Artificial Intelligence for Medical Imaging & MRI

Magnetic resonance imaging (MRI) is a cornerstone of modern medical diagnosis, yet its broader clinical impact remains constrained by fundamental limitations—long acquisition times, high operational costs, and a strong dependence on specialized infrastructure. Addressing these challenges requires not only advances in hardware, but also computational paradigms that can extract more information from less data while maintaining clinical reliability. Our research follows a coherent trajectory toward this goal, evolving from physics-informed reconstruction to generalizable AI systems capable of operating under real-world constraints.

Accelerated MRI Reconstruction: From Data Scarcity to Reliable Imaging: ^{1, 2, 3} We begin by addressing one of the most immediate bottlenecks in MRI: slow acquisition and incomplete data. Accelerated MRI relies on undersampling in k-space, which significantly reduces scan time but introduces artifacts and information loss. To overcome this, we develop reconstruction methods that integrate deep learning with MRI physics, including hybrid architectures that jointly model k-space and image-domain information, as well as Bayesian approaches that leverage statistical priors to guide reconstruction. Our work demonstrates that high-quality images can be recovered even under aggressive undersampling while preserving structures critical for diagnosis. Importantly, we also investigate robustness under degraded conditions, particularly for clinically relevant tasks such as stroke detection. By incorporating distributionally robust learning strategies, our models maintain reliable diagnostic performance even when image quality is compromised, underscoring the importance of coupling reconstruction with downstream clinical objectives rather than treating them as independent problems.

Data-Efficient and Transferable Learning: ^{4, 5} Building on these reconstruction advances, we turn to a broader challenge: limited and heterogeneous medical data. Many clinical scenarios lack sufficient labeled datasets, and imaging conditions vary widely across scanners, institutions, and patient populations. To address this, we develop data-efficient learning frameworks that reduce reliance on large, task-specific annotations. Diffusion-based reconstruction models enable high-quality image recovery through novel degradation processes in k-space, while neural style transfer methods enable field-strength translation, generating high-quality images from low-field inputs without requiring paired training data. These approaches emphasize transferability and flexibility, demonstrating that models can generalize across imaging domains, anatomical regions, and acquisition conditions. This reflects a shift from optimizing for individual datasets to designing systems that remain robust under real-world variability.

Toward Unified and Generalizable Models: ⁶ A central question that emerges from this progression is whether MRI models can move beyond task-specific designs toward unified architectures. To this end, we introduce transformer-based frameworks such as the Magnetic Resonance Image Processing Transformer (MR-IPT), pretrained on diverse MRI datasets and applied across multiple reconstruction settings. By leveraging a shared backbone, these models learn universal feature representations that generalize across acceleration factors, sampling patterns, and perturbations. This marks a conceptual transition from task-specific learning to representation learning, where a single model can adapt to multiple scenarios without retraining from scratch. The result is enhanced robustness, scalability, and efficiency compared to conventional deep learning approaches.

From Reconstruction to Multi-Task Medical AI: ⁷ While reconstruction addresses image quality, clinical practice requires models that can also interpret and analyze images across tasks. A key limitation of earlier approaches is that generalization is often confined to a single task domain. To overcome this, our recent work explores whether pretrained models can extend across multiple downstream medical imaging tasks, even under limited labeled data. Using large-scale self-supervised pretraining on over 31 million MRI slices, we develop a unified transformer framework that supports MRI sequence classification, skull stripping (brain extraction), and multi-region anatomical segmentation. These models achieve strong performance in few-shot settings, demonstrating that large-scale pretraining can significantly reduce the need for extensive annotations. This progression reflects a broader shift—from reconstructing images to understanding them, and ultimately to enabling flexible, task-agnostic AI systems.

Integration of Physics, Learning, and Clinical Validation: Across all stages of this work, a unifying theme is the integration of physical modeling, statistical reasoning, and modern machine learning. Rather than treating MRI as a purely data-driven problem, we embed domain knowledge to ensure that models remain interpretable, stable, and firmly grounded in imaging principles. Equally important is clinical validation: our work is evaluated not only through quantitative metrics, but also in clinically relevant settings, including stroke diagnosis and cross-dataset generalization. Through this integration, methodological advances are translated into robust, clinically meaningful improvements, bridging the gap between algorithmic innovation and real-world medical practice.

Outlook: Toward Scalable and Practical Medical AI: The work highlighted here traces a clear progression from accelerating MRI acquisition to developing generalizable reconstruction models, and ultimately to enabling scalable, deployable AI systems for medical imaging. Looking ahead, we seek to advance these frameworks toward multimodal imaging that integrates MRI with complementary modalities such as CT and PET; toward disease-focused applications including dementia and neurodegenerative disorders; and toward clinical translation through sustained collaboration with radiologists and healthcare systems. By uniting advances in machine learning with domain knowledge and clinical insight, our vision is to bridge the gap between AI innovation and medical practice—enabling imaging systems that are not only reliable and scalable, but also broadly accessible, adaptive, and impactful across diverse clinical settings.