Tong Wang, Yue Sun, Limei Wang, Gang Li, Weili Lin, Li Wang
{"title":"LifespanFix: Semi-supervised cortical correction across the human lifespan.","authors":"Tong Wang, Yue Sun, Limei Wang, Gang Li, Weili Lin, Li Wang","doi":"10.1016/j.media.2026.104389","DOIUrl":"https://doi.org/10.1016/j.media.2026.104389","url":null,"abstract":"<p><p>Accurate characterization of the cerebral cortex in brain magnetic resonance imaging (MRI) is essential for reliable morphometric analyses across the human lifespan. However, the cortex's thin and highly convoluted architecture, together with imaging artifacts, frequently introduces topological and anatomical defects in segmentation that can bias downstream analyses. Here we present LifespanFix, a lifespan-robust cortical correction model that directly fixes topological and anatomical defects in segmentation. Specifically, we develop a semi-supervised framework in which a generation network produces template-constrained pseudo-labels that serve as structurally plausible targets for training a correction network (LifespanFix). During inference, only the trained LifespanFix model is deployed, enabling efficient end-to-end correction of cortical defects without requiring the generation network or the time-intensive registration used during training. Our framework was trained on 3,189 subjects spanning the human lifespan - from infancy to older adulthood - across nine public datasets. For evaluation, we assembled an independent lifespan testing cohort comprising 1,058 subjects from five datasets, including label maps generated by different segmentation methods. LifespanFix consistently outperforms state-of-the-art correction approaches, producing more accurate and anatomically plausible cortical segmentations across diverse ages, scanners, and imaging protocols, thereby enabling more reliable large-scale lifespan neuroimaging analyses.</p>","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"116 ","pages":"104389"},"PeriodicalIF":14.0,"publicationDate":"2026-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149408088","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}
Shijia Zhou, Le Cao, Lei Mou, Hang Wang, Jinkui Hao, Xuening Zhang, Hongshuo Li, Leilei Yuan, Weicheng Liao, Thomas H. Julian, Meng Wang, Yonghuai Liu, William Robert Kwapong, Yanda Meng, Jianyang Xie, Yalin Zheng, Huazhu Fu, Jiang Liu, Bin Sheng, Alejandro F. Frangi, Bo Wu, Yitian Zhao
{"title":"Memory-guided mutual information optimization in retinal imaging for early stroke detection and subtyping","authors":"Shijia Zhou, Le Cao, Lei Mou, Hang Wang, Jinkui Hao, Xuening Zhang, Hongshuo Li, Leilei Yuan, Weicheng Liao, Thomas H. Julian, Meng Wang, Yonghuai Liu, William Robert Kwapong, Yanda Meng, Jianyang Xie, Yalin Zheng, Huazhu Fu, Jiang Liu, Bin Sheng, Alejandro F. Frangi, Bo Wu, Yitian Zhao","doi":"10.1016/j.media.2026.104382","DOIUrl":"https://doi.org/10.1016/j.media.2026.104382","url":null,"abstract":"Effective stroke prevention relies on identifying at-risk individuals before neurological symptoms arise, yet this is severely limited by the absence of scalable screening tools. To address this, we developed RetCEVD, a diagnostic framework that uses rapid, non-invasive retinal Optical Coherence Tomography Angiography (OCTA) for auxiliary assessment of ICVD. At its core, RetCEVD introduces a memory-guided mutual information optimization strategy that maximizes shared disease-relevant information within the same diagnostic category while suppressing redundant or confounding dependencies across different categories. This dual information constraint encourages compact, discriminative representations of subtle retinal microvascular abnormalities. Trained and validated on a multi-center cohort of 24,438 retinal images, RetCEVD addresses three primary tasks: auxiliary detection of existing ICVD, etiological subtyping (large artery atherosclerosis vs. small vessel disease), and current symptomatic-status classification (asymptomatic vs. symptomatic LAA). The model achieved AUCs of 0.95 internal and 0.94 external for ICVD detection, 0.83 internal and 0.81 external for etiological