Generic deep learning (DL) networks for image restoration like denoising and interpolation lack mathematical interpretability, require voluminous training data to tune large parameter sets, and are fragile in the face of covariate shift. To address these shortcomings, we build interpretable networks by unrolling variants of a graph-based optimization algorithm of different complexities. Specifically, for a general linear image formation model, we first formulate a convex quadratic programming (QP) problem with a new $\ell _{2}$ -norm graph smoothness prior called gradient graph Laplacian regularizer (GGLR) that promotes piecewise planar (PWP) signal reconstruction. To solve the posed unconstrained QP problem, instead of computing a linear system solution straightforwardly, we introduce a variable number of auxiliary variables and correspondingly design a family of ADMM algorithms. We then unroll them into variable-complexity feedforward networks, amenable to parameter tuning via back-propagation. More complex unrolled networks require more labeled data to train more parameters, but have better overall performance. The unrolled networks have periodic insertions of a graph learning module, akin to a self-attention mechanism in a transformer architecture, to learn pairwise similarity structure inherent in data. Experimental results show that our unrolled networks perform competitively to generic DL networks in image restoration quality while using only a fraction of parameters, and demonstrate improved robustness to covariate shift.

Due to the complexity and low contrast of medical images, existing methods face significant challenges in detecting weak objects with large size variations. To mitigate this issue, we propose a Multiscale Fusion Network with Boundary Preservation (MFBP-Net) for medical image object detection. Specifically, MFBP-Net incorporates a Dynamic Multi-Dilated Feature Fusion (DMDF) mechanism to adaptively adjust the receptive field. To maintain geometric continuity during feature reconstruction, a Direction-aware Upsampling (DAU) strategy is employed to reinforce the representation of slender structures and weak boundaries. Additionally, a Dynamic Multi-Scale Feedback Head (DMSFH) is introduced to establish a semantic feedback pathway that coordinates shallow texture cues with deep contextual representations. By integrating these complementary components, MFBP-Net captures both local detail and global context. Experiments on Br35H and Endo-PR demonstrate that MFBP-Net outperforms current advanced one-stage detection methods and exhibits superior stability and adaptability for detecting and precisely localizing fine structures in complex medical images.

Learning image representations with deep self-supervised models is an important task in computer vision, which aims to establish beneficial and general representations from unlabeled images. However, existing efforts train models mainly on high-level features, neglecting lower-level features and their global spatial information, thus limiting the discriminative power of the learned representations. In this work, we propose a representational learning model based on restricted attention feature fusion network (RAFFNet) to improve the quality and generalization of the learned image representations. Specifically, to fully exploit the features in the deep network, we use a self-supervised model on multi-level features to learn more general representations. Meanwhile, a new feature fusion strategy with a dual attention mechanism of channel and space is used for multi-level features, enabling the model training to obtain more important and comprehensive feature information. Furthermore, in order to better extract global spatial information, we devise a simple but effective attentional weighted mask, which restricts the weight of spatial attention and prevents the model from focusing only on local features with high attention weights. Experiments on four public classification datasets, CIFAR-10, CIFAR-100, Tiny ImageNet and ImageNet-1%, and two object detection datasets, PASCAL VOC and COCO, demonstrate that the proposed RAFFNet has better representation performance and generalization ability than most state-of-the-art image representation learning algorithms.

Due to the irregularity of spatial structures and the dynamic variation of signals across both space and time, spatio-temporal graph signal denoising is a typical ill-posed problem. This paper proposes a lightweight and interpretable neural network framework based on mixed graph structures and deep unfolding algorithms for spatio-temporal graph signal denoising. Specifically, we construct and jointly leverage temporal directed graphs and spatial undirected graphs to characterize the temporal correlations and spatial consistency of traffic data. Gradient-informed graph Laplacian and graph total variation are introduced to quantify and promote graph signal smoothness jointly. Furthermore, we design an iterative optimization algorithm based on the Alternating Direction Method of Multipliers (ADMM) and unfold its iterative steps into a feed-forward neural network based on end-to-end training. Experimental results demonstrate that the proposed method achieves significant performance in traffic time-series denoising tasks, exhibiting superior robustness and low parameter complexity.

