DTI and MRI) and genetic data. Secondly, the self-representation learning method is used to dynamically capture sample similarities, thereby effectively mining the category information between samples. Finally, the l2,0 norm is used for dynamic feature selection, and the relevant features are enhanced by utilizing the similarity between features, while irrelevant and redundant features are excluded. In this way, the intrinsic struc tural properties of the data are learned, and then the classification and regression tasks are performed through Support Vector Machine(SVM). Our extensive experiments on the Parkinson's Progression Markers Initiative (PPMI) dataset demonstrate that this algorithm significantly improves both PD classification(94.07% accuracy for PD vs. NC at baseline) and clinical score regression performance(1.10 Mean Absolute Error for GDS prediction at baseline).

Infrared-visible video fusion under extremely low-light conditions is critically important yet remains underexplored, largely due to the scarcity of high-quality datasets and challenges posed by spatiotemporal uncertainty and modality bias. To address the dataset shortage, we built a dataset of 4,739 infrared and visible registration video pairs captured under extremely low-light conditions, spanning 5 scene types and 17 subcategories. Further, we proposed an Uncertainty-guided Spatiotemporal Consistency Fusion Network, termed USCFNet, for the infrared-visible video fusion. At each layer of the encoder, an Entropy-Gated SpatioTemporal Attention (EGSTA) module is introduced to capture temporal instability and spatial reliability variations through entropy-aware attention modulation, thereby enhancing feature spatiotemporal consistency. The refined infrared and visible features are then fused via a Difference-Guided Fusion (DGF) module, which adaptively exploits their content and edge differences to improve structural integrity and detail clarity. By progressively connecting DGF modules from shallow to deep layers, the network achieves the synergistic fusion of shallow textures and deep semantics. Subsequently, the output of the last DGF module is fused with the modality features of the last layer through a hierarchical mixture-of-experts fusion module. This module enables the balanced integration of modality information while preserving fine local details. Finally, the fusion feature is fed into the decoder to produce the final fused video. Extensive experiments on our dataset and two public datasets show that USCFNet outperforms competing methods, achieving lower distortion and stronger spatiotemporal consistency. The source code and dataset are available at https://github.com/Zhaocheng1/ELVID.

Existing multi-organ segmentation methods usually rely on large and fully labeled datasets for training. However, medical image datasets are typically decentralized by privacy constraints and partially labeled due to the high costs of full annotation in clinical practice, resulting in label inconsistency across medical centers. Federated learning offers privacy-preserving decentralized training, but the label inconsistency leads to significant divergence in local model parameters across medical centers, thereby hindering the achievement of the global optimum. To resolve this issue, an effective and communication-efficient Federated Learning under Reliable Supervision (FedRS) is proposed, which ensures: i) the local models are trained with reliable supervisory information through the proposed Less-Forgetting and Less-Constraint loss functions, thereby reducing the divergence in local model parameters; and ii) the global model is aggregated based on the consistency of predictions between each local model (after local training) and the global model (received before training), thereby enhancing the reliability of the global model. Extensive experimental results on nine publicly available 3D abdominal CT image datasets show that our FedRS outperforms localized, centralized, and state-of-the-art federated learning methods on both in-federation and out-of-federation datasets, demonstrating its effectiveness and strong generalization capability. In particular, our FedRS only utilizes a model with only 4.1M parameters as its backbone, thereby significantly reducing its communication cost. The source code is publicly available at https://github.com/luohy812/FedRS.

Accurate diagnosis of Parkinson's disease (PD) remains challenging due to heterogeneous symptoms and subtle neuroanatomical changes. Although multimodal Magnetic Resonance Imaging (MRI) provides complementary information, existing methods often fail to effectively model cross-modal interactions and brain anatomical connectivity. We propose a cross-modal graph attention network with self-supervised contrastive pretraining for PD diagnosis. Contrastive learning extracts robust volumetric representations from limited data, while a bidirectional cross-modal attention module enables deep interaction between T1-weighted MRI and quantitative susceptibility mapping features. An anatomically-informed graph attention network further models inter-regional dependencies to enhance diagnostic reliability and interpretability. Experiments on a hospital cohort demonstrate superior performance over representative methods.

Multimodal medical image fusion aims to integrate complementary information from different imaging modalities. This paper proposes a Spatial-Frequency Cooperative Fusion Network, termed SFCFNet, for more effective multimodal medical image fusion. SFCFNet follows an encoder-decoder structure, with two image encoders and one decoder. It employs a Spatial-Frequency Cooperative Fusion (SFCF) block to connect the two image encoders within each layer. Specifically, the SFCF block comprises a frequency-interaction module and a spatial-compensation module. The frequency interaction module employs a discrete Fourier transform to extract phase and amplitude components from both modalities, and then aggregates them separately across modalities using modality-guided gating units to capture global semantics. To extract the local details, the spatial compensation module formulates a refinement mechanism in the spatial domain that adaptively compensates for texture discrepancies between different modalities. Extensive experiments show the effectiveness and superiority of the proposed network over state-of-the-art methods.