subtyping, and 0.75 internal and 0.76 external for current symptomatic-status classification. The two-year ALAA follow-up provided supplementary longitudinal validation for subsequent stroke events (AUC, 0.71). By comprehensively evaluating the model’s clinical interpretability, including its learned pathological features and alignment with established clinical patterns, we provide novel insights into retina-based stroke evaluation. Ultimately, RetCEVD may serve as an efficient, low-cost screening and triage tool to identify individuals who may benefit from confirmatory neurovascular imaging and specialist assessment.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"14 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389944","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}
Biwen Meng, Qian Da, Xi Long, Jiahao Wang, Ruochen Liu, Frauke Wilm, Katharina Breininger, Yuqi Li, Chuanguang Yang, Jiayan Huang, Shuoyu Xu, Huajia Wang, Xing Lu, Pengzhou Cai, Ze Zhao, Shashank Gupta, Nameeta Shah, Adrian Galdran, Prasad Dutande, Suhas Gajre, Ho Heon Kim, Young Shin Ko, Abdul Qayyum, Steven A Niederer, Yinglan Kuang, Jimin Xiao, Angelos Stefanidis, Shuihua Wang, Guangliang Cheng, Yanfei Zuo, Yuexiang Li, Chaofu Wang, Linlin Shen, Jingxin Liu
{"title":"The COSAS 2024 challenge: Cross-Organ and Cross-Scanner Adenocarcinoma Segmentation for domain generalization in computational pathology","authors":"Biwen Meng, Qian Da, Xi Long, Jiahao Wang, Ruochen Liu, Frauke Wilm, Katharina Breininger, Yuqi Li, Chuanguang Yang, Jiayan Huang, Shuoyu Xu, Huajia Wang, Xing Lu, Pengzhou Cai, Ze Zhao, Shashank Gupta, Nameeta Shah, Adrian Galdran, Prasad Dutande, Suhas Gajre, Ho Heon Kim, Young Shin Ko, Abdul Qayyum, Steven A Niederer, Yinglan Kuang, Jimin Xiao, Angelos Stefanidis, Shuihua Wang, Guangliang Cheng, Yanfei Zuo, Yuexiang Li, Chaofu Wang, Linlin Shen, Jingxin Liu","doi":"10.1016/j.media.2026.104390","DOIUrl":"https://doi.org/10.1016/j.media.2026.104390","url":null,"abstract":"Histopathological assessment remains the gold standard for cancer diagnosis. However, automated histopathology models can be sensitive to domain shifts arising from organ-specific morphological heterogeneity and scanner-associated acquisition variability. To investigate these complementary sources of variation, we established the Cross-Organ and Cross-Scanner Adenocarcinoma Segmentation (COSAS) 2024 Challenge, comprising separate cross-organ and cross-scanner segmentation tasks. The benchmark includes 580 pixel-level annotated histopathology patches spanning six organ sites and six scanner platforms. Thirty international teams submitted valid solutions using a range of convolutional, transformer-based, and hybrid architectures together with different training, augmentation, and inference strategies. Separately, we evaluated conventional segmentation networks and foundation-model-based baselines as organizer-run references. Performance varied substantially across domains and between the two tasks. The relative strengths of the reference models were task dependent, with no model family uniformly strongest across both settings, while the leading challenge submissions achieved competitive performance through different combinations of augmentation, multi-domain training, model design, and inference-time refinement. Overall, COSAS provides a controlled benchmark for studying cross-domain adenocarcinoma segmentation and offers practical insights into the design of AI systems that can better handle the biological and technical variability encountered in diagnostic workflows.