Multimodal ultrasound imaging, combining B-mode ultrasound, shear wave velocity, and shear wave time, plays a crucial role in breast tumor diagnosis by providing detailed insights into tumor characteristics. However, challenges such as inter-modal feature misalignment and attention drift complicate accurate segmentation due to variations in how each modality captures tissue properties. To address issues, we propose the MusoMamba framework and the MUB2025 dataset. MusoMamba enhances segmentation accuracy through bidirectional alignment and region-specific feature enhancement, utilizing the Modality-Collaborative Learning and Modality-Guided Enhancement modules. The MUB2025 dataset, comprising paired images across the three modalities from 506 cases, supports detailed analysis and model training. MusoMamba achieves strong results on MUB2025, with a 72.67
To address the issues of poor handling of ground object change boundary details in complex scenes and insufficient synergy caused by the decoupling of semantic and change subtasks in existing semantic change detection methods for remote sensing images, this paper proposes an improved algorithm based on contrastive learning. A Cross-Layer Attention Fusion FPN Decoder is designed. Through a multi-stage feature fusion strategy, it integrates low-level spatial details and high-level contextual information to generate more discriminative feature representations. Additionally, a Pixel-Level Change Modeling with Classification Features method is proposed, which shifts change modeling from the decision layer to the feature fusion stage. A dynamic change-aware loss is introduced to establish a collaborative training mechanism between the semantic segmentation and change detection tasks, enhancing the model’s sensitivity to change regions. In contrastive learning, a boundary hard sample mining strategy is proposed to prioritize the optimization of category boundary pixels, strengthening the model’s ability to recognize complex boundaries. The method is validated on two public datasets. Experimental results show that, compared with existing methods, the proposed method achieves significant improvements on multiple evaluation metrics for the tasks of change detection and semantic segmentation.
Although vision foundation models like SAM2 excel at natural image segmentation, their direct application to medical image segmentation faces severe domain gap challenges due to unique imaging principles, complex noise characteristics, and specialized semantic requirements. We propose AMoE-SAM2UNet, a cross-domain medical image segmentation framework that integrates an Adaptive Mixture of Experts Decoder (AMoE-Decoder) with the SAM2 encoder. The AMoE-Decoder comprises three specialized expert modules: the Multi-Scale Adaptive Expert (MSAE) captures lesion scale variations through parallel dilated convolutions and adaptive fusion operations; the Boundary Refinement Expert (BRE) sharpens blurred boundaries by enhancing high-frequency information via wavelet transform; and the Foreground Enhancement Expert (FEE) leverages multi-branch convolutions and spatial-channel attention to enhance foreground-background discrimination, addressing low contrast and complex background. Moreover, a lightweight gating network dynamically combines expert outputs based on input characteristics. Extensive experiments on five diverse medical imaging modalities—dermoscopy (ISIC 2017), ultrasound (TN3K), endoscopy (ClinicDB), CT (COVID), and histology (GLAS)—demonstrate that AMoE-SAM2UNet achieves consistent Dice coefficient improvements ranging from 1.8
Fluorescein Fundus Angiography (FFA), although widely used to visualize the retinal vasculature to monitor Diabetic retinopathy (DR), typically requires the intravenous injection of sodium fluorescein, a fluorescent contrast agent, making it an invasive procedure. In contrast, Color Fundus (CF) images are much more accessible and cheaper to collect. Generation of FFA based on CF thus become an effective approach to address such problem. However, previous works and existing image generation methods often struggle to preserve fine-grained details and anatomical integrity in terms of FFA generation. To address this limitation, we propose the Guidance-Controllable Mamba Diffusion Model (OTMamba) capable of generating high-fidelity FFA images directly from CF photographs. In contrast to existing GAN-based and diffusion-based medical image generation approaches that suffer from high computational complexity, limited scalability, and poor transferability, our proposed method demonstrates enhanced performance and clinical relevance in ophthalmology image translation. Experimental results confirm the model’s superiority in synthesizing diagnostically valuable FFA images, providing a promising non-invasive alternative for diabetic retinopathy screening and management.
As typical complementary modalities, infrared and visible images integrate spectral information to enhance perceptual quality and scene understanding, particularly under challenging imaging conditions. However, resolution disparity between modalities, which is caused by sensor limitations and hardware heterogeneity, poses significant challenges for effective fusion. To address the issue mentioned above, we propose a Mixture-of-Experts (MoE)-based fusion framework specifically designed for heterogeneous-resolution infrared and visible image pairs, named HeRIF. Specifically, the dual-branch encoder-decoder from HeRIF reconstructs low-resolution semantics and high-resolution residuals, with the fusion model exploiting cross-modal high-resolution guidance to refine low-resolution outputs and improve visual fidelity. The restoration module incorporates two MoE units, each dedicated to processing the output features of a distinct encoder, thereby facilitating high-resolution detail restoration tailored to inputs of varying resolutions. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in terms of fusion quality and downstream tasks, providing a scalable and robust solution for multi-resolution image fusion.
Rapid development and application of diffusion models has led to a surge in diffusion generated video content, increasing the risk of fake information spreading. Therefore, it is imperative to develop a generalized and robust detector for diffusion generated videos. However, existing detection methods lack generalized and robust representations, failing to generalize across videos created by multiple diffusion models. In this paper, we point out that text-vision embedding provides a generalized and robust representation for diffusion generated video detection, building on the success of text-image contrastive learning. Visual embeddings from text-visual contrastive pre-trained models inherently separate real and diffusion generated videos. Leveraging this insight, we propose a generalized detection framework that exploits contrastive embeddings between visual content and text semantics to capture generation artifacts. A comprehensive evaluation across multiple video generation models shows our detector’s superior generalized capability. On unseen diffusion video models, it achieves an average accuracy 0.9229. The code is available at https://github.com/1129ljc/T2VE .
Existing methods for cropland change detection are hindered by interference from coupled high and low frequency features and bi-temporal feature misalignment. To address these issues, we propose the Dual-Frequency Decoupling and Dynamic Fusion Network (DFDFusion), a dual-stream CNN-Transformer framework incorporating a three-stage synergistic optimization mechanism. First, the Dual-Frequency Domain Decomposition and Enhancement Module (DF ^3 -AE) decouples high-frequency textures and low-frequency morphologies using multi-scale dynamic weighting, enhancing feature representation for fragmented boundaries and macro contours. Second, the Bi-temporal Multi-Scale Attention Fusion Module (BMSAF) aligns bi-temporal features via channel-spatial attention, effectively mitigating seasonal interference. Finally, the Hierarchical Cooperative Attention Module (HCAM) progressively refines features through channel reweighting, region focusing, and pixel refinement, addressing edge blurring and semantic fragmentation. Experiments on the CLCD, PX-CLCD, and LEVIR-CD datasets demonstrate DFDFusion’s superior performance, achieving F1 scores of 74.90
The automatic generation of radiology reports has emerged as an effective solution to address time-consuming clinical tasks and accurately identify key pathological features in X-ray images, attracting significant attention in recent years. Cross-modal mapping between images and text, a critical component for generating high-quality reports, presents considerable challenges due to the lack of corresponding annotations. In this study, we propose a Cross-modal Memory Alignment Framework with Disease Aware Contrastive Learning to generate coherent and informative reports by aligning visual and textual features. Specifically, we first embed both image and text features into a shared space through parametric query vectors, ensuring tight alignment between images and reports. The memory-aligned embeddings are then obtained by querying a memory matrix, where the query signals are derived from visual features. These embeddings subsequently guide the visual-text feature alignment during the report generation process. For model optimization, we integrate reinforcement learning with knowledge distillation techniques to achieve efficient knowledge transfer through iterative training. Extensive experiments on the IU X-ray and MIMIC-CXR benchmark datasets demonstrate that our method generates more accurate and clinically relevant reports, which further demonstrates the effectiveness of the proposed alignment method.
Semi-Supervised Learning (SSL) enhances medical image segmentation by leveraging unlabeled data to alleviate the scarcity of annotated samples. Although vision foundation models (e.g., the Segment Anything Model, SAM) demonstrate strong generalization and feature extraction capabilities, their integration with SSL remains underexplored. To address this, we propose a novel dual-branch SSL-SAM framework that enables iterative pseudo-label refinement through a co-training mechanism, harnessing labeled and unlabeled data. For enhanced segmentation precision, we introduce Cross-Path Prompt Fusion (CPF), which facilitates dynamic interaction between prompt-based segmentation paths, enhancing feature representation. Additionally, we develop Frequency Enhanced Pseudo-label Filtering (FEPF), a denoising strategy that suppresses high-frequency noise while preserving anatomical boundaries, improving pseudo-label reliability. Comprehensive experiments on four medical image datasets demonstrate that our approach consistently outperforms existing SSL and SAM-based methods, particularly in data-scarce scenarios. Code is available at: https://github.com/Heirudy/Dual-SAM .
A core challenge in knowledge graph reasoning lies in reconciling topological structure awareness with semantic correlation modeling. When updating knowledge graphs based on embedding representations from local or global perspective, traditional approaches often encounter significant limitations due to noisy neighborhood interference, excessive local smoothing, and a lack of long-range dependency modeling. To address these issues, we propose DNPrune-GT, a Dynamic Neighbor Pruning Graph Transformer that facilitates multi-granularity reasoning through a semantically guided hierarchical architecture. Specifically, DNPrune-GT introduces a dynamic pruning mechanism based on semantic similarity to construct context-aware subgraphs by retaining the top-k most relevant neighbors at each layer, effectively suppressing noise propagation. Furthermore, we leverage the modularity-optimized Leiden algorithm to partition the graph into functional communities, enabling a hierarchical reasoning process that transitions from local aggregation to community-level reorganization and global interaction. To further enhance the model’s capacity to capture long-range semantic dependencies across communities, we design a topology-aware graph Transformer module that incorporates community positional encodings and multi-order adjacency matrices. Extensive experiments on the FB15k-237 and WN18RR datasets demonstrate the superiority of DNPrune-GT, achieving mean reciprocal rank (MRR) scores of 0.385 and 0.497, respectively surpassing the best-performing Graph Transformer baselines by 3.5
Semi-supervised medical image segmentation (SSMIS) has shown promise in model training for limited medical labeling data. However, the dominant teacher-student modeling approach is subject to confirmation bias due to erroneous pseudo-labeling effects. In this research we present the Adaptive Diversity-Alternated Teaching (ADAT) framework, which is based on the average teacher approach and involves one student and two teachers. The framework integrates labeled and unlabeled data into a dual-teacher network for adaptive alternating instruction, leveraging inter-teacher prediction discrepancies to dynamically guide multi-perspective student learning. Specifically, ADAT dynamically coordinates dual-teacher instruction through predictive entropy weights in the Diverse Alternate Information (DAI) Exchanging Module. Hybrid images are generated by foreground-background swapping between labeled/unlabeled data batches, ensuring teaching diversity via distinct teacher inputs. Meanwhile, the Adaptive Discrepancy Module (ADM) categorizes teacher predictions into consensus and discrepancy regions for targeted student learning. By comparing the ratio of these parts, we adaptively adjust the learning degree of student at congruent and discrepant parts. The Dice Score is 2.51
Most existing low-light image enhancement (LLIE) methods mainly focus on adjusting overall brightness of images. However, low-light images typically suffer from a low signal-to-noise ratio (SNR), making them highly prone to noise and visual artifacts after enhancement. These artifacts severely degrade the visual quality and naturalness of enhanced results. To address this problem, we propose a novel SNR-Aware Mamba-Transformer (SMT) network for high-quality low-light enhancement. Specifically, we introduce an SNR-aware Mamba module for effective global noise modeling, guided by pixel-level SNR cues. Additionally, we design an eight-directional mamba scanning mechanism to enable multi-perspective contextual feature aggregation, which further improves noise robustness and spatial coherence. To restore fine-grained details, we incorporate a multi-scale attention aggregation module that efficiently fuses features at different levels. Extensive experiments on eight public benchmark datasets demonstrate that our SMT not only achieves visually superior results but also outperforms or matches state-of-the-art methods on multiple quantitative image quality metrics.
Real-time dynamic reconstruction and visualization of endoscopic scenes are critical for advancing digestive medicine applications. Existing methods based on Gaussian Splatting often suffer from geometric inconsistencies under complex occlusions due to inadequate geometric constraints during explicit modeling. To overcome these limitations, we propose Endo2DGS, a novel reconstruction framework that enables efficient 4D dynamic modeling through three key innovations. First, we simplify the 3D Gaussian ellipsoid to a 2D Gaussian ellipse, ensuring robust geometric consistency across multiple views. Second, we introduce learnable deformation parameters for each Gaussian primitive, enabling precise representation of tissue surface deformations. Third, we develop a depth-guided normal estimation approach that enforces geometry-constrained optimization, significantly improving shape recovery in occluded regions. Our framework achieves an optimal balance of visual quality, computational efficiency, and geometric accuracy, demonstrating substantial clinical potential. Experimental results on the EndoNeRF dataset show that Endo2DGS delivers superior visual quality (38.294 PSNR), real-time performance (356.79 FPS), low GPU memory consumption (1.6 GB), and high geometric fidelity (1.875 DMSE).
Fine-needle aspiration biopsy (FNAB) is a fundamental diagnostic procedure for thyroid nodules, in which accurate judgment of puncture positioning, specifically determining whether the needle tip is correctly located within the nodule, is essential for reliable pathological assessment. However, evaluating puncture positioning during FNAB remains challenging due to the complex and subtle visual cues in ultrasound images. To address this, we propose UL-TPNet, an ultra-lightweight deep learning-based network for Thyroid Puncture Positioning Detection guided by Nodule Location. With only 0.3M parameters, UL-TPNet achieves precise identification of the needle tip’s position relative to the nodule, providing real-time and accurate assistance to clinicians during FNAB procedures. To validate the effectiveness of our approach, we constructed the Thyroid Nodule Ultrasound (TNUS) dataset, which includes fine-grained annotations of puncture positioning. Experimental results demonstrate that UL-TPNet achieves an accuracy of 99.31
Camouflage, a vital survival strategy in nature, allows organisms to evade predators through environmental mimicry. However, the high similarity between camouflaged objects and backgrounds in color, texture, and contour poses significant detection challenges. While recent studies have achieved promising progress, they mainly focus on spatial features and lack multi-scale and cross-domain fusion. To this end, we propose WaveCamoNet, a cross-domain fusion model for camouflaged object detection. The model extracts frequency domain features by wavelet transform and fuses them with multi-scale spatial features to enhance semantic representation. Also, we design a texture enhancement module to refine high-frequency details and suppress background noise. Experiments on three challenging benchmark datasets demonstrate that our WaveCamoNet significantly outperforms the existing state-of-the-art CNN-based methods under four widely-used evaluation metrics.
Accurate early diagnosis of colorectal cancer (CRC) is crucial for patient survival. However, multi-modal analysis using Narrow Band Imaging (NBI) and Chromoendoscopy Imaging (CI) is often challenged by data scarcity and ineffective information fusion. In this paper, a novel unsupervised data augmentation method, called Pathology Information-Guided Diffusion Model (PIG-Diff), was proposed to address the issue of CRC data scarcity. In PIG-Diff, a General Segmentation Large Model (GSLM) is introduced to generate pathological segmentation masks, which are then combined with text prompts as dual priors to guide diffusion models in generating high-quality images. Moreover, to improve the classification accuracy of CRC, a Dual-branch Residual Wavelet Attention Network (DB-ResWANet) was proposed, which integrates a Wavelet Attention Module (WaveletAtt) for enhancing NBI feature details and a Spatial Feature Alignment Module (SFAM) for adaptive NBI-CI feature fusion. Furthermore, an auxiliary modality classification task is employed to enhance modality-aware representations. Experimental results on the CRC magnifying endoscopy dataset indicate that the image generation quality of the PIG-Diff surpasses that of other existing image generation methods. The DB-ResWANet method, which was trained with PIG-Diff augmented data also outperformed various representative medical image classification methods. Our code and dataset is available at https://github.com/helinrui/DB-ResWANet .
Hybrid Transformer-Mamba is a promising foundational deep model for medical image segmentation. Existing hybrid models are constructed by empirically interleaving Transformer and Mamba blocks, which fail to mitigate the feature biases between different deep models and lack a comprehensive understanding of local, long-range, and longer-range features in medical images. To address these issues, this paper proposes a novel hybrid model with U-Net architecture, called TM-UNet. The encoder of TM-UNet is implemented using hybrid CNN-Transformer-Mamba (CTM) blocks. Each CTM block comprises two parallel Transformer branches and CNN-Mamba branch, interacted by a dual cross-attention module. The CTM block achieves more effective multi-hierarchical feature representation than standard cascading structure, owing to its parallel branch design and dual cross-attention mechanism. Unlike typical U-Net with symmetrical encoder-decoder, the decoder of TM-UNet is implemented using convolutional layer and cascaded fusion module. Additionally, an attention based feature fusion block is developed to self-adjust features for final prediction. Extensive experiments on the ISIC2018, BUSI, Kvasir-SEG, and LiTS datasets demonstrate the superiority of TM-UNet. Specifically, TM-UNet achieves 1.07 https://github.com/yongchangxu/TM-Unet .
Text image editing is a critical task in computer vision and graphics research, with diverse applications such as advertising design and social media content creation. Generating high-quality text images with readability and consistency using Diffusion Models (DFMs) poses a challenge due to their inadequate handling of text styles and glyphs. In this paper, we present DiffCTE (Consistent Visual Text Editing with High Style Fidelity via Diffusion Model), a novel approach that enhances the generation of readable and stylistically consistent text images by providing DFMs with improved style condition guidance. Leveraging a vision language model (VLM) consisting of a Vision Encoder, Style Abstractor, and Text Decoder, the Style Abstractor effectively extracts style condition and represents style information across languages, encompassing aspects such as glyph, font, and color. By integrating the style condition into DFMs, DiffCTE enables the generation of visually coherent and stylistically consistent text. Specifically, the generated text achieves style transfer by maintaining consistency with the reference style. Our comprehensive experiments demonstrate the superior performance of DiffCTE compared to existing methods, setting new directions for future research in text image editing with DFMs.
Tooth segmentation plays a crucial role in oral disease diagnosis, orthodontic treatment planning, and computer-aided diagnosis (CAD). However, the low-contrast and high-noise characteristics of cone-beam computed tomography (CBCT) images, along with blurred anatomical boundaries, weak semantic connectivity, and the lack of effective frequency-domain information, make accurate segmentation challenging for existing methods. To address this challenge, we propose DDWFNet, a multi-scale segmentation network that integrates dense connections with wavelet transform. The model features two key innovations: (1) The Dynamic Wavelet Transform Convolution (DWTC) module enhances feature representation in low-contrast, noisy regions through multi-scale frequency-domain analysis. (2) The Quaternary Wavelet Hybrid-subband Attention (QWHA) mechanism suppresses background noise and strengthens semantic connections through cross-scale feature fusion and adaptive weighting. We validated DDWFNet on both the 3D CBCT and ISIC-2018 datasets. The results show that DDWFNet improves the IoU by 5.99