Retinal image segmentation is essential for analyzing retinal structures like vessels and diagnosing retinopathy. However, the inherent intricacy of the retina, along with annotation scarcity and data heterogeneity, presents prevalent challenges in creating accurate and generalizable deep learning models. Diffusion models, while initially developed for image generation, have recently shown great promise for visual perception by leveraging the learned internal representations. However, these diffusion representations, which spread across network blocks (space) and diffusion timesteps (time), potentially suffer from issues like stochastic semantic distortion and cumulative structural blurring, compromising their semantic fidelity to the source image. In this paper, by delving into the generalization property of diffusion models, we propose a novel anchoring inversion strategy to derive diffusion representations that are semantically faithful to the source image from the deterministic trajectory. Furthermore, we introduce a time-space frequency-aware aggregation interpreter (T&S-FreqAgg) to aggregate the multi-scale and multi-timestep diffusion representations in a frequency-aware way for Domain Generalizable Semantic Segmentation (DGSS). Extensive experiments on nine public retinal image datasets demonstrate the superiority of our proposed framework, DiffDGSSv2, over state-of-the-art methods. Our code will be available at: https://github.com/Xyporz/DiffDGSSv2.

Myopic choroidal neovascularization (mCNV) is a common complication of high myopia. This disease will spread to the macular region of the retina, which has a serious impact on vision. Accurate segmentation of mCNV in retinal optical coherence tomography (OCT) plays a significant role in evaluating the condition of the disease and guiding the formulation of the treatment. However, accurate segmentation is challenging due to the blurred boundary, low contrast, tiny lesion areas, and high interference in OCT images. To this end, we propose a boundary-aware U-Net structure using two-path Transformers (i.e., BUT-Net) to accomplish lesion segmentation in mCNV using OCT images. Specifically, the encoder is built by two-path feature extractors, in which a fourstage Transformer is used to extract global features with long-distance dependency so that the network can express the morphological structure and tiny lesions. Four mixed blocks are designed by ResNet-34 and Transformer modules to excavate contextual and deep semantic features. The two-path fusion (TPF) module is utilized to fuse the two-path features by channel and spatial attention mechanism, which retains the local and global features of the fused features. For the decoder, we devise a dual attention gate (DAG) module to enhance the more complex channel dependencies in the up-sampling process. In addition, the boundary information is employed to refine the generated masks by weak supervision. The extensive experiments conducted on an inhouse OCT dataset and two public different modality datasets demonstrate that our proposed method achieves superior segmentation performance. Our code is available at https://github.com/shehare9517/BUT_Net.

Appendicitis is a common gastrointestinal disease, but appendiceal cancer is a rare malignancy. The similarities between the two diseases often lead to missed diagnosis in clinical practice, resulting in serious consequences. To address this problem, we propose a multi-module hybrid attention network (MHAN) to tackle the challenge of classifying appendiceal cancer and appendicitis in CT images. Our framework comprises four innovative modules. The multi-level feature extraction (MFE) enhances the multi-scale representation of small lesions. The hypergraph-based feature interaction (HFI) captures high-order cross-layer dependencies. The agent attention fusion (AAF) supports cross-layer recursive semantically guided enhancement. Finally, a reinforcement learning-based Dynamic loss optimization (RDLO) improves robustness in imbalanced data problem. We validate the proposed model on both internal and external datasets, and the results demonstrate significant improvements in lesion representation and diagnostic performance.

Accurate segmentation of lymphoma and metastatic tumor regions in ultrasound imaging is clinically crucial but remains challenging due to inherent modality artifacts, low lesion-to-parenchyma contrast, and particularly the heterogeneous echogenicity obscuring precise lesion boundary delineation. The scarcity of pixel-level expert annotations further impedes fully supervised deep learning solutions. To address this, we propose BGPCNet, a novel semi-supervised segmentation network that effectively leverages both limited labeled data and abundant unlabeled data by synergistically integrating Frequency Consistency Module(FCM) and Boundary Guided Patch Contrast(BGPC). Specifically, the FCM establishes robust structural coherence across different frequency sub-bands. It employs Discrete Wavelet Transform (DWT) to decompose spatial features into four frequency components, enhances each using Visual State Space (VSS) blocks inspired by VMamba, and enforces consistency between student and teacher model features in both spatial and frequency domains via a dedicated frequency consistency loss. Concurrently, the BGPC module explicitly tackles ambiguous boundary regions by formulating a contrastive learning strategy. It treats patches containing structure boundaries and central structure patches as positive and negative samples for each other, driving the model to learn discriminative features that better distinguish boundary pixels from the background and internal regions. Built upon the Mean Teacher framework with supervised pre-training, BGPCNet optimizes a combined loss function incorporating supervised loss, pixel-wise consistency loss, frequency consistency loss, and BGPC loss. Comprehensive evaluations on a large-scale private clinical superficial lymph node dataset and the public TN3k thyroid nodule dataset demonstrate BGPCNet’s superiority. It significantly outperforms state-of-the-art semi-supervised methods across key metrics.