Retinal vessel segmentation requires not only high pixel-wise precision but also faithful preservation of vascular connectivity. Conventional pixel-wise supervision may lead to fragmented centerlines and locally inconsistent vessel paths under low contrast, imaging noises, and complex thin bifurcations. To address these issues, we propose an edge-guided residual mixing and directional-enhanced atrous fusion network (EDNet). Based on the nnU-Net baseline, EDNet focuses on boundary and direction-aware representations within the encoder–decoder pipeline. Specifically, Edge-guided Residual Mixer (EGRM) is integrated into encoder stages to strengthen edge-related and axis-consistent responses while suppressing isotropic interference. Directional-Enhanced Atrous Mixer (DEAM) is placed before the deepest skip connection to compactly fuse multi-angle priors, global semantics, and atrous multi-scale context for decoder-friendly deep features. Extensive experiments on DRIVE, CHASE_DB1, and FIVES demonstrate that the proposed EDNet outperforms the competing methods and produces visually more coherent vessel trees with fewer breakpoints, particularly in thin and highly tortuous regions.

Federated learning (FL) enables multiple clients to collaboratively train machine learning (ML) models without exposing their private data. The recent surge in poisoning attacks and privacy leakage against FL has driven the development of secure federated learning (SFL) solutions. However, existing SFL schemes show inadequate performance when confronted with non-independent and identically distributed ( non-IID) data. In addition, traditional SFL architectures are prone to single point of failure (SPOF) issues due to the heavy computational and communication burdens imposed on a single server. In response to these issues, a novel blockchain-based secure federated learning (BSFL) framework is proposed in this paper. Specifically, we devise a proof of verification (PoV) consensus protocol to identify poisoning attacks under non-IID situations, while preventing the waste of computational and communication resources. Subsequently, we present a personalized differential privacy (PDP) mechanism, which achieves comprehensive privacy protection with lower noise levels. Furthermore, the integration of the blockchain with the proposed reward mechanism overcomes SPOF and fosters constructive participation through transparent processes. Formal theoretical analysis demonstrates the security, privacy, and efficiency of our framework. Extensive experimental evaluations indicate that BSFL exhibits strong resilience against various poisoning attacks and achieves better model accuracy compared to existing SFL solutions.

Diffusion models have made significant progress in high-fidelity image synthesis and editing, yet high-quality object removal remains challenging. Existing methods lack fine-grained semantic control over image-conditioning signals. Moreover, generating editing masks from cross-attention maps often leads to irrelevant regions in the masks. In this paper, we propose a training-free, dual-branch diffusion-control framework that improves object removal performance by decomposing the image-conditioning signals and optimizing the editing mask. Specifically, we introduce a semantics-aware IP-Adapter attention modulation strategy that leverages the decomposability of image-conditioning embeddings. This strategy dynamically rescales cross-attention maps throughout multi-scale denoising to attenuate target features and strengthen background consistency. In parallel, we design a noun-aware mask optimization mechanism that parses non-target nouns in the prompt to locate distractor regions and progressively suppresses their spurious responses, producing precise editing masks. Experiments show that the framework improves object-removal stability and controllability, and outperforms existing methods in mask purity, semantic alignment, and background integration.

Developing a precise parathyroid identification model is difficult due to the intricacies of thyroid surgical environments, including obstruction of surgical instruments, light variations, extrusion, and deformation. Given that traditional detectors (such as Faster R-CNN) rely on limited receptive fields, they cannot effectively solve the problems of occlusion, deformation, and data scarcity in PG detection. This research presents an adaptive capsule graph neural network (ALCapsule-GNN) for parathyroid gland (PG) detection, a new model that integrates capsule networks with graph convolutional networks (GNN) to improve the efficacy of PG object detection tasks in medical images. The model employs a capsule network-based Backbone to extract multi-scale features and constructs an adaptive graph using a dynamic graph generator, where nodes represent capsules and edges encode spatial-semantic relationships. The hierarchical graph capsule network incorporates an attention mechanism to refine the graph structure while leveraging a shared weight mechanism to minimize redundant computations. Iterative message passing enhances target detection performance and improves the robustness to occlusion and deformation. ALCapsuleGNN efficiently captures spatial relationships and contextual information among objects while preserving the model’s computational efficiency and adaptability. Comprehensive trials in a proprietary parathyroid data set demonstrate that our method achieves exceptional results in automated parathyroid detection.

