Computerized Medical Imaging and Graphics最新文献

{"title":"Self-supervised isotropic reconstruction for abnormality detection in anisotropic MRI","authors":"Jui-Yo Hsu,&nbsp;Pin-Hsun Lian,&nbsp;Tzu-Yi Chuang,&nbsp;Yi-Hwa Chen,&nbsp;Gary Han Chang","doi":"10.1016/j.compmedimag.2026.102786","DOIUrl":"10.1016/j.compmedimag.2026.102786","url":null,"abstract":"<div><div>Accelerating musculoskeletal magnetic resonance imaging (MRI) while preserving diagnostic detail remains challenging because acquiring fully‑isotropic ground‑truth volumes is clinically costly. In routine practice, anisotropic scans with reduced through-plane resolution degrade multiplanar visualization and slice-by-slice review in reformatted planes, obscure subtle abnormalities spanning only a few slices, and limit automated three-dimensional (3D) analyses that assume comparable spatial resolution across axes. We present a two‑stage, fully self‑supervised pipeline that learns directly from anisotropic scans—obviating any paired high‑resolution data—and converts highly anisotropic (8:1) turbo‑spin‑echo volumes into isotropic images and 3D abnormality maps. Unlike prior self-supervised super-resolution methods, Stage 1 uses a single forward multi-view generative adversarial network (GAN) with patch-based contrastive and adversarial objectives rather than a backward/cycle-consistency approach. Stage 2 leverages an anatomy-conditioned denoising-diffusion model for healthy counterfactual generation, yielding voxel-wise lesion maps without external annotations.</div><div>On 2225 Osteoarthritis Initiative knee scans from five different imaging centres, the framework reduced Fréchet inception distance from 407.4 → 254.4 (coronal) and 429.9 → 266.9 (axial), achieved the best Kernel Inception Distance (KID) / Learned Perceptual Image Patch Similarity (LPIPS) scores among competing unsupervised methods, and was preferred in 65–67% of blinded orthopedist comparisons. Crucially, isotropic enhancement propagated to downstream tasks: femur–tibia segmentation F1 scores increased and previously confluent bone‑marrow lesions were separated into discrete entities, enabling precise volumetric quantification. Robustness experiments demonstrated consistent gains across five imaging centers, synthetic noise/contrast perturbations, and transfer of the resolution-enhancement module to two additional MRI protocols, supporting robustness across sites and acquisition protocols.</div><div>By eliminating the need for ground‑truth isotropic images while surpassing state‑of‑the‑art unsupervised super‑resolution in both perceptual quality and clinical utility, our method may facilitate retrospective cohort studies and prospective scan-time reduction in heterogeneous knee MRI settings, with preliminary transferability to additional protocols.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"133 ","pages":"Article 102786"},"PeriodicalIF":5.5,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148363894","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Anatomically and biochemically guided deep image prior for sodium MRI denoising","authors":"Haider Ali,&nbsp;Ramona Woitek,&nbsp;Siegfried Trattnig,&nbsp;Olgica Zaric","doi":"10.1016/j.compmedimag.2026.102791","DOIUrl":"10.1016/j.compmedimag.2026.102791","url":null,"abstract":"<div><div>Sodium (²³Na) magnetic resonance imaging (MRI) provides valuable metabolic information but is fundamentally limited by low signal-to-noise ratio (SNR) and long acquisition times. While classical and learning-based denoising methods such as NLM, BM3D, TV, and RD-DIP can produce visually plausible results, they may alter the underlying sodium signal distribution, potentially affecting structural fidelity and quantitative consistency. To address these challenges, we propose DIP-Fusion, a Deep Image Prior (DIP)-based framework for sodium MRI denoising that exploits complementary anatomical information from proton (¹H) MRI and biochemical information from ²³Na MRI via a novel fused proton–sodium prior within a directional total variation (dTV) regularization scheme. The framework optimizes a variational loss combining data fidelity, fused dTV regularization, gradient consistency, and bias-field correction to ensure stable denoising. We evaluate DIP-Fusion on healthy volunteers and breast cancer patients under multiple acquisition settings, comparing against classical denoising methods (NLM, BM3D, TV) and a recent deep learning-based Rician denoising approach (RD-DIP). Results demonstrate consistent improvements over all baselines. In healthy subjects, DIP-Fusion achieves PSNR gains of up to +2.36 dB and SSIM improvements of approximately 5%, with reduced perceptual error. In patient datasets, it further improves image quality, achieving PSNR gains of up to +3.35 dB while preserving sodium-specific signal characteristics. Under varying Rician noise levels, DIP-Fusion shows improved robustness compared to RD-DIP, maintaining higher denoising accuracy and structural fidelity, particularly in high-noise regimes. Overall, the proposed fusion-based prior stabilizes DIP optimization and enables more reliable sodium MRI denoising under challenging noise conditions.