Hayit Greenspan, MICCAI Fellow, 2024

Hayit Greenspan was named a MICCAI Fellow in 2024 in recognition of her outstanding contributions to machine learning methods in medical image computing and her tireless support of the MICCAI Society. She has served in MICCAI conference leadership roles, including Program Chair for MICCAI 2023, made significant editorial contributions to the MICCAI book series, and participated in conference workshop committees. She is also a member of the MICCAI 2027 (Auckland, New Zealand) Organizing Committee, serving as the Keynote Chair.

Hayit Greenspan is a Professor of Biomedical Engineering and heads the Medical Image Processing Lab in the School of Biomedical Engineering at Tel-Aviv University. In 2021, she joined the Icahn School of Medicine at Mount Sinai, NY, as Co-Director of the AI and Emerging Technologies (AIET) PhD concentration in the Graduate School of Biomedical Sciences at the Icahn School of Medicine at Mount Sinai in New York. Dr. Greenspan is also a Co-Founder and Chief Scientist of RADLogics Inc., a company that focuses on bringing newly developed AI image analysis tools to radiologists and oncologists for clinical use.   

Dr. Greenspan received the B.S. and M.S. degrees in Electrical Engineering from the Technion and her Ph.D. from CALTECH. She was a Postdoc with the CS Division at U.C. Berkeley, following which she joined Tel-Aviv University.  From 2008 until 2011, she was a visiting Professor at Stanford University, in the Department of Radiology, Faculty of Medicine.  She was also a visiting researcher at IBM Research in the Multi-modal Mining for Healthcare group, in Almaden CA.

Dr. Greenspan has over 250 publications in leading international journals and conferences and has received several awards and patents. She is a member of several journals and conference program committees, including SPIE medical imaging, IEEE_ISBI, and MICCAI.  She served as an Associate Editor for the IEEE Transactions on Medical Imaging (TMI).  In 2016 she was the Lead Co-editor for a Special issue on Deep Learning in Medical Imaging in IEEE TMI. In 2017 she co-edited an Elsevier Academic Press book on Deep learning for Medical Image Analysis (sold 1000 copies:” Most successful Title Published” in 2018). Dr. Greenspan is ranked among the top 1-2% of cited researchers in medical imaging and biomedical engineering (Stanford University ranking). She was named among the “Top 30 Women AI leaders in Drug Discovery and Advanced Healthcare,” by Deep Knowledge Analytics.

In the following Q&A, Dr. Greenspan generously shares her insights on generative AI in medical imaging, its clinical applications, and the impact the MICCAI community has had on her career.

Q: When did you first become involved with the MICCAI Society?

A: My journey with MICCAI began around 2001. At the time, I was transitioning from generic computer vision into the medical domain, and MICCAI felt like the definitive home for that intersection. I remember being struck by the community’s unique ability to blend rigorous mathematical theory with clinical application. I spent those early years absorbing the specialized knowledge found in MICCAI’s workshops and building a network of colleagues who have since become lifelong collaborators. Over the years, I’ve moved from an inspired attendee to a contributor, and eventually into leadership roles like Program Chair. It’s been a privilege to grow alongside this community and help shape the very forums that once mentored me.

Q: Being named a MICCAI Fellow in 2024 is a significant recognition. What has the MICCAI community meant to you throughout your career?

A: The MICCAI community has been my primary academic home and my most significant partnership. Early in my career, it was the benchmark against which I challenged myself and my research; today, it is the venue I look forward to yearly and to which I submit my group’s most impactful work. Beyond the publications, MICCAI represents a unique convergence of methodological rigor and clinical relevance—a shared mission I hold with my closest colleagues. Having served in roles ranging from reviewer to Program Chair, I’ve seen firsthand how this community fosters innovation.

Being named a Fellow feels like a perfect full-circle moment. It is a privilege to be recognized by a community that has given me so much and that has been a key part of my career. To be named a MICCAI Fellow is not just a personal honor; it’s a reflection of a career-long commitment to a field that continues to redefine the future of healthcare.

Q: What first sparked your interest in applying deep learning to medical image analysis? Was there a defining moment that shaped your direction?

A: My research has always lived at the intersection of pattern recognition and medical image analysis. However, the true 'inflection point' occurred between 2013 and 2015. At that time, our community was deeply skeptical about deep learning; the prevailing wisdom was that we simply didn't have enough annotated medical imaging data to make neural networks-based analysis viable.

