CVIR: The new editors for AI

April 5, 2024

Artificial intelligence (AI) is evolving rapidly and is becoming more important within medicine and IR by the day. It has become imperative for CVIR to include this field, and the journal has added a new section dedicated to AI to the editorial board this year.

CVIR invited three esteemed experts in the field of AI to join the new section: Dr. Bradford J. Wood, Prof. Dania Daye, and Dr. Tom Lévy-Boeken. Their extensive expertise, diverse perspectives and ongoing contributions to AI in IR make them invaluable to the Journal’s team. We spoke to the new editors to hear what sparked their interest in AI and what CVIR authors and readers can look forward to reading from them in the future.

CIRSE: Can you tell us about yourself, your hospital, and your personal interest in AI and how it affects your daily practice?

Wood: I am lucky to lead a team of smart, eclectic, multi-disciplinary IR MDs, PhDs, software engineers, chemical engineers, image processing and regulatory science experts, and trainees with a clinical arm in IR, and a research arm in the Center for Interventional Oncology in the Intramural Research Program of the National Institutes of Health (USA), the National Cancer Institute, the National Institute of Biomedical Imaging and Bioengineering, and the NIH Clinical Center, the world’s largest hospital dedicated solely to clinical and translational research. AI is a tool that is becoming integrated in every corner of what we do, in hidden ways as well as more obvious ones. Today, we use AI for image reconstruction, segmentation, registration, and optimization of pathways, and endpoints. AI identifies and selects specific biopsy targets and classifies lesions on prostate MRI, and can actually predict a tumour margin or treatment plan that is not always available to the human naked eye. What is most exciting though is what is yet to come. These are super exciting times to be in IR.

Daye: I am currently an academic interventional radiologist at Massachusetts General Hospital and an Assistant Professor of Radiology at Harvard Medical School with a joint appointment in the MGH/HST Martinos Center for Biomedical Imaging where I lead the Precision Interventional and Medical Imaging (PIMI) Lab. My research centres on the applications of machine learning and computer vision for precision medicine. I am an engineer by training. I completed my MD/PhD at the University of Pennsylvania, where I obtained my PhD in Bioengineering. Part of my research there focused on using machine learning for breast cancer diagnosis, risk prediction, and prognostication. I have used the skills I gained during my PhD to focus my current research endeavours on the applications of AI in IR. I am very excited about all the opportunities this new technology will bring to advance the care of the patients we serve in interventional radiology.

We are currently in the very early stages of seeing AI affecting our daily practice in interventional radiology. While there are a number of applications on the market today to aid in care coordination and patient selection for IR procedures, we are just scratching the surface. I can’t wait to see how AI will change our practice over the coming decades.

Lévy-Boeken: Let me start by expressing how thrilled I am to be joining the editorial board. Being part of a new section is just like AI in IR: few landmarks, many blank pages, and hopefully a bright future! I joined Professor Sapoval’s team three years ago. Our team is fully dedicated to interventional radiology and consists of three (and soon four) interventional units. We are based in Paris, at the European Hospital Georges Pompidou, within APHP (greater Paris public hospitals).

I come from a mathematical background and later entered the medical world, which explains why I am working both at the university hospital and in a math lab. My main area of clinical practice is digestive oncology, and we are trying to bring AI towards our interventions to enhance every step of our work.

To be completely honest, AI does not affect my daily practice yet. This is probably true for many physicians: we attend great congress presentations, read impressive publications, yet every day cases seem like… every day cases. I have to admit this is part of the fun, taking part in something that has yet to come true. Regarding the preclinical research we are conducting, I am personally interested in automation and how AI might affect the core of image-guided interventions where we basically perceive and act simultaneously. If you look at AI and radiology, there is a tremendous gap between diagnostic research (whether institutional or industrial) and interventional research. You can’t blame radiologists for focusing on images at first, but we now have the capability to go beyond from personalized multimodal patient selection towards smart robotics.

CIRSE: AI is rapidly evolving. How have you personally perceived this with regards to AI in IR?

Wood: AI is transforming medicine as we know it, at a light-speed pace. What will come in its place is a personalized and predictive paradigm for the coolest medical specialty on the planet. No other field combines imaging skills, clinical care, high technologies, digital skills, computer science, engineering sciences, and now data science in an exquisite symphony of man and machine. IR is ripe for AI, but it will take time to fully realize the immense potential. Imagine an AI tumour board that helps select the best pathway, knowing who will benefit from what tool, the best place to biopsy, the digital pathology combined with genomics, and proteomics will help determine the treatment plan that will be delivered with AI assisted robotics along AI optimized pathways with AI selected catheters and probes to fit the anatomy. AI models pick the detector obliquity, vessels and sequence of tracked robotic needles, with next to no radiation, and will let us know when we are done with the procedure and when to follow up. Decision points in the procedure will be informed by whatever high level evidence is available, with references and likelihood of outcomes spoken to us during key parts of the procedure. Video-game medicine! IR is the coolest!

