Cape Town - 2026 ISMRM-ISMRT Annual Meeting and Exhibition

Foundation Models

Oral

Analysis Methods

Tuesday, 12 May 2026

Hall 1A

08:20 - 10:10

Moderators: Onat Dalmaz & Aya Ghoul

Session Number: 401-02

No CME/CE Credit

Similar Sessions

This session will focus on the development, validation, and benchmarking of foundation models applied to a diverse array of MRI tasks.

08:20		401-02-001. TumorCLIP: Lightweight Vision–Language Fusion for Explainable MRI-Based Brain Tumor Classification Yishaoying Jia, Jinfu Niu, Zongyu Li, Jia Guo Xi'an Jiaotong-Liverpool University, China Impact: This study demonstrates that lightweight vision-language fusion enhances MRI-based brain tumor classification, offering an interpretable and efficient diagnostic framework that improves subtype recognition and may facilitate broader clinical adoption of explainable AI in medical imaging.
08:31		401-02-002. A Unified Vision-Language Foundation Model for Multi-Task MRI Application Xingxin He, Aurora Rofena, Yifan Hu, Ruimin Feng, Zhehao Liao, Valerio Guarrasi, Paolo Soda, Zhaoye Zhou, Albert Jang, Fang Liu Harvard Medical School, Boston, United States of America Impact: OmniMRI is a unified vision-language foundation model trained on large-scale heterogeneous MRI data, performing reconstruction, segmentation, detection, diagnosis, and report generation in one system to enhance automation, efficiency, and generalization across diverse protocols, anatomies, and tasks in the MRI workflow.
08:42		401-02-003. BrainMR Specialist: A Foundation Model of Brain MRI for Diverse Downstream Applications Juhyung Park, Minjun Kim, Rokgi Hong, Jaehyeon Koo, Roh-Eul Yoo, Seung Hong Choi, Jongho Lee Seoul National University, Seoul, Korea, Republic of Impact: Our brain-specialized foundation model provides a single, data-efficient, and scalable deep learning model for diverse clinical and research applications, reducing the need for task-specific models, especially in limited available data.
08:53		401-02-004. BrainDFMAE: A Unified Foundation Model for Aging-Brain sMRI via Deformation-Aware Pretraining Xinmei Qiu, Kehan Li, Yuzhu He, Zehua Ren, Fan Wang, Jianhua Ma, Chunfeng Lian Xi'an Jiaotong University, Xi'an, China Impact: BrainDFMAE establishes a unified foundation model for aging-brain sMRI through deformation-aware pretraining, significantly enhancing capabilities in precise brain mapping, early diagnosis, and longitudinal progression prediction. This provides a powerful tool for advancing personalized treatment strategies and improving patient outcomes.
09:04		401-02-005. Site effects persist in MRI foundation models: insights from BrainIAC embeddings Rafael Navarro, Santiago Aja-Fernández, Ángel Guerrero, Rodrigo de Luis García Hospital Clínico Universitario de Valladolid, Valladolid, Spain Impact: Foundation-model representations carry residual site information. Aiming for invariance at training, with intensity or contrast augmentations and domain-aligned objectives, may strengthen generalization across scanners and acquisition settings, which is critical for real-world multi-centre deployments.
09:15		401-02-006. Foundation model for cardiac perfusion MRI enables 10-fold reduction in labeled dataset size for deep-learning analysis M. Berk Sahin, Zhuoan Li, Khalid Youssef, Arian Sohi, Dilek Yalcinkaya, Luis Zamudio, Michael Elliott, Venkateshwar Polsani, Matthew Tong, Dipan Shah, Orlando Simonetti, Abolfazl Hashemi, Behzad Sharif Purdue University, West Lafayette, United States of America Impact: We propose the largest-scale cardiac perfusion MRI foundation model trained on > 600,000 unlabeled multi-center images, achieving state-of-the-art performance for automatic segmentation with over 10-fold fewer manual labels, reducing reliance on manual annotation, and providing a reusable model for other tasks.
09:26		401-02-007. KIMRA: K-space–Image Multimodal Representation Alignment for Comprehensive Cardiac Analysis Yundi Zhang, Sevgi Gokce Kafali, Daniel Rueckert, Jiazhen Pan Technical University of Munich and TUM University Hospital, Munich, Germany Impact: This work establishes a scalable foundation for cardiac screening by deriving comprehensive representations directly from undersampled k-space, enabling efficient cardiac function assessment and advancing cardiovascular research through the rich physiological information preserved in the raw acquisition domain.
09:37		401-02-008. A Vision-Language Foundation Model for Automated Segmentation of Cardiac Contours in Cine MRI Mingzhen Li, Hanyu Su, Jiaqi Guo, Lexiaozi Fan, Neda Tavakoli, Santiago López-Tapia, Aggelos Katsaggelos, Daniel Kim Northwestern University, Chicago, United States of America Impact: CardiVLSM bridges vision-language reasoning with foundation segmentation, enabling fully automated, prompt-free segmentation of cine MRI. Its strong generalizability across datasets indicates a path toward scalable, clinically deployable AI tools for cardiac function assessment without requiring manual intervention.
09:48		401-02-009. A Foundation Model-Driven Framework for Automated QA of Medical Imaging AI Solutions Jayesh Tripathi, Gurunath Reddy Madhumani, Sajith Rajamani, Vanika Singhal, Bhavatharani Sundararajan, Garvita Basantani, Sneha Sree C, Alvin Ronnie, Tisha Abraham, Chitresh Bhushan, Dattesh Dayanand Shanbhag GE HealthCare, Bengaluru, India Impact: The proposed template matching framework enables scalable, automated quality assessment of AI-based images, reducing manual review burden and supporting clinical deployment. It provides infrastructure for continuous monitoring and regulatory compliance, with broad applicability across imaging tasks requiring similarity based validation.
09:59		401-02-010. Self-Supervised Representation Learning of Brain MRI Using DINOv3: From Anatomical Features to Cross-Domain Generalization Wei Jiang, Tianyi Ding, Hongli Chen, Wenjia Song, Zhuang Xiong, Yang Gao, Nan Ye, Feng Liu, Hongfu Sun The University of Queensland, Brisbane, Australia Impact: This study shows that self-supervised vision transformers can learn anatomical features from unlabeled MRI, enabling label-efficient and domain-robust medical analysis, highlighting their potential to bridge the gap between large-scale unlabeled imaging datasets and practical clinical applications.