In 2015, my group was among the first to explore what we now call transfer learning—leveraging ImageNet pre-trained CNNs for medical tasks. I remember the 'Eureka' moment: we saw a nearly 10% leap in performance across classification tasks. That shift from skepticism to empirical success was transformative. The community was able to explore fine-tuning and, eventually, data augmentation schemes (such as Generative Adversarial Networks (GANs)) to overcome the data scarcity that had held the field back. It has been incredible to watch a domain that was once met with doubt become the standard for our entire society.

Q: You were an early leader in deep learning for medical imaging. When you look at the field today, what excites you most — and what still needs work?

A: What excites me most today are a few domains:  First, the shift toward Foundation Models as the next generation of transfer learning. We are moving from millions of generic images to specialized models that can be adapted with just a few clinical annotations—making few-shot and zero-shot learning a reality for clinically-motivated downstream tasks.

I am also deeply invested in the Generative AI space, specifically in cross-modality information transfer. I see immense potential in using generative models to extract 'hidden' insights from accessible, affordable modalities like X-rays, achieving the diagnostic power typically reserved for CT or MRI. I call this 'Enhanced Early Detection'—using AI to find signals earlier than ever before. Finally, Multimodal Fusion is a new frontier.  By merging images with text (clinical notes and reports) and 1D signals, among other data, we can aim to create a truly holistic representation of the patient, allowing us to move from simple diagnosis to complex prognostic modeling and personalized treatment planning.

What still needs work: We need better robustness to domain shifts and more sophisticated continual learning frameworks. Furthermore, we need to move beyond generic accuracy and develop task-specific performance metrics that truly respect the clinical context and the stakes of the medical domain.

Q: You trained in electrical engineering and later moved into radiology and medical imaging. How did that interdisciplinary transition influence your research perspective?

A: My transition from Electrical Engineering to Radiology-focused research changed my perspective from being model-driven to being problem-driven. In EE, we are trained to build powerful mathematical tools; in focusing on medically motivated tasks, I learned to understand the clinical workflow and the 'language' of the domain.

I have found that while many groups can build sophisticated models, the real breakthrough often lies in the problem definition. By understanding the medical context, I can identify the subtle nuances—the true challenges and workflow bottlenecks—that others might miss. This perspective allows for a much more fruitful collaboration with clinicians, ensuring that the unique models we develop aren't just technically impressive but are relevant and deployable in real-world healthcare settings.

Q: What do you see as the biggest gap between AI research in medical imaging and real-world clinical deployment?

A: I see the gap as a series of misalignments in three key areas.

First is the Metric-Task Gap. In research, we often optimize Dice coefficients or AUCs, but in the clinic, the only metrics that matter are patient outcomes, diagnostic confidence, and workflow efficiency. We must start asking the right questions. For example, Does this model reduce the time to treatment?

Second is the Technical-Clinical Language Barrier. Successful deployment requires a deep, continuous dialogue with clinical partners. We need a common language between the engineer and the doctor to understand the nuances of the clinical workflow, ensuring our models are configurable and adaptable to physicians’ specific needs.

Finally, there is the Reliability Gap, specifically concerning domain shifts and explainability. To gain clinical trust, we need robust, generalizable architectures and explainability that enables medical experts to understand why a model flags a specific region.

Q: What advice would you give young researchers who want to work at the intersection of AI and healthcare?

A: My advice would be to technically stay at the absolute forefront of AI modeling,while simultaneously immersing themselves in the clinical world.

Here is how I suggest approaching it:

• Form Deep Partnerships with Clinicians: Don't just look for data; look for collaborators. Spend time in the clinic, learn the 'pain points' of the medical staff, and understand the workflow.
• Solve for Scarcity: Be aware that the biggest hurdle isn't just model architecture; it’s the reality of limited labeled data. Focus your research on how to do more with less through self-supervision, generative modeling, or transfer learning.
• Redefine Success: Don't just settle for standard AI metrics. Work with your clinical partners to define task-specific metrics that reflect real-world clinical utility that is tightly coupled with patient outcomes.

• Go Deep on a 'Vertical': While it’s tempting to be a generalist, find a specific clinical domain—be it oncology, cardiology, or neuroimaging—and go deep. Understanding the nuances of one vertical is what allows you to build truly unique, relevant models. That is where real breakthroughs happen.