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"133 ","pages":"Article 102791"},"PeriodicalIF":5.5,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148297086","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Continual test-time adaptation via weight averaging of feature augmentations in cross-domain medical image segmentation","authors":"Kyuri Kim,&nbsp;Hyungjoo Cho,&nbsp;Yoonho Na,&nbsp;Jimin Lee,&nbsp;Sung-Joon Ye","doi":"10.1016/j.compmedimag.2026.102793","DOIUrl":"10.1016/j.compmedimag.2026.102793","url":null,"abstract":"<div><div>Continual Test-Time Adaptation (CTTA) effectively mitigates domain shifts that occur across medical institutions by allowing models to adapt at test time without revisiting the source data. Existing CTTA methods exploit transform invariance through pixel-level augmentation, but these approaches tend to produce inconsistent results in online environments, particularly critical in semantic segmentation tasks. To address these issues, we propose FAWA, a CTTA framework that combines feature-based augmentation and a weighted averaging strategy. FAWA augments features by retrieving statistical information from a continuously updated memory bank, which generates diverse perturbations of input target samples. Each transformed sample is independently fine-tuned, and the resulting parameter sets are merged through continuous weighted averaging. This process leads the model to converge to a stable minimum and mitigates the accumulated error. Experiments on Brain CT and Prostate MRI benchmarks indicate that FAWA achieves competitive performance while ensuring robustness during extended adaptation streams.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"133 ","pages":"Article 102793"},"PeriodicalIF":5.5,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148400059","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A lightweight network for segmenting tree-like structures in medical images","authors":"Irfan Dwiki Bhaswara,&nbsp;Sonit Singh,&nbsp;Arcot Sowmya,&nbsp;Erik Meijering","doi":"10.1016/j.compmedimag.2026.102796","DOIUrl":"10.1016/j.compmedimag.2026.102796","url":null,"abstract":"<div><div>Segmentation of medical images containing tree-like structures is critical for disease diagnosis, simulation modelling and surgical planning. Although many deep learning methods have achieved impressive performance in this domain, challenges remain due to the elongated and complex geometry of such structures, which are prone to disconnection during segmentation. Most existing methods rely on convolutional neural networks, which struggle to capture long-range dependencies inherent in tree-like structures. Transformer-based approaches offer improvements in this respect but suffer from quadratic computational complexity as input size increases. Recently, Mamba has emerged as a promising alternative, offering comparable performance to Transformers with linear computational complexity, and has started gaining attention in medical image segmentation. However, prior work has mainly focused on organ or lesion segmentation. In this work, we propose CoorMa, a lightweight U-Net-shaped model that integrates Coordinate Attention and Mamba Aggregation Modules for the segmentation of tree-like structures in medical images. Coordinate Attention enhances directional feature extraction, while the Mamba Aggregation Modules fuse features across scales to better capture long-range dependencies. Additionally, we introduce a skeleton-based artefact removal algorithm to eliminate spurious segmentation artefacts while preserving structural integrity. We evaluate our method on four 2D datasets (DCA1, CHUAC, DRIVE, CHASE_DB1) and one 3D dataset (Binary Airway Segmentation). Experimental results show that CoorMa outperforms state-of-the-art methods, achieving the highest evaluation metrics with significantly fewer parameters.