Daye: Today, most of the applications of AI remain in diagnostic radiology. The number of approved algorithms has been increasing exponentially year over year. In IR, we have been a lot slower to adopt this new technology in our day-to-day practice. This is due to the many challenges posed by the imaging data we have in IR. Our IR datasets tend to be a lot smaller than our diagnostic counterparts. There is limited standardization of the intra-procedural data collected and the views obtained. Many of our IR applications require integration with the electronic health record for maximal impact. That said, recent advances in synthetic data and the establishment of data sharing consortia are opening new avenues for research. We hope this will accelerate the development of new AI algorithms in IR.

Lévy-Boeken: AI is rapidly evolving. One of the main challenges in IR is the lack of large databases for what we do, and the lack of adequate labelling. Specialized centres provide specialized interventions, one centre focuses on SIRT while another is well-known for lymphatics. Even meta-analyses on mainstream treatments such as liver ablation do not go beyond a few thousand patients. This might seem like a lot, but foundation models for diagnostic imaging are now being trained on hundreds of thousands of cases. Since most models initially required stereotyped labelled datasets, IR was not suitable for early AI applications. This deepened the contrast with the rapidly expanding applications for non-interventional radiology.

The most recent breakthroughs in calculation capabilities lead to few- or even zero-shot learning models: after training, these models are able to execute tasks that were not even present in the database. This is huge for IR: we can now adapt high-performance models to our specific and specialized activity. This will certainly accelerate the interest in AI stemming from our IR community for applications built on small datasets.

Another breakthrough is accessibility. Chat GPT is a solid example of how anyone can now expect to try new models without any prior computational knowledge. This will ultimately attract more IRs towards these solutions.

AI in robotics is a game-changer. Eventually, our clinical practice might be deeply transformed by upcoming technological advancements in this field.

CIRSE: You are about to start work on guidelines for authors about AI manuscripts – what is the background behind this initiative?

Wood: Making AI-empowered IR the best we can be, will require patience, focused attention to rapid deployment of new data science tools, large and seamless data and video annotations, and most importantly, shared data sets and standardized processes that respect privacy and local regulations. We need the IR community to ask questions and solve problems that are relevant, add value, and fit a business plan. This won’t be easy, but Dania, Tom, and Klaus are super teammates! Clinicians need to work elbow to elbow with data scientists and computer scientists to address unmet clinical problems. We need to also help establish methods and principles for AI in IR that match and customize for IR the rapidly emerging constructs from academic data science and medicine. The time is now.

Daye: It is important to ensure robustness and reproducibility in the  algorithms that are being published to make sure they have the potential for an eventual translational into clinical care. Establishing an author checklist for AI articles will provide guidance for authors and allow them to submit the highest quality manuscript possible for publication, and to ensure higher impact of their research. We are currently in the process of putting together an AI article checklist that we hope will help authors in their submissions and will allow us to ensure the robustness and quality of the research published in this space in order to further advance the field.

Lévy-Boeken: From a publication perspective, such guidelines could secure some basic principles. For example, ethics are of the utmost importance because this research is data-driven.

The idea is to bring everyone together on this joint venture based on shared values and interests, not to create a set of rigid rules that might be dissuasive or excluding. Good models are important, but the clinical impact and reproducibility will always be at the centre.

This initiative stems from the feeling that it is incredibly hard to determine the quality of a model based on a publication. It takes thorough reviewing, and readers might not have the time to go through the models themselves. Guidelines will help ensure quality in a way readers can be confident that the technical aspects were verified. The idea is to generate interest for all interventional radiologists in a trustworthy fashion.

CIRSE: What are your hopes for the future? What goals do you aim to accomplish in this position during your term of office?

Wood: CVIR is on the cutting edge of our science and our discipline. We have a great opportunity in the coming decade to implement AI in IR to improve our practices, and impact our patients. I hope to help motivate and shine a light for early-career clinical IR MD and MD PhDs as to what the future might look like. To get there will require open arms and broad collaborations across continents, including new ways to share data while respecting privacy and ethics and avoiding new bias. We need to establish how IR is different from diagnostic radiology in terms of AI, and teach IR clinicians just enough data science to be impactful. There is strength in numbers and unity, so we hope to partner with other IR and DR advocacy groups, societies and journals to build our team larger and more unified globally. Collective decisions we make as a specialty this year will determine the health of IR for the coming decade.

Daye: I have no doubt that AI will transform how IR is practised over the coming decades. AI will eventually allow us to choose the right device for the right patient for the most optimal outcome. Advances in Generative AI models will also allow us to optimize the way we deliver care to our IR patients. I foresee that there will soon be IR applications that will show an effect on patient outcomes.

I am very excited to be serving in this position and to help support the development of this new area for the journal. I look forward to working with Brad and Tom to further develop this new section.

Lévy-Boeken: Our specialty depends on innovation and on the rapid adoption of techniques. I believe CVIR plays an essential role in promoting new therapies throughout our community, and ensuring interventional radiologists are able to perform these new interventions in the safest way.

My hope for the future is that readers also get to try the models. I believe we are getting closer to generalizable models. Large language models are now used by everyone, we could ask in the near future that every team who publishes an AI-related article immediately provides usable interfaces, not only raw codes that demand so many prerequisites. Just like a new navigation software or a new ablation probe, my hope is that everyone gets a chance to read about it and eventually decide for themselves how impactful or insignificant the proposed solution is.