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"45 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389943","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":"Hierarchical causal construction for semantic-granularity-aligned counterfactual medical image generation","authors":"Zhizhong Fu, Wei Zhou, Zhaoyang Jiang, Yudong Lin, Gaoming Li, Yuke Cao, Yifu Hou, Xiaorong Ding, Qiang Yan, Yifan Chen","doi":"10.1016/j.media.2026.104391","DOIUrl":"https://doi.org/10.1016/j.media.2026.104391","url":null,"abstract":"Counterfactual image generation seeks to answer how a medical image should change when a clinical semantic factor is altered, and is important for model interpretation, lesion localization, and controllable image editing. This work addresses the semantic granularity mismatch in existing counterfactual generation methods and formulates it as the problem of semantic-granularity alignment (SGA). Specifically, counterfactual queries typically target macro-level semantics, such as lesion morphology, tissue intensity, or risk phenotype, whereas generative models operate on micro-level visual units, including pixels, patches, textures, and local boundaries. Without an explicit correspondence between these two levels, macro-level semantic interventions cannot be reliably translated into changes in the relevant local structures. To address this issue, we propose hierarchical causal construction (HCC), which uses causal abstraction to establish a hierarchical mapping between micro-level visual units and macro-level clinical semantics. We further integrate HCC with causal representation learning (CRL) to develop HCC-CRL, which first encodes an image into micro-level causal nodes, and then aggregates them into macro-level causal nodes and semantic slots through a learnable soft quotient mapping. Counterfactual interventions are subsequently performed at the macro level, and the intervened states are projected back into the micro-level visual space for counterfactual image generation. Within a unified hierarchical causal representation, the framework jointly models target semantic changes, preservation of non-target information, and local visual consistency. Experiments on Morpho-MNIST, the Breast Ultrasound Images (BUSI) dataset, and a microvascular invasion (MVI) prediction task show that HCC-CRL generates counterfactual images with well-directed semantic changes and improved structural preservation, while maintaining a robust balance among intervention success, identity preservation, and image quality.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"282 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389946","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}
Huangxin Xu, Qizhe Liu, Jue Hou, Yi Ding, Long Lei, Ying Hu, Zenan Wang
{"title":"Tuda: Tucker-parameterized unified dual-stream asymmetric architecture for label-free 3D subcellular structure prediction","authors":"Huangxin Xu, Qizhe Liu, Jue Hou, Yi Ding, Long Lei, Ying Hu, Zenan Wang","doi":"10.1016/j.media.2026.104384","DOIUrl":"https://doi.org/10.1016/j.media.2026.104384","url":null,"abstract":"Label-free prediction of three-dimensional subcellular fluorescence from transmitted-light microscopy enables scalable and less invasive live-cell analysis. However, building a unified subcellular structure prediction (SSP) model remains difficult: each training volume is only partially labeled, target organelles exhibit highly heterogeneous morphologies, and many structures require long-range topological coherence beyond local appearance. We propose TUDA, a unified dual-stream framework in which all components are conditioned on a shared learnable task embedding, enabling organelle-specific specialization throughout the network. The local stream uses a Tucker decomposition-based hypernetwork to synthesize organelle-specific convolutional kernels from the task embedding, achieving compact yet diverse feature extraction within a single shared model. Complementing this, the global stream introduces Flash-TAS (Flash Transformer with Topology-Aware Shift), a Transformer branch that captures long-range structural dependencies while preserving voxel-level spatial correspondence. A Fusion Bridge adaptively integrates both streams for voxel-wise fluorescence regression. On a public benchmark covering 12 subcellular structures, TUDA achieves state-of-the-art SSP performance, improving relative <mml:math altimg=\"si156.svg\" display=\"inline\"><mml:msup><mml:mrow><mml:mi>R</mml:mi></mml:mrow><mml:mrow><mml:mn>2</mml:mn></mml:mrow></mml:msup></mml:math> over RepMode by 3.9% while using 75% fewer parameters, with notable gains on topologically complex targets such as actin filaments and tight junctions.