Medical image analysis最新文献

{"title":"Fundus image quality assessment in retinopathy of prematurity via multi-label graph evidential network","authors":"Donghan Wu ,&nbsp;Wenyue Shen ,&nbsp;Lu Yuan ,&nbsp;Heng Li ,&nbsp;Huaying Hao ,&nbsp;Juan Ye ,&nbsp;Yitian Zhao","doi":"10.1016/j.media.2026.103959","DOIUrl":"10.1016/j.media.2026.103959","url":null,"abstract":"<div><div>Retinopathy of Prematurity (ROP) is a leading cause of childhood blindness worldwide. In clinical practice, fundus imaging serves as a primary diagnostic tool for ROP, making the accurate quality assessment of these images critically important. However, existing automated methods for evaluating ROP fundus images face significant challenges. First, there is a high degree of visual similarity between lesions and factors that influence quality. Second, there is a paucity of trustworthy outputs and interpretable or clinical-friendly designs, which limit their reliability and effectiveness. In this work, we propose a ROP image quality assessment framework, termed Q-ROP. This framework leverages fine-grained multi-label annotations based on key image factors such as artifacts, illumination, spatial positioning, and structural clarity. Additionally, the integration of a label graph network with evidential learning theory enables the model to explicitly capture the relationships between quality grades and influencing factors, thereby improving both robustness and accuracy. This approach facilitates interpretable analysis by directing the model’s focus toward relevant image features and reducing interference from lesion-like artifacts. Furthermore, the incorporation of evidential learning theory serves to quantify the uncertainty inherent in quality ratings, thereby ensuring the trustworthiness of the assessments. Trained and tested on a dataset of 6677 ROP images across three quality levels (i.e. acceptable, potentially acceptable, and unacceptable), Q-ROP achieved state-of-the-art performance with a 95.82% accuracy. Its effectiveness was further validated in a downstream ROP staging task, where it significantly improved the performance of typical classification models. These results demonstrate Q-ROP’s strong potential as a reliable and robust tool for clinical decision support.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103959"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146048255","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Generative data-engine foundation model for universal few-shot 2D vascular image segmentation","authors":"Rongjun Ge ,&nbsp;Xin Li ,&nbsp;Yuxing Liu ,&nbsp;Chengliang Liu ,&nbsp;Pinzheng Zhang ,&nbsp;Jiong Zhang ,&nbsp;Jian Yang ,&nbsp;Jean-Louis Dillenseger ,&nbsp;Chunfeng Yang ,&nbsp;Yuting He ,&nbsp;Yang Chen","doi":"10.1016/j.media.2026.103996","DOIUrl":"10.1016/j.media.2026.103996","url":null,"abstract":"<div><div>The segmentation of 2D vascular structures via deep learning holds significant clinical value but is hindered by the scarcity of annotated data, severely limiting its widespread application. Developing a universal few-shot vascular segmentation model is highly desirable, yet remains challenging due to the need for extensive training and the inherent complexities of vascular imaging. In this work, we propose <strong>UniVG</strong> (Generative Data-engine Foundation Model for Universal Few-shot 2D Vascular Image Segmentation), a novel approach that learns the compositionality of vascular images and constructing a generative foundation model for robust vascular segmentation. UniVG enables the synthesis and learning of diverse and realistic vascular images through two key innovations: <em>1) Compositional learning</em> for flexible and diverse vascular synthesis: It decomposes and recombines vascular structures with varying morphological features and diverse foreground-background configurations to generate richly diverse synthetic image-label pairs. <em>2) Few-shot generative adaptation</em> for transferable segmentation: It fine-tunes pre-trained models with minimal annotated data to bridge the gap between synthetic and real vascular domains, synthesizing authentic and diverse vessel images for downstream few-shot vascular segmentation learning. To support our approach, we develop UniVG-58K, a large dataset comprising 58,689 vascular images across five imaging modalities, facilitating robust large-scale generative pre-training. Extensive experiments on 11 vessel segmentation tasks cross 5 modalties (only with 5 labeled images on each task) demonstrate that UniVG achieves performance comparable to fully supervised models, significantly reducing data collection and annotation costs. All code and datasets will be made publicly available at <span><span>https://github.com/XinAloha/UniVG</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103996"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146209648","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"BCIRT: Backscattering-corrected implicit representation tomography","authors":"Chuanhao Zhang ,&nbsp;Yangxi Li ,&nbsp;Jianping Song ,&nbsp;Yingwei Fan ,&nbsp;Guochen Ning ,&nbsp;Yu Shen ,&nbsp;Canhong Xiang ,&nbsp;Fang Chen ,&nbsp;Hongen Liao","doi":"10.1016/j.media.2026.104000","DOIUrl":"10.1016/j.media.2026.104000","url":null,"abstract":"<div><div>Optical coherence tomography (OCT) A-scan backscattering signals provide depth-resolved textural information about internal structures. However, conventional OCT imaging is limited by refraction-induced distortion and speckle noise, hindering fine detail resolution. While multi-angle imaging systems alleviate these issues through incoherent compounding of backscattering signals, in vivo applications face challenges: limited angular coverage during surface scanning degrades backscatter intensity compounding quality, and the absence of angular information introduces artifacts in multi-view position-intensity alignment. Furthermore, excessive smoothing during speckle suppression obscures fine textures. Consequently, reconstructing ultra-fine structures from limited-angle, sparse-view measurements remains a critical challenge. To address this, we present Backscattering-Corrected Implicit Representation Tomography (BCIRT), a framework for reconstructing multi-angle low-coherence signals. We also develop a dedicated limited-angle imaging system for intraoperative BCIRT deployment. BCIRT formulates cross-view backscattering signals as a continuous function of spatial position, utilizing implicit neural representation (INR) for fitting. A physics-informed iterative mechanism inversely models ray propagation to determine corrected ray paths, enhancing the neural representation’s robustness against distortions. Leveraging these corrected paths, we introduce a dual dynamic line mixer and a contrastive-guided discriminative deblurring module to achieve high-resolution microstructure reconstruction with reduced speckle noise. Extensive experiments on biological samples and surgical resected samples demonstrate that our method achieves state-of-the-art performance, highlighting its potential for clinical applications and biomedical research.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 104000"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146777316","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhancing feature fusion of U-like networks with dynamic skip connections","authors":"Yue Cao,&nbsp;Quansong He,&nbsp;Kaishen Wang,&nbsp;Jianlong Xiong,&nbsp;Zhang Yi,&nbsp;Tao He","doi":"10.1016/j.media.2026.104010","DOIUrl":"10.1016/j.media.2026.104010","url":null,"abstract":"<div><div>U-like networks have become fundamental frameworks in medical image segmentation through skip connections that bridge high-level semantics and low-level spatial details. Despite their success, conventional skip connections exhibit two key limitations: inter-feature constraints and intra-feature constraints. The inter-feature constraint refers to the static nature of feature fusion in traditional skip connections, where information is transmitted along fixed pathways regardless of feature content. The intra-feature constraint arises from the insufficient modeling of multi-scale feature interactions, thereby hindering the effective aggregation of global contextual information. To overcome these limitations, we propose a novel Dynamic Skip Connection (DSC) block that fundamentally enhances cross-layer connectivity through adaptive mechanisms. The DSC block integrates two complementary components: (1) Test-Time Training (TTT) module: This module addresses the inter-feature constraint by enabling dynamic adaptation of hidden representations during inference, facilitating content-aware feature refinement. (2) Dynamic Multi-Scale Kernel (DMSK) module: To mitigate the intra-feature constraint, this module adaptively selects kernel sizes based on global contextual cues, enhancing the network’s capacity for multi-scale feature integration. The DSC block is architecture-agnostic and can be seamlessly incorporated into existing U-like network structures. Extensive experiments demonstrate the plug-and-play effectiveness of the proposed DSC block across CNN-based, Transformer-based, hybrid CNN-Transformer, and Mamba-based U-like networks. The code is available at <span><span>https://github.com/BlackJack-Cao/U-like-Networks-with-DSC</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 104010"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147334541","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"IUGC: A benchmark of landmark detection in end-to-end intrapartum ultrasound biometry","authors":"Jieyun Bai ,&nbsp;Yitong Tang ,&nbsp;Xiao Liu ,&nbsp;Jiale Hu ,&nbsp;Yunda Li ,&nbsp;Xufan Chen ,&nbsp;Yufeng Wang ,&nbsp;Chen Ma ,&nbsp;Yunshu Li ,&nbsp;Bowen Guo ,&nbsp;Jing Jiao ,&nbsp;Yi Huang ,&nbsp;Kun Wang ,&nbsp;Lifei Li ,&nbsp;Yuzhang Ma ,&nbsp;Xiaoxin Han ,&nbsp;Haochen Shao ,&nbsp;Zi Yang ,&nbsp;Qingchen Liu ,&nbsp;Yuchen Hu ,&nbsp;Shuo Li","doi":"10.1016/j.media.2026.103960","DOIUrl":"10.1016/j.media.2026.103960","url":null,"abstract":"<div><div>Accurate intrapartum biometry plays a crucial role in monitoring labor progression and preventing complications. However, its clinical application is limited by challenges such as the difficulty in identifying anatomical landmarks and the variability introduced by operator dependency. To overcome these challenges, the Intrapartum Ultrasound Grand Challenge (IUGC) 2025, in collaboration with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), was organized to accelerate the development of automatic measurement techniques for intrapartum ultrasound analysis. The challenge featured a large-scale, multi-center dataset comprising over 32,000 images from 24 hospitals and research institutes. These images were annotated with key anatomical landmarks of the pubic symphysis (PS) and fetal head (FH), along with the corresponding biometric parameter-the angle of progression (AoP). Ten participating teams proposed a variety of end-to-end and semi-supervised frameworks, incorporating advanced strategies such as foundation model distillation, pseudo-label refinement, anatomical segmentation guidance, and ensemble learning. A comprehensive evaluation revealed that the winning team achieved superior accuracy, with a Mean Radial Error (MRE) of 6.53 ± 4.38 pixels for the right PS landmark, 8.60 ± 5.06 pixels for the left PS landmark, 19.90 ± 17.55 pixels for the FH tangent landmark, and an absolute AoP difference of 3.81 ± 3.12° This top-performing method demonstrated accuracy comparable to expert sonographers, emphasizing the clinical potential of automated intrapartum ultrasound analysis. However, challenges remain, such as the trade-off between accuracy and computational efficiency, the lack of segmentation labels and video data, and the need for extensive multi-center clinical validation. IUGC 2025 thus sets the first benchmark for landmark-based intrapartum biometry estimation and provides an open platform for developing and evaluating real-time, intelligent ultrasound analysis solutions for labor management.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103960"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146033892","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Robust non-rigid image-to-patient registration for contactless dynamic thoracic tumor localization using recursive deformable diffusion models","authors":"Dongyuan Li ,&nbsp;Yixin Shan ,&nbsp;Yuxuan Mao ,&nbsp;Puxun Tu ,&nbsp;Haochen Shi ,&nbsp;Shenghao Huang ,&nbsp;Weiyan Sun ,&nbsp;Chang Chen ,&nbsp;Xiaojun Chen","doi":"10.1016/j.media.2026.103948","DOIUrl":"10.1016/j.media.2026.103948","url":null,"abstract":"<div><div>Deformable image-to-patient registration is essential for surgical navigation and medical imaging, yet real-time computation of spatial transformations across modalities remains a major clinical challenge-often being time-consuming, error-prone, and potentially increasing trauma or radiation exposure. While state-of-the-art methods achieve impressive speed and accuracy on paired medical images, they face notable limitations in cross-modal thoracic applications, where physiological motions such as respiration complicate tumor localization. To address this, we propose a robust, contactless, non-rigid registration framework for dynamic thoracic tumor localization. A highly efficient Recursive Deformable Diffusion Model (RDDM) is trained to reconstruct comprehensive 4DCT sequences from only end-inhalation and end-exhalation scans, capturing respiratory dynamics reflective of the intraoperative state. For real-time patient alignment, we introduce a contactless non-rigid registration algorithm based on GICP, leveraging patient skin surface point clouds captured by stereo RGB-D imaging. By incorporating normal vector and expansion-contraction constraints, the method enhances robustness and avoids local minima. The proposed framework was validated on publicly available datasets and volunteer trials. Quantitative evaluations demonstrated the RDDM’s anatomical fidelity across respiratory phases, achieving an PSNR of 34.01 ± 2.78 dB. Moreover, we have preliminarily developed a 4DCT-based registration and surgical navigation module to support tumor localization and high-precision tracking. Experimental results indicate that the proposed framework preliminarily meets clinical requirements and demonstrates potential for integration into downstream surgical systems.