Cape Town - 2026 ISMRM-ISMRT Annual Meeting and Exhibition

AI Frontiers in Image Analysis

Flash Presentation

Analysis Methods

Monday, 11 May 2026

Roof Terrace

13:50 - 15:06

Moderators: Valiyeh Ansarian Nezhad & SAGNIK GHOSH

Session Number: 331-02

No CME/CE Credit

Similar Sessions

This session showcases modern AI approaches for MRI analysis tasks, ranging from generative approaches to foundation models.

Skill Level: Basic,Intermediate,Advanced

13:50		331-02-001. BridgeMamba: Frequency–Spatial Bridging for Undersampled MRI Segmentation Hongli Chen, Pengcheng Fang, 雨夏陈, Tianyi Ding, Wei Jiang, Zhifeng Chen, Chunyi Liu, Yang Gao, Fangfang Tang, Feng Liu, Shanshan Shan The University of Queensland, Brisbane, Australia Impact: We propose BridgeMamba, a dual-stream Mamba model with a standard spatial branch and a frequency-aware branch. This study enables the reliable analysis of undersampled MRI by segmenting directly without reconstruction, improving clinical workflow efficiency.
13:52		331-02-002. Evaluating DeepSeek-OCR Embeddings on Single-Slice MRI: From 0.064 T to 3 T Long Wang, Zhihao Zhang, Zechen Zhou, Ajit Shankaranarayanan Subtle Medical Inc, Menlo Park, United States of America Impact: Consistent embeddings across 64 mT and 3 T can facilitate MRI harmonization, automated quality control, and information compression, improving reliability and efficiency across field strengths and imaging sites.
13:54		331-02-003. Personalized Specific Absorption Rate Prediction in Ultra-High Field MRI Based on Cycle-Consistent Generative Adversarial Net Yizhi Cui, Shao Che, zhuoxv cui, Shahzeb Hayat, Nan Li, Enhua Xiao, Peiyu He, Dong Liang, Ye Li Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China Impact: This research supports dynamic, personalized optimization of MRI scan parameters. This enhances safety and image quality in high-SAR risk scenarios (e.g., ultra-high field, pediatric imaging), unlocking advanced sequence potential.
13:56		331-02-004. Set deep learning for protocol generalisation in machine-learning-based brain microstructure estimation Leevi Kerkelä, Antoine Legouhy, Nina Kraguljac, Gary Zhang University College London, London, United Kingdom Impact: We show that set-deep-learning-based parameter estimation generalises across acquisition protocols, enabling fast, single-model microstructure mapping without retraining.
13:58		331-02-005. Coarse-to-Fine Meta-Reweighting with Dynamic Retrieval for Pediatric Brain Tumor Segmentation Abdulkhalek Al-Fakih, Kanghyun Ryu, Mohammed Al-masni Sejong University, Seoul, Korea, Republic of Impact: This framework leverages adult MRI data without compromising pediatric specificity, offering a scalable strategy for accurate tumor segmentation in rare or underrepresented patient populations.
14:00		331-02-006. AlignPET: Structure-Aligned MRI-to-PET Synthesis via Variational Autoregression Model for Ischemic Brain Lesions Yifei Chen, Guanyu Zhou, Yuanhan Wang, Mingxuan Liu, Xuguang Bai, Jialan Zheng, Bixiao Cui, Jie Lu, Qiyuan Tian Tsinghua University, Beijing, China Impact: AlignPET demonstrates a transformative approach to functional brain imaging by generating PET-equivalent maps directly from MRI. It enables low-cost, radiation-free metabolic visualization, offering potential to democratize access to functional neuroimaging for both clinical and research applications.
14:02		331-02-007. Generative and Graph-Based Modelling of 4D Flow MRI for Quantitative Hemodynamic Recovery Assessment in Stroke Himanshu Singh, Chirag Sharma, Yug Kaushik, Sparsh Singh, Vishnu V. Y., S Senthil Kumaran All India Institute of Medical Sciences, New Delhi, India Impact: This work establishes a computational bridge between cardiovascular and neurological recovery, using 4D Flow as a predictive biomarker tool. Integrating generative and graph-based modelling, with interpretable vascular phenotyping that could guide personalized stroke rehabilitation and future neurovascular outcome prediction frameworks
14:04		331-02-008. Physics-Informed Neural Networks Improve Time-Encoded Pseudo-Continuous Arterial Spin Labeling Perfusion Quantification Alessandro Giupponi, Chiara Da Villa, Giulio Ferrazzi, Francesca Benedetta Pizzini, Matthias van Osch, Mattia Veronese, Marco Castellaro University of Padova, Padova, Italy Impact: Physics-Informed Neural Networks enable accurate, data-driven quantification of cerebral perfusion from arterial spin labelling. By embedding the Buxton model within the network, this approach improves noise robustness and physiological consistency compared to conventional Bayesian estimation, advancing quantitative brain imaging reliability.