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"14 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389947","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":"MoCE-GAN: Mixture of Complexity Experts with generative adversarial networks for image restoration in sparse-view and limited-view photoacoustic tomography","authors":"Zhihao Chen, Ting Feng, Dean Ta","doi":"10.1016/j.media.2026.104388","DOIUrl":"https://doi.org/10.1016/j.media.2026.104388","url":null,"abstract":"Photoacoustic tomography (PAT) is a hybrid imaging modality that combines optical contrast with ultrasonic penetration depth for biomedical applications. Sparse-view and limited-view sampling strategies reduce acquisition time and system complexity but introduce reconstruction artifacts that degrade image quality and clinical utility. Although deep learning methods have advanced sparse-view PAT restoration, most rely on uniform architectures that allocate a fixed computational budget to every input sample, leaving them poorly matched to the varying degradation severity across samples. We propose MoCE-GAN, which couples a Mixture-of-Complexity-Experts (MoCE) generator—selecting experts according to sample-level degradation complexity via global feature statistics and a frequency embedding—with a PatchGAN discriminator that enforces local perceptual fidelity. Experiments on embryo numerical simulation, bone numerical simulation, <ce:italic>in vivo</ce:italic> mouse, and <ce:italic>in vivo</ce:italic> human datasets evaluate sparse-view settings with 16, 32, or 64 transducers and limited-view settings with 45°, 90°, or 180° angular coverage, depending on the acquisition geometry. Compared with four representative image-domain methods (PAT-ADN, FD-UNet, SA-UNet, GAID-PAT), MoCE-GAN raises PSNR over the best baseline by an average of 1.77 dB across the 16 dataset–configuration combinations under configuration-specific training, and by 2.10 dB under mixed training, where a single model covers all configurations and yields PSNR 0.24 dB lower on average than the configuration-specific models. SSIM and LPIPS are each best in most configurations. Beyond quantitative gains, restorations remain anatomically consistent under severe undersampling, supporting accelerated or reduced-view PAT acquisition scenarios.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"64 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389960","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}
Zhicheng He, Xiaohan Xing, Hongqiu Wang, Kun Yuan, Guankun Wang, Yanheng Li, Haoxuan Che, Nassir Navab, Hongliang Ren, Long Bai
{"title":"SurgPace: Complexity-aware progressive distillation for surgical multimodal large language models","authors":"Zhicheng He, Xiaohan Xing, Hongqiu Wang, Kun Yuan, Guankun Wang, Yanheng Li, Haoxuan Che, Nassir Navab, Hongliang Ren, Long Bai","doi":"10.1016/j.media.2026.104386","DOIUrl":"https://doi.org/10.1016/j.media.2026.104386","url":null,"abstract":"Multimodal large language models (MLLMs) have shown strong potential in robotic surgery for question answering, scene understanding, and grounded assistance, but their high computational cost limits real clinical deployment. Knowledge distillation provides a practical way to build compact surgical MLLMs, yet direct distillation is not well suited to surgical data because surgical samples vary greatly in difficulty, visual ambiguity, and required clinical context, making uniform teacher-student transfer unstable. In addition, the role of supervision changes during training: early stages require stronger teacher guidance for multimodal alignment, while later stages rely more on task supervision for effective optimization. To address these challenges, we propose SurgPace, a complexity-aware continuous curriculum framework for progressive distillation of surgical vision-language models. SurgPace estimates sample complexity using predictive confidence and Lorentz geometric consistency, and uses the fused score to guide a continuous easy-to-hard training process. We further introduce hierarchical representation alignment, which