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"133 ","pages":"Article 102796"},"PeriodicalIF":5.5,"publicationDate":"2026-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148400032","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Artificial intelligence empowered coronary artery imaging: A review","authors":"Mingzhi Wang ,&nbsp;Junxin Chen ,&nbsp;Hao Li ,&nbsp;Li-bo Zhang ,&nbsp;Wei Wang","doi":"10.1016/j.compmedimag.2026.102782","DOIUrl":"10.1016/j.compmedimag.2026.102782","url":null,"abstract":"<div><div>Cardiovascular disease is the leading cause of death worldwide, with coronary artery disease the most prevalent cause. Although artificial intelligence has advanced medical imaging, few reviews focus specifically on AI-powered coronary artery imaging. This review provides a systematic and critical analysis of AI applications in coronary artery imaging across three domains: measurement, including centerline extraction, vessel segmentation, and 3D reconstruction; functional assessment, including fractional flow reserve, wall shear stress, and myocardial perfusion; and disease diagnosis. Web of Science, Google Scholar, and MEDLINE were searched for 2016–2025. From 9950 records, 90 studies were selected after duplicate removal, title and abstract screening, and full-text eligibility assessment using predefined criteria and quality appraisal. Seven imaging modalities are covered: computed tomography, magnetic resonance imaging, nuclear imaging, digital subtraction angiography, intravascular ultrasound, optical coherence tomography, and synthetic data. Reported performance varies by modality, task, and dataset. Challenges for clinical implementation include dataset generalizability, computational requirements, interoperability, and explainable AI. The review discusses emerging technologies such as multimodal fusion, self-supervised learning, federated learning, and foundation models, and acknowledges current limitations. It aims to inform researchers and clinicians and support future translation of AI in coronary artery imaging.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102782"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148151267","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An interpretable multimodal model integrating clinical, spectral CT imaging, and deep learning analysis of intra- and peritumoral regions for preoperative prediction of perineural invasion in gastric cancer: A prospective, multicenter study","authors":"Yuncong Zhao ,&nbsp;Lihong Chen ,&nbsp;Lili Wang ,&nbsp;Wenjing Yue ,&nbsp;Deye Zeng ,&nbsp;Yaru You ,&nbsp;Caiming Weng ,&nbsp;Weixing Li ,&nbsp;Yilin Chen ,&nbsp;Lin Lin ,&nbsp;Haiping Xu ,&nbsp;Yunjing Xue","doi":"10.1016/j.compmedimag.2026.102776","DOIUrl":"10.1016/j.compmedimag.2026.102776","url":null,"abstract":"<div><h3>Objective</h3><div>Perineural invasion (PNI) is crucial for risk stratification and treatment planning. This study aims to develop a novel multimodal model for preoperative prediction of PNI in gastric cancer.</div></div><div><h3>Methods</h3><div>A total of 250 patients with pathologically confirmed gastric cancer from multiple centers were enrolled and divided into a training cohort (n = 138), an internal validation cohort (n = 59), and an external validation cohort (n = 53). Spectral CT parameters from both intra- and peritumoral regions were acquired using the GSI viewer software, while deep learning features were extracted via a ResNet-50 architecture. Clinical features were screened through univariate and multivariate regression analyses. Based on machine learning, we developed a spectral parameter model, a deep learning feature model, a clinical prediction model, and a multimodal fusion model (MFM). The predictive performance of the models was evaluated using calibration curves, receiver operating characteristic curves, and decision curve analysis (DCA). Model interpretability was achieved by employing Shapley Additive exPlanations (SHAP).</div></div><div><h3>Results</h3><div>Compared with individual models, MFM demonstrated the best predictive performance, with area under the curve (AUC) values of 0.926, 0.885, and 0.871 in the training, internal validation, and external validation cohorts, respectively. DCA confirmed that MFM provided a higher net benefit over a wide range of threshold probability. Calibration curve showed MFM had better predictive consistency, and the Hosmer–Lemeshow test indicated good fit across all cohorts (all p-values &gt; 0.05). SHAP effectively interpreted the decision-making process of the model.</div></div><div><h3>Conclusion</h3><div>The novel multimodal fusion model, which integrates clinical, spectral CT parameters, and deep learning features from intra- and peritumoral regions, demonstrated outstanding performance in preoperatively predicting PNI in gastric cancer.