While deep learning techniques, such as Convolutional neural networks (CNNs), show significant potential in medical applications, real-time detection of parathyroid glands (PGs) during complex surgeries remains insufficiently explored, posing challenges for surgical accuracy and outcomes. Previous studies highlight the importance of leveraging prior knowledge, such as shape, for feature extraction in detection tasks. However, they fail to address the critical multi-scale variability of PG objects, resulting in suboptimal performance and efficiency. In this paper, we propose an end-to-end framework, MSWF-PGD, for Multi-Scale Weighted Fusion Parathyroid Gland Detection. To improve accuracy and efficiency, our approach extracts feature maps from convolutional layers at multiple scales and re-weights them using cluster-aware multi-scale alignment, considering diverse attributes such as the size, color, and position of PGs. Additionally, we introduce Multi-Scale Aggregation to enhance scale interactions and enable adaptive multi-scale feature fusion, providing precise and informative locality information for detection. Extensive comparative experiments and ablation studies on the parathyroid dataset (PGsdata) demonstrate the proposed framework’s superiority in accuracy and real-time efficiency, outperforming state-of-the-art models such as RetinaNet, FCOS, and YOLOv8.

Lightweight super-resolution (SR) has garnered attention for balancing performance and efficiency in resource-constrained environments. In lightweight SR tasks, traditional CNN-based methods are constrained by limited receptive fields, leading to suboptimal SR performance. In contrast, ViT-based models achieve remarkable results but suffer from significant computational burden due to the self-attention mechanism. In this paper, we propose adaptive Pixel Classification and equivalent Large Kernels Network (PCLKN), a novel lightweight SR model that addresses the limitations of traditional CNN-based and ViT-based methods. PCLKN utilizes equivalent large kernels to expand the receptive field while relying solely on convolutional operations, significantly reducing computational overhead. Additionally, it integrates global priors with spatial and channel attention to enhance feature extraction and leverages adaptive pixel classification to utilize similar pixel information for reconstruction. Experimental results on benchmark datasets demonstrate that PCLKN achieves superior SR performance with an excellent trade-off between performance and computational complexity.

Hyperspectral image (HSI) denoising is a critical preprocessing step in numerous HSI applications. However, existing model-based and deep learning methods struggle to balance model complexity and interpretability. To address local feature modeling and parameter scale control, this paper proposes a hyperspectral image denoising method based on deep algorithm unfolding with Laplacian regularization and spatial-spectral hybrid total variation. Specifically, we combine graph Laplacian regularization with hybrid spatial-spectral total variation to jointly model the local features and global spatial-spectral characteristics of HSI. The deep unfolding network architecture, based on deep algorithm unfolding, leverages the layer independence and stable convergence of the ADMM iterations, enabling the construction of an end-to-end network model with a small number of parameters. Experimental results on both simulated and real datasets demonstrate that the proposed model achieves significant performance improvements in HSI denoising compared to existing techniques.

In complex visual scenarios, people can understand occluded objects through contextual reasoning and association. However, existing video object detection (VOD) methods could not work well. They treat all frames equally during feature aggregation and ignore the interdependence of objects between adjacent frames. This results in excessive noise during long-term feature alignment. To address these issues, we propose a task-specific spatiotemporal context-aware decoupling (TSCD) network, which adapts the global semantic information and local localization information needed for classification and regression tasks, respectively. Furthermore, a spatiotemporal context-aware feature matching (CAFM) module is utilized to refine the regression features while avoiding inter-frame feature mixing. Effective object associations are used to improve the regression of occluded object locations. Extensive experimental results on the ImageNet VID dataset and the OVIS dataset show that our proposed TSCD has significant advantages in occlusion scenes. The code is available at https://github.com/KaihongLi/TSCD.