With the explosive growth of the digital content industry and the increasingly urgent need for copyright protection, digital watermarking technology has become a core technical means for safeguarding related rights. This paper describes a novel multi-domain digital watermarking scheme based on Arnold-Logistic encryption and chaotic evolutionary optimization algorithm. The breakthrough of this method lies in constructing a multi-domain embedding framework by integrating the non-subsampled Contourlet transform (NSCT), discrete Hankel transform (DHT), and Hadamard transform (HT) to achieve maximum energy concentration, employing a dual-chaos encryption mechanism of scrambling followed by diffusion to enhance security, and utilizing a chaotic evolutionary optimization algorithm to adaptively determine the optimal quantization step size. The experimental results demonstrate that the watermarked images achieve an average Peak Signal-to-Noise Ratio (PSNR) greater than 40 dB and an average Structural Similarity Index Measure (SSIM) higher than 0.97, which indicates that the proposed method has good imperceptibility. Meanwhile, the extracted watermark maintains the NC generally above 0.97 and the Bit Error Rate (BER) below 0.05, indicating excellent robustness. Additionally, the encryption system boasts a key space as large as 2306, effectively resisting brute-force and statistical analysis attacks.

Parkinson's disease (PD) is an irreversible neurodegenerative disorder that significantly impacts patients' lives. Accurate early diagnosis prediction is crucial for providing timely treatment to delay disease progression. However, current diagnostic methods predominantly rely on the experience and judgment of clinicians, introducing subjectivity and a lack of standardized, quantitative measures. Sparse subspace learning, as a machine learning technique, can extract critical information from multimodal data while addressing issues such as noise, high-dimensional complexity, and class imbalance. Our study utilizes longitudinal, multimodal neuroimaging data collected at multiple time points to develop a diagnostic model for PD. The approach involves extracting latent local features and leveraging deep learning techniques to generate a comprehensive global feature subset. Adaptive sparse selection is employed to reduce feature redundancy. Finally, support vector machine is used for classification and regression tasks, specifically for PD diagnosis and disease progression score prediction. Extensive experiments were conducted on the PPMI dataset, achieving an accuracy of 90.78% for Scan Without Evidence of Dopaminergic Deficit (SWEDD) vs. Normal Control (NC) classification, 83.79% for PD vs. NC, and 91.50% for PD vs. SWEDD. The results demonstrate that the proposed method improves PD classification and prediction performance, showing promise for early diagnostic applications.

Federated Class Incremental Learning (FCIL) enables distributed clients to collaboratively train a global model based on their private sequential tasks without compromising data privacy. Currently, some FCIL methods have been proposed, and most are designed based on deep models. However, enabling these FCIL models to converge requires numerous communication rounds, significantly increasing communication costs. Recently, the Broad Learning System (BLS), an effective and efficient shallow model, was proposed and adapted for CIL tasks [i.e., BLS-Class Incremental Learning (CIL)]. BLS-CIL exhibits fast updates and high retainability. However, it requires prior knowledge of when new class data arrives and cannot be directly used in federated scenarios due to the global catastrophic forgetting in FCIL. Thus, an innovative Federated Class incremental learning method based on BLS (FedCBLS) is proposed, which extends BLS-CIL within the federated scenario and provides three advantages: 1) high accuracy from the local perspective, achieved by integrating BLS-CIL with a newly designed automatic decision-making (ADM) method to detect novel classes and learn them incrementally for local clients; 2) high accuracy from the global perspective, attained through the newly proposed local model refinement (LMR) and global model projection (GMP) methods, mitigating global catastrophic forgetting stemming from heterogeneous data across clients; and 3) extremely low communication costs due to the newly derived closed-form solutions without iterative optimization for both local and global models. Comprehensive experimental results show that our FedCBLS outperforms the state-of-the-art (SOTA) FCIL methods by up to 8.15%, while drastically reducing communication costs to 1% of SOTA’s. Our code is available at https://github.com/dujie-szu/FedCBLS.git