aligns the student with the teacher at the feature, logit, and probability levels, while gradually shifting the training focus from teacher imitation to task supervision. This design separates sample-level curriculum from objective-level scheduling, leading to more stable and effective distillation. We evaluate SurgPace on Surg-396 K, a benchmark built from four public surgical datasets and covering five dialogue paradigms. SurgPace consistently outperforms strong MLLM distillation baselines, demonstrating its effectiveness for building compact, memory-efficient surgical MLLMs with faster decoding under controlled benchmarks.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"16 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389945","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}
Hanna Hoffmann, Setareh Bady, Claas de Boer, Max Kirchner, Jan Egger, Rainer Röhrig, Frank Hölzle, Lennart Johannes Gruber, Kunpeng Xie, Marlon Neuhaus, Victor Alves, Guilherme Barbosa, Leonardo Barroso, João Carvalho, Hao Chen, Gabriella d’Albenzio, André Ferreira, Nuno Gomes, Yuichiro Hayashi, Kousuke Hirasawa, Rebecca Hisey, Seungjae Hong, Seoi Jeong, Tiago Jesus, Daehong Kang, Satoshi Kasai, Shunsuke Kikuchi, Takayuki Kitasaka, Satoshi Kondo, Hyoun-Joong Kong, Youngbin Kong, Atsushi Kouno, Shlomi Laufer, Kyu Eun Lee, Bining Long, Nooshin Maghsoodi, Hiroki Matsuzaki, Evangelos Mazomenos, Ori Meiraz, Kensaku Mori, Marina Music, Masahiro Oda, Roi Papo, Jieun Park, Rafael Piexoto, Saeid Rezaei, Mariana Ribeiro, Soyeon Shin, Yang Shu, Idan Smoller, Danail Stoyanov, Yihui Wang, Xinkai Zhao, Sebastian Bodenstedt, Isabel Funke, Stefanie Speidel, Behrus Hinrichs-Puladi
{"title":"OSS: Open suturing skills vision-based assessment challenge 2024–2025","authors":"Hanna Hoffmann, Setareh Bady, Claas de Boer, Max Kirchner, Jan Egger, Rainer Röhrig, Frank Hölzle, Lennart Johannes Gruber, Kunpeng Xie, Marlon Neuhaus, Victor Alves, Guilherme Barbosa, Leonardo Barroso, João Carvalho, Hao Chen, Gabriella d’Albenzio, André Ferreira, Nuno Gomes, Yuichiro Hayashi, Kousuke Hirasawa, Rebecca Hisey, Seungjae Hong, Seoi Jeong, Tiago Jesus, Daehong Kang, Satoshi Kasai, Shunsuke Kikuchi, Takayuki Kitasaka, Satoshi Kondo, Hyoun-Joong Kong, Youngbin Kong, Atsushi Kouno, Shlomi Laufer, Kyu Eun Lee, Bining Long, Nooshin Maghsoodi, Hiroki Matsuzaki, Evangelos Mazomenos, Ori Meiraz, Kensaku Mori, Marina Music, Masahiro Oda, Roi Papo, Jieun Park, Rafael Piexoto, Saeid Rezaei, Mariana Ribeiro, Soyeon Shin, Yang Shu, Idan Smoller, Danail Stoyanov, Yihui Wang, Xinkai Zhao, Sebastian Bodenstedt, Isabel Funke, Stefanie Speidel, Behrus Hinrichs-Puladi","doi":"10.1016/j.media.2026.104359","DOIUrl":"https://doi.org/10.1016/j.media.2026.104359","url":null,"abstract":"Achieving high levels of surgical skill through effective training is essential for optimal patient outcomes. Automated, data-driven skill assessment holds significant potential to improve surgical training. While machine learning-based methods are increasingly popular for assessing skills in minimally invasive surgery, their application to open surgery remains limited. We present the results of a dedicated MICCAI challenge designed to benchmark and advance vision-based skill assessment in open surgery.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"30 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389961","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}
Xiaoqian Chen, Minxiao Wang, Vanessa L. Wildman, Ryan James Sanford, Lei “Richard” Qiu, Aparna H. Kesarwala, Xiaofeng Yang