The great success of diffusion models in the text-to-image field has driven the increasing demand for fine-grained local image editing. One of which is changing the viewpoint of objects to given positions in accordance with 3D geometric principles. How to keep the surrounding region unchanged and maintain structural and semantic consistency when editing the designated objects is a challenging yet widely applicable task. However, existing methods often fail to maintain correct geometric structure and editing efficiency simultaneously. To this end, we explore the geometric structure changes of images when the viewpoint changes from the perspective of 3D camera projection and propose a geometry-aware local viewpoint editing approach that requires neither 3D reconstruction nor model training, and performs editing solely at a single timestep in the latent space of diffusion models. Central to our approach is constructing latent-space location mappings across different viewpoints by integrating multi-view geometry theory with the absolute depth information. To address assignment conflicts and latent feature missing problems while enhancing detail fidelity, we design an occlusion reasoning mechanism and a foreground-background aware bilateral interpolation strategy. Additionally, a consistency-preserving strategy is introduced to enhance alignment with the original image. Extensive experiments on image datasets demonstrate the overall advantages of our approach in structural consistency and runtime efficiency.

This paper investigates the problem of circular structure detection, which is of great significance in medical detection scenarios. Existing detection methods usually treat this problem as a special case of object detection and adapt existing object detectors. However, these methods ignore the inherent properties of graphs, leading to poor model-checking performance and weak generalization ability. Compared with complex shape targets, the geometric properties of circles are simpler, so the graphics parameters can be found in the complex parameter space through specific feature extractors for shape detection. This paper presents an end-to-end deep learning framework for circular object detection in medical images. Leveraging deep features extracted by convolutional neural networks (CNNs), the proposed method integrates a dynamically differentiable Hough transform module to effectively map image-domain features to the parameter space of circles, defined by their center coordinates and radii. This parameter space's aggregation and projection mechanism concentrates features onto potential circular trajectories, generating distinct peak responses. The proposed approach achieves an efficient and fully differentiable detection pipeline by transforming object detection into a peak localization task, eliminating the need for traditional non-maximum suppression (NMS). In addition, this paper also uses the circular IoU loss function for circular objects. The experiment shows that the dynamic deep circular Hough transform and circular IoU loss function have better detection effects on medical circular objects.

Conventional model-based image denoising optimizations employ convex regularization terms, such as total variation (TV) that convexifies the ℓ_0-norm to promote sparse signal representation. Instead, we propose a new non-convex total variation term in a graph setting (NC-GTV), such that when combined with an ℓ_2-norm fidelity term for denoising, leads to a convex objective with no extraneous local minima. We define NC-GTV using a new graph variant of the Huber function, interpretable as a Moreau envelope. The crux is the selection of a parameter a characterizing the graph Huber function that ensures overall objective convexity; we efficiently compute a via an adaptation of Gershgorin Circle Theorem (GCT). To minimize the convex objective, we design a linear-time algorithm based on Alternating Direction Method of Multipliers (ADMM) and unroll it into a lightweight feed-forward network for data-driven parameter learning. Experiments show that our method outperforms unrolled GTV and other representative image denoising schemes, while employing far fewer network parameters.