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102776"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148058124","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Artificial intelligence for real-time motion tracking in MRI-guided radiotherapy: A systematic review","authors":"Shengqi Chen ,&nbsp;Zilin Wang ,&nbsp;Jianrong Dai ,&nbsp;Guohua Wu ,&nbsp;Jiayun Chen","doi":"10.1016/j.compmedimag.2026.102783","DOIUrl":"10.1016/j.compmedimag.2026.102783","url":null,"abstract":"<div><h3>Background and Purpose</h3><div>Intrafraction motion compromises accurate dose delivery in MRI-guided radiotherapy (MRIgRT), motivating the adoption of artificial intelligence (AI). This systematic review aims to evaluate the performance of AI-driven motion tracking approaches and identify barriers to clinical translation.</div></div><div><h3>Methods</h3><div>PubMed and Web of Science were searched from January 1, 2020, to January 1, 2026. Studies were selected via a two-stage screening process based on predefined criteria. Key information was extracted, and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) was used to assess methodological quality.</div></div><div><h3>Results</h3><div>Twenty-eight studies were included. Nineteen focused on target localization through registration-, segmentation-, reconstruction-, or machine learning-based approaches, and nine addressed motion prediction. Reported performance generally indicated sub-2 mm localization or prediction errors in many settings, segmentation accuracy of at least 0.83 in 12 studies, and inference times ranging from &lt; 0.1–420 ms, although one volumetric reconstruction study required 8000 ms. Current research exhibits a pronounced anatomical bias toward the liver (n = 23). Only two studies used established public datasets, five reported external validation, and six released source code. The mean CLAIM score was 24.00 ± 2.85.</div></div><div><h3>Conclusion</h3><div>AI-driven motion tracking has substantially advanced real-time motion management in MRIgRT, but clinical translation remains limited by anatomical bias, small and heterogeneous datasets, scarce external validation, heterogeneous evaluation metrics, and insufficient deployment and prospective dosimetric validation. Future progress will require multi-center open datasets, anatomy-aware model design, workflow-level optimization, and prospective clinically oriented validation.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102783"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148151418","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"RFD: A Reducing Feature Discrepancy method for unsupervised cross-modality SAM adaptation","authors":"Ji Xia ,&nbsp;Zhehan Shen ,&nbsp;Wei Xia ,&nbsp;Yulin Pu ,&nbsp;Xueyang Zou ,&nbsp;Weiming Liu ,&nbsp;Dian Ding ,&nbsp;Chengyan Wang ,&nbsp;Guangtao Xue ,&nbsp;Shunjie Dong ,&nbsp;Ruokun Li","doi":"10.1016/j.compmedimag.2026.102781","DOIUrl":"10.1016/j.compmedimag.2026.102781","url":null,"abstract":"<div><div>Vision foundation models, such as Segment Anything Model (SAM), have achieved great success in medical image segmentation tasks under full supervision, but require accurate masks. However, SAM suffers from significant performance degradation when applied to unsupervised cross-modality medical image segmentation. To maintain satisfactory performance and reduce the masking cost, we propose a Reducing Feature Discrepancy (RFD) method for SAM in cross-modality medical image segmentation, including Reweighting Unbalanced Feature Adaptation (RUFA) and Structural Prototype-based Contrastive Learning (SPCL). Specifically, we integrate medical adapters into the SAM encoder to improve its feature capture capabilities. SPCL introduces the structural prototypes with the structural information distance instead of simple node embeddings, improving feature discrimination in latent spaces and achieving accurate pixel-prototype transport. We also propose Inexact Prototype-based Pixel Transport (IPPT) in SPCL for the dynamic selection and allocation of prototypes, leading to efficient and robust online clustering. Furthermore, RUFA distinguishes image samples utilizing features from the SAM latent space, contingent upon their similarity or dissimilarity, thereby promoting inter-class discrimination and intra-class compactness with small uncertainty. Our experimental results demonstrate a significant improvement over recent state-of-the-art unsupervised domain adaptation (UDA) works and SAM-based methods on four public datasets, validating its effectiveness and generalization.