The ESR’s AI blog was honored to speak with Michail Klontzas this week, to add to our growing series “On Artificial Intelligence”. Klontzas is currently an Assistant Professor of Radiology at the University of Crete in Heraklion, Greece, as well as an Editorial Board Member for the ESR’s flagship journal, European Radiology, in the Musculoskeletal section. Klontzas has a wealth of knowledge on the topic of artificial intelligence due to his experience in academia and his various roles, including being a member of the Trainee Editorial Board of the RSNA journal “Radiology: Artificial Intelligence” and participating in the steering committee of CLAIM guidelines for medical imaging AI. Let’s dive in!
What is your background/experience with artificial intelligence and what first attracted you to the topic?
I became interested in artificial intelligence during my first PhD at Imperial College London, where I worked on data analysis and bioinformatics for bone tissue engineering. Then during my radiology residency, I was fascinated by the potential of radiomics and deep learning to assist in diagnosis and predict treatment outcomes, which drove me to complete a second PhD in artificial intelligence for bone marrow imaging. During my PhDs, I was trained in coding and have worked on radiomics and deep learning for musculoskeletal and oncological imaging. A