{"title":"Joint-Embedding Self-Supervised Representation Learning for Robust Cone-Beam CT Segmentation","authors":"Xiaoqian Chen, Minxiao Wang, Vanessa L. Wildman, Ryan James Sanford, Lei “Richard” Qiu, Aparna H. Kesarwala, Xiaofeng Yang","doi":"10.1016/j.media.2026.104393","DOIUrl":"https://doi.org/10.1016/j.media.2026.104393","url":null,"abstract":"Cone-beam CT (CBCT) segmentation is challenged by reconstruction artifacts and limited reliable annotations. We investigate joint-embedding predictive architecture (JEPA)-style self-supervised learning for volumetric CBCT segmentation using CBCT-LGR, which extends our preliminary slice-wise conference study through fully three-dimensional local-global encoding, learnable semantic tokens, and a combined objective comprising local-global alignment, semantic alignment, and predictive consistency. We compare CBCT-LGR with supervised baselines (nnU-Net and STU-Net), generic self-supervised methods (BYOL-Med3D and Models Genesis), and foundation models (SAM, MedSAM, and Anatomix) using harmonized patient-level splits, preprocessing, and evaluation metrics while retaining each method’s stated supervision setting. For representation-learning methods, frozen encoders are evaluated using supervised decoders trained with the same 10-subject labeled-data budget per dataset. Across three datasets, CBCT-LGR outperforms the generic self-supervised baselines on dataset-average overlap and boundary metrics. On CBCTLiTS, supervised and fine-tuned foundation-model approaches achieve higher absolute DSC under clean reconstruction conditions. Across five controlled reconstruction-quality levels, however, CBCT-LGR exhibits the smallest maximum relative DSC decline among all nine methods, with declines of 1.5% for liver and 14.5% for tumor, and achieves the highest numerical tumor DSC at the two most degraded levels. Results on Dataset 1 quantify agreement with imperfect reference annotations and do not establish true anatomical accuracy or a definitive method ranking. These findings define the capabilities and limitations of JEPA-style pretraining for volumetric CBCT segmentation.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"12 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149389948","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}
Hyungjin Chung, Dohun Lee, Zihui Wu, Byung-Hoon Kim, Katherine L. Bouman, Jong Chul Ye
{"title":"ContextMRI: Enhancing compressed sensing MRI through metadata conditioning","authors":"Hyungjin Chung, Dohun Lee, Zihui Wu, Byung-Hoon Kim, Katherine L. Bouman, Jong Chul Ye","doi":"10.1016/j.media.2026.104353","DOIUrl":"https://doi.org/10.1016/j.media.2026.104353","url":null,"abstract":"Compressed sensing MRI seeks to accelerate MRI acquisition processes by sampling fewer <mml:math altimg=\"si1.svg\" display=\"inline\"><mml:mi>k</mml:mi></mml:math>-space measurements and then reconstructing the missing data algorithmically. The success of these approaches often relies on strong priors or learned statistical models. While recent diffusion model-based priors have shown great potential, previous methods typically ignore clinically available metadata (e.g. patient demographics, imaging parameters, slice-specific information). In practice, metadata contains meaningful cues about the anatomy and acquisition protocol, suggesting it could further constrain the reconstruction problem. In this work, we propose ContextMRI, a text-conditioned diffusion model for MRI that integrates granular metadata into the reconstruction process. We train a pixel-space diffusion model directly on minimally processed, complex-valued MRI images. During inference, metadata is converted into a structured text prompt and fed to the model via CLIP text embeddings. By conditioning the prior on metadata, we unlock more accurate reconstructions and show consistent gains across multiple datasets, acceleration factors, and undersampling patterns. Our experiments demonstrate that increasing the fidelity of metadata—ranging from slice location and contrast to patient age, sex, and pathology—systematically boosts reconstruction performance. This work highlights the untapped potential of leveraging clinical context for inverse problems and opens a new direction for metadata-driven MRI reconstruction. Code is available at <ce:inter-ref xlink:href=\"https://github.com/DoHunLee1/ContextMRI\" xlink:type=\"simple\">https://github.com/DoHunLee1/ContextMRI</ce:inter-ref>.","PeriodicalId":18328,"journal":{"name":"Medical image analysis","volume":"21 1","pages":""},"PeriodicalIF":10.9,"publicationDate":"2026-11-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149358735","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}