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102781"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148158869","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A longitudinally-consistent deep framework for joint subcortical segmentation and registration of infant brains","authors":"Liangjun Chen ,&nbsp;Zhengwang Wu ,&nbsp;Fenqiang Zhao ,&nbsp;Ya Wang ,&nbsp;Weili Lin ,&nbsp;Li Wang ,&nbsp;Gang Li","doi":"10.1016/j.compmedimag.2026.102779","DOIUrl":"10.1016/j.compmedimag.2026.102779","url":null,"abstract":"<div><div>Subcortical segmentation of infant brain MR images is fundamental for studying neurodevelopment but remains challenging due to rapid brain development, low tissue contrast, and ambiguous anatomical boundaries. Conventional age-specific methods typically segment longitudinal scans independently, which ignores crucial temporal guidance and frequently leads to longitudinally inconsistent results. To address these issues, we propose C2FSRnet, a longitudinally-consistent coarse-to-fine network for joint subcortical segmentation and registration. Inspired by the reciprocal benefits between these tasks, we develop a framework to simultaneously perform affine/deformable registration and segmentation of longitudinal scans. In the coarse stage, the registration and segmentation branches share a joint encoder, enabling them to assist each other in capturing more informative and generalized features. Specifically, we jointly input multiple longitudinal scans from each subject to train the framework to learn within-subject anatomical correspondences, providing explicit longitudinal guidance and ensuring consistency. Additionally, signed distance maps are incorporated as spatial contextual guidance into the fine-stage network in a multi-path manner to achieve refined segmentation and registration. We validated C2FSRnet on the UNC/UMN Baby Connectome Project (393 scans from 124 subjects) and the developing Human Connectome Project (20 scans from 10 subjects) datasets. Compared to 15 state-of-the-art methods, our framework consistently achieves superior accuracy and longitudinal consistency in both tasks, demonstrating its robust performance for characterizing early subcortical development in both healthy and clinical populations.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102779"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148058167","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"CGHNet: Cross-Guided 2D–3D Hybrid Network with attention mechanism for focal liver lesion classification","authors":"Zhuoao Li ,&nbsp;Saya Sailike ,&nbsp;Yuzhe Li ,&nbsp;Xiaogang Zhang ,&nbsp;Xuequn Shang ,&nbsp;Bolin Chen","doi":"10.1016/j.compmedimag.2026.102780","DOIUrl":"10.1016/j.compmedimag.2026.102780","url":null,"abstract":"<div><div>Accurate differential diagnosis of focal liver lesions (FLLs) is pivotal for clinical treatment planning and prognosis. Although multi-sequence MRI serves as the primary imaging modality, precise characterization remains challenging due to complex lesion heterogeneity and subtle inter-class similarities. Existing computational models face an inherent dilemma: slice-based 2D models lack Z-axis spatial continuity, whereas volume-based 3D models often compromise intricate intra-slice semantics to capture global structures. To overcome these limitations, we propose the Cross-Guided 2D–3D Hybrid Network (CGHNet). CGHNet effectively synergizes these complementary paradigms by employing a 2D Vision Transformer (ViT) to model long-range intra-slice dependencies and a 3D CNN to preserve holistic volumetric contexts. To resolve the severe semantic misalignment between these heterogeneous representations, we introduce a Cross-Guided Fusion Mechanism (CGFM). This module employs a reciprocal gating strategy that utilizes macroscopic global contexts from one branch to dynamically suppress background noise and recalibrate feature responses in the other. Furthermore, an Adaptive Decision Fusion (ADF) module is designed to function as an instance-aware arbiter, selectively weighting branch-specific predictions via residual correction to robustly rectify diagnostic uncertainties. Extensive experiments on the LLD-MMRI2023 dataset demonstrate that CGHNet achieves competitive performance, yielding an accuracy of 82.7% and an AUC of 96.5%. These results compare favorably with reproduced representative baselines under the same experimental setting, highlighting the potential of CGHNet as an assistive tool for focal liver lesion classification.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"132 ","pages":"Article 102780"},"PeriodicalIF":4.9,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148018593","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}