Digital watermarking technology is of vital importance for copyright protection. However, the existing methods can rarely simultaneously achieve the goals of invisibility, robustness, real-time performance, and security. This paper proposes a blind color image watermarking scheme based on Hankel kernel analysis space modulation. The Direct Current (DC) component of discrete Hankel transform is represented as a linear combination of the first column pixels, and the weights are determined by the roots of the zero-order Bessel function. Therefore, the embedding process is simplified to only perform minimum norm modification on the first column pixels, avoiding frequency transform, and using the lotus effect algorithm to find the optimal quantization step size. To resist strong scaling attacks, a blind synchronization mechanism based on implicit template matching and local block search is designed. The experimental results show that the average PSNR exceeds 40.3272 dB, and the SSIM exceeds 0.9639; in twenty attack categories, the average NC reaches 0.9962, and the BER is 0.0183, demonstrating excellent robustness under conventional, geometric, and composite attacks. Embedding and extraction only take 0.0561 s and 0.0421 s, respectively; the key space reaches 2432, effectively resisting statistical attacks and brute-force attacks. The proposed scheme successfully meets the comprehensive requirements of high invisibility, strong robustness, real-time processing, and high security in practical applications.

Cardiac segmentation is crucial in cardiac diagnosis. While many deep learning methods have achieved impressive results, they often require extensive annotated data, which is scarce due to the need for clinical experts to perform the annotations. Typically, only the end-diastolic (ED) and end-systolic (ES) phases of a cardiac cycle are annotated, leaving a substantial amount of data unannotated and underutilized. As a dynamic organ, the temporal information from other frames within the cardiac cycle is vital, even if only ED and ES are segmented. However, the lack of annotations for these frames makes it challenging to utilize this information effectively. Additionally, deformation during cardiac motion leads to discrepancies between frames, complicating their direct use. To tackle these issues, we propose a Pyramid Progressive Image Mapping Network (PPIM-Net), which maps unannotated frames to the spatial distribution of the ED and ES frames, thus leveraging unannotated data while addressing deformation. We introduced a Pyramid Progressive Image Mapping (PPIM) module that aligns unannotated frames with target frames using progressive window similarity registration. To enhance edge accuracy, we incorporated an Edge Correction Module (ECM) that refines edge information through window scanning. By mapping unannotated frames to annotated frames, we can supervise them using standard frame annotations, thereby improving the extraction of key structural details. We validated our method on a private dataset and two public datasets, achieving encouraging results that demonstrate effectiveness in utilizing unannotated data.

Accurate segmentation of cardiac structures in echocardiography is crucial for the early diagnosis of congenital heart disease. Existing methods are limited by the local receptive field of CNNs or the unidirectional scanning of Mamba, resulting in insufficient feature extraction. We propose a multi-directional UMamba network incorporating prototype priors for multi-view segmentation. The network includes a spatio-temporal graph attention module (S-T GAT) to capture dynamic dependencies, four-directional UMamba (FUMamba) for comprehensive structural modeling, and cross-view correlation attention (Cross Attention) to facilitate multi-perspective feature fusion, a prototype manager (Prototype) enhances feature stability under limited annotation. Experiments on multiple datasets demonstrate that our method outperforms other state-of-the-art approaches, achieving a Dice score of 93% on the A4C view and 92% on the PSAX view.

Neuroimaging AI remains constrained by a model-per-disease paradigm—where separate frameworks are trained for individual disorders—limiting knowledge transfer and direct cross-disorder comparisons. While resting-state fMRI provides a powerful subject-specific functional connectivity (FC) fingerprint, it suffers from low spatial resolution and limited clinical accessibility. Conversely, structural MRI (sMRI) is widely available and spatially detailed, yet existing representations typically rely on predefined features or disease labels, failing to explicitly encode the individual-level functional organization.We propose SF-DisenNet, a self-supervised framework that uses each subject’s own FC matrix as a neurobiological supervision signal to learn function-aligned sMRI representations without disease labels. A ResNet-DC patch encoder and an atlas-guided Anatomical Mapping Unit (AMU) aggregate local patches into AAL-90 regional embeddings, whose pairwise relationships define a representation-based structural connectivity (SC) matrix aligned with the subject’s FC. An independence regularization further promotes spatially disentangled regional representations. After one pretraining stage on UK Biobank, the encoder is transferred to downstream tasks without further updating. SF-DisenNet achieves 95.1% cross-modal fingerprint matching between sMRI-derived SC and fMRI-derived FC. The AMU exhibits emergent hemispheric lateralization across all 45 AAL-90 anatomical pairs with a median effective patch count close to one. On longitudinal ADNI data, the structural fingerprint achieves 83.5% 24-month re-identification accuracy; its drift (ΔSC) is 2.1× larger in mismatched subjects and correlates with hippocampal atrophy rate. Ultimately, this unified feature space supports fine-tuning-free transfer to AD/MCI, ASD, PD, and SWEDD tasks, enabling cross-disorder analysis within a single sMRI framework. The source code is available at https://github.com/k-Jayus/BRAIN.