14:06		331-02-009. UnA2LGENet: A Generalizable SAM-Adaptor for Multicenter LGE of Myocardial Infarction Across 1000+ patients Xiuzheng Yue, Jing Qi, Miao Hu, Yinyin Chen, Hang Jin, Tao Li, Kunlun He Chinese PLA General Hospital, Beijing, China Impact: By coupling SAM proposals with adaptive refinement and on-the-fly domain adaptation, UnA2LGENet delivers fast, reproducible multi-center LGE quantification—unlocking robust infarct metrics at scale for trials, registries, and outcome modeling.
14:08		331-02-010. A Unified AI Tool for Clinical Brain MRI Super-Resolution, Outpainting, Skull-Stripping, and Segmentation Across the Lifespan Zehua Ren, Yuzhu He, Xinmei Qiu, Kehan Li, Haifeng Wang, Ruyi Xiao, Chunfeng Lian, Jianhua Ma, Fan Wang Xi'an Jiaotong University, Xi'an, China Impact: Our Unified AI tool takes low-resolution clinical scan as input with potential incomplete FOV and outputs high-resolution, skull-stripped, and outpainted brain image, alongside an anatomically consistent tissue segmentation. This streamlined and unified method empowers researchers and clinicians in their work.
14:10		331-02-011. Augmentrum: A Data Augmentation Package for MR Spectroscopy John LaMaster, Julian Merkofer, Kay Igwe Technical University of Munich (TUM), Munich, Germany Impact: Augmentrum is a data augmentation package that maximizes the utility of limited in-vivo data for deep learning applications while preserving physically plausible MRS variability. This package reduces the required amount of training data while improving model performance and generalizability.
14:12		331-02-012. Reliability of an Automated Quantification Tool for Brain Sagging Signs in Spontaneous Intracranial Hypotension Pei-Yun Wu, Yen-Feng Wang, Po-Hsun Su, Shuu-Jiun Wang, Chia-Feng Lu National Yang Ming Chiao Tung University, Taipei, Taiwan Impact: Our automated quantification tool of brain sagging signs may transform subjective spontaneous intracranial hypotension assessment into an objective measurement, improving diagnostic consistency and workflow. It establishes a foundation for automated scoring, reporting, treatment monitoring, and follow-up to enhance patient management.
14:14		331-02-013. Application of Multi-Teacher Distillation for Enhancing the Effectiveness of Medical Image Segmentation via Foundation Models Qing Li, Yizhe Zhang, Ying-Hua Chu, Mo Yang, Haoyang Zhang, Junhong Liu, Wang Liao, Shuo Wang, Chengyan Wang Human Phenome Institute, Fudan University, Shanghai, China Impact: This study integrates the concept of multi-teacher distillation for the first time and preliminary attempt to enhance the performance of SFMs, potentially reducing costs of time, resources. More importantly, it provides a paradigm shift for future applications of SFMs.
14:16		331-02-014. Zero-Shot Physics-Informed Neural Networks for Robust Multi-Vendor DSC-MRI Perfusion Quantification in Glioblastoma Patients Ziyu Fu, Puneet Kumar, Mahsa Servati, Chinmay Mokashi, Natenael Semmineh, Nazanin Majd, Vinaykumar Puduvalli, C. Chad Quarles MD Anderson Cancer Center, Houston, United States of America Impact: This work introduces a physics-driven, pretraining-free tissue residue function estimation strategy that stabilizes DSC-MRI deconvolution across vendors and patient heterogeneity, representing a key step toward robust, generalizable perfusion quantification in neuro-oncology.
14:18		331-02-015. Etiological Classification of White Matter Lesions Using Foundation Models: A Human-AI Collaboration Exploration Shan Lv, Zebin Gao, Yuerong Lizhu, Ao Feng, Cong Zhang, Siyao Xu, Tianyu Gao, Feng Xu, Yaou Liu Beijing Tiantan Hospital, Capital Medical University, Beijing, China Impact: This study introduces a multimodal foundation model that dynamically integrates MRI data to accurately distinguish white matter lesion etiologies, offering a powerful tool for early differential diagnosis and personalized treatment planning in diverse white matter disorders.
14:20		331-02-016. Do We Need Massive Pretraining? A Comparative Study of Foundation and Transformer Models for Abdominal MRI Segmentation Sandeep Kaushik, Seyed Iman Zare Estakhraji, Dattesh Dayanand Shanbhag, Zhijian Yang, Noel DSouza, Chitresh Bhushan, Gurunath Reddy Madhumani, Vanika Singhal, Erhan Bas, Marc Lebel GE HealthCare, Waukesha, United States of America Impact: This study questions the necessity of massive pretraining for clinical MRI segmentation, showing comparable performance with limited data. It enables cost‑efficient workflows and motivates research into minimal data strategies for scalable, practical AI in medical imaging.