ObjectiveTimely diagnosis of early gastric cancer (EGC) is significantly associated with patient prognosis, but traditional endoscopic diagnosis relies on the physician’s experience and has certain limitations. This study comprehensively evaluated the accuracy of artificial intelligence (AI) in the diagnosis of EGC through meta-analysis and compared the performance ability of different AI models.MethodsPubMed, Embase, Web of Science Cochrane Library, and China National Knowledge Infrastructure databases were systematically searched (established until January 2025), and studies evaluating the accuracy of AI models in the diagnosis of EGC were included, requiring reporting of sensitivity and specificity, or providing data for calculating these indicators. Data were extracted independently by two reviewers, and sensitivity and specificity were pooled using a bivariate random effects model, and subgroup analysis was performed by AI model type. The primary outcome measures were the summary sensitivity, specificity, and area under the curve (AUC) of all AI models.ResultsOf 26 studies involving 43,088 patients were included. Meta-analysis results showed that the summary sensitivity of the AI model was 0.90 (95%CI: 0.87-0.93), the specificity was 0.92 (95%CI: 0.87-0.95), and the AUC was 0.96 (95%CI: 0.94-0.98), respectively. Subgroup analysis showed that the sensitivity of deep convolutional neural network (DCNN) was higher than that of traditional CNN (0.94 vs 0.89), while the specificity was almost equivalent (0.91 vs 0.91). In dynamic video verification, the AUC of the AI model reached 0.98, which was significantly better than the clinician level (AUC 0.85-0.90).ConclusionThe AI model, especially the DCNN architecture, showed excellent accuracy in the diagnosis of EGC. Future research should focus on the dynamic effect of the model, improvement of interpretability, and multicenter prospective validation.Systematic Review Registrationhttps://www.crd.york.ac.uk/PROSPERO/view/CRD420251003071, identifier CRD420251003071.

Long-term time series forecasting holds significant applications in various fields. Although existing deep-learning models have made great progress, they still fall short in dealing with complex periodic patterns and residual prediction. This paper proposes an improved model named MultiCycleNet, which enhances the performance of CycleNet by introducing multiperiod modeling and residual prediction mechanisms. Specifically, we design an explicit cycle extraction mechanism based on modular mapping and channel-specific learnable embeddings, enabling the model to learn and represent multiple periodicities. Furthermore, we propose a residual prediction scheme combining one-dimensional average pooling and multi-scale denoising. The theoretical motivation lies in separating low-frequency trends from high-frequency noise, ensuring stable residual learning. Extensive experiments on electricity, transportation, finance, and tobacco datasets demonstrate that MultiCycleNet achieves a better balance between prediction accuracy, robustness, and computational efficiency.

Image hiding through the application of invertible neural network (INN) represents a significant branch within the realm of deep image hiding methodologies, characterized by a compact network architecture and a streamlined parameter count. Nonetheless, when juxtaposed with autoencoder-based approaches, existing INN methods often result in inferior image quality. To surmount this challenge, this paper introduces a novel masking-based image hiding paradigm, establishes a new spatial domain transformation for images, and refines the Swin-transformer block. By integrating these innovations, an INN architecture is crafted that is particularly adept for deep image hiding, termed NiNet. The experimental results demonstrate that NiNet can remarkably address the problem of image hiding. In terms of steganographic image quality, NiNet outperforms the current state-of-the-art (SOTA) algorithms by 0.26 dB on the DIV2K dataset, 1.49 dB on the COCO dataset, and 0.39 dB on the ImageNet dataset. Regarding the quality of secret image recovery, NiNet surpasses the SOTA algorithms by 2.06 dB on DIV2K, 1.98 dB on COCO, and 0.50 dB on ImageNet.

Node classification tasks have seen considerable progress with the use of Graph Neural Networks (GNNs). However, they cannot work well on imbalanced node classification and tend to prioritize the majority classes with more labeled instances while overlooking the minority classes with fewer labeled instances. Existing solutions focus on generating new nodes to augment the training set, which may disrupt the original topological structure of the graph, so GNNs may not achieve optimal classification results. To address this issue, we introduce GraphMMC, a reliable and flexible strategy to generate pseudo-labels that can be easily integrated with various GNNs, which will augment the imbalance training set to a class-balanced set without generating new nodes. We use the similarity between the unlabeled nodes and the minority classes to correction the lowconfidence pseudo-labels generated by GNNs to obtain reliable pseudo-labels. Our experiments demonstrate that the proposed method outperforms state-of-the-art baselines on several class-imbalanced datasets.