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103948"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145956932","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Generating synthetic MRI scans for improving Alzheimer’s disease diagnosis","authors":"Rosanna Turrisi,&nbsp;Giuseppe Patané","doi":"10.1016/j.media.2026.103947","DOIUrl":"10.1016/j.media.2026.103947","url":null,"abstract":"<div><div>Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia. Magnetic Resonance Imaging (MRI) combined with Machine Learning (ML) enables early diagnosis, but ML models often underperform when trained on small, heterogeneous medical datasets. Transfer Learning (TL) helps mitigate this limitation, yet models pre-trained on 2D natural images still fall short of those trained directly on related 3D MRI data. To address this gap, we introduce an intermediate strategy based on synthetic data generation. Specifically, we propose a conditional Denoising Diffusion Probabilistic Model (DDPM) to synthesise 2D projections (axial, coronal, sagittal) of brain MRI scans across three clinical groups: Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD. A total of 9000 synthetic images are used for pre-training 2D models, which are subsequently extended to 3D via axial, coronal, and sagittal convolutions and fine-tuned on real-world small datasets. Our method achieves 91.3% accuracy in binary (CN vs. AD) and 74.5% in three-class (CN/MCI/AD) classification on the 3T ADNI dataset, outperforming both models trained from scratch and those pre-trained on ImageNet. Our 2D ADnet achieved state-of-the-art performance on OASIS-2 (59.3% accuracy, 57.6% F1), surpassing all competitor models and confirming the robustness of synthetic data pre-training. These results show synthetic diffusion-based pre-training as a promising bridge between natural image TL and medical MRI data.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103947"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146032814","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Diversity-driven MG-MAE: Multi-granularity representation learning for non-salient object segmentation","authors":"Chengjin Yu ,&nbsp;Bin Zhang ,&nbsp;Chenchu Xu ,&nbsp;Dongsheng Ruan ,&nbsp;Rui Wang ,&nbsp;Huafeng Liu ,&nbsp;Xiaohu Li ,&nbsp;Shuo Li","doi":"10.1016/j.media.2026.103971","DOIUrl":"10.1016/j.media.2026.103971","url":null,"abstract":"<div><div>Masked Autoencoders (MAEs) have grown increasingly prominent as a powerful self-supervised learning paradigm. They are capable of effectively leveraging inherent image prior information and are gaining traction in the field of medical image analysis. However, their application to feature representations of the non-salient objects, such as microvasculature, accessory organs, and early-stage tumors–is fundamentally limited by dimensional collapse problem, which diminishes feature diversity critical for non-salient structure discrimination. To address this, we propose a Multi-Granularity Masked Autoencoder (MG-MAE) framework for feature diversity learning: (1) We extend the conventional MAE into a multi-granularity framework, a global branch reconstructs global pixels, with a local branch recovering Histogram of Oriented Gradients (HOG) features, enabling hierarchical representation of both coarse-grained and fine-grained patterns; (2) Critically, in the local branch, a diversity-enhanced loss function incorporating Nuclear Norm Maximization (NNM) constraint to explicitly mitigate feature space collapse through orthogonal embedding regularization; and (3) A Dynamic Weight Adjustment (DWA) strategy that dynamically prioritizes hard-to-reconstruct regions via entropy-driven gradient modulation. Comprehensive evaluations across five clinical benchmarks–CCTA139, BTCV, LiTS, ACDC, and MSD Pancreas Tumour datasets–demonstrate that MG-MAE achieves statistically significant improvements in Dice Similarity Coefficient (DSC) scores for non-salient object segmentation, outperforming state-of-the-art methods. The code is available at <span><span>https://github.com/zhangbbin/mgmae</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103971"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146134049","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DSFNet: Dual-source and spatiotemporal-feature fusion network for bedside diagnosis of lung injuries with electrical impedance tomography","authors":"Zhiwei Li ,&nbsp;Yang Wu ,&nbsp;Kai Liu ,&nbsp;Yingqi Zhang ,&nbsp;Bai Chen ,&nbsp;Hao Wang ,&nbsp;Jiafeng Yao","doi":"10.1016/j.media.2026.104003","DOIUrl":"10.1016/j.media.2026.104003","url":null,"abstract":"<div><div>Electrical Impedance Tomography (EIT) is a promising tool for non-invasive and real-time lung monitoring, but the data heterogeneity and low spatial resolution limit its ability to diagnose lung injuries. To address these challenges, we propose DSFNet, a dual-source and spatiotemporal-feature fusion network that integrates EIT spatiotemporal boundary voltages and ventilation images to classify four lung conditions, including healthy (HE), pneumothorax (PN), pleural effusion (PE), and pneumonia (PM). The temporal dynamics modeling (TDM) module and multi-head self-attention (MHSA) module are designed to improve the temporal feature extraction and representation of DSFNet. We construct a novel EIT simulation dataset describing pathological respiratory patterns and introduce a hybrid data learning strategy that combines simulation data (SD) and experimental data (ED) to address the small sample problem and improve the accuracy of model classification. The DSFNet trained with the SD + 25 % ED pattern achieved an accuracy of 97.78 % and 96.55 % on the dynamic phantom dataset and the clinical human dataset, respectively, demonstrating its excellent performance and robustness. The SHAP analysis further revealed the feature contributions of the input data. This study provides an effective approach for bedside lung injury diagnosis based on multi-source EIT data.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 104003"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146777317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multimodal sparse fusion transformer network with spatio-temporal decoupling for breast tumor classification","authors":"Jiahao Xu ,&nbsp;Shuxin Zhuang ,&nbsp;Yi He ,&nbsp;Haolin Wang ,&nbsp;Zhemin Zhuang ,&nbsp;Huancheng Zeng","doi":"10.1016/j.media.2026.103966","DOIUrl":"10.1016/j.media.2026.103966","url":null,"abstract":"<div><div>Accurate analysis of tumor morphology, vascularity, and tissue stiffness under multimodal ultrasound imaging plays a critical role in the diagnosis of breast cancer. However, manual interpretation across multiple modalities is time-consuming and heavily dependent on the radiologist’s expertise. Computer-aided classification offers an efficient alternative, yet remains challenging due to significant modality heterogeneity, inconsistent image quality, and redundant information across modalities. To address these issues, we propose a novel Multimodal Sparse Fusion Transformer Network (MSFT-Net). First, a Spatio-Temporal Decoupling Attention architecture (STDA) is introduced to disentangle and extract dynamic and static features from different modalities along spatial and temporal dimensions, capturing modality-specific motion and morphological characteristics independently. Second, the Mixed-Scale Convolution Module (MSCM) obtains tumor features at multiple scales, enhancing geometric detail representation and improving receptive field coverage. Third, the Sparse Cross-Attention Module (SCAM) adaptively retains the most effective query-key interactions between modalities, thereby facilitating the aggregation of high-quality features for robust multimodal information fusion. MSFT-Net is trained and tested on a curated dataset comprising multimodal breast tumor videos collected from 458 patients, including ultrasound (US), superb microvascular imaging (SMI), and strain elastography (SE), and its generalizability is further validated on the public BraTS'21 MRI dataset. Extensive experiments demonstrate that MSFT-Net achieves superior performance in multimodal breast tumor classification compared to state-of-the-art methods, providing fast and reliable support for radiologists in diagnostic tasks.</div></div>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"110 ","pages":"Article 103966"},"PeriodicalIF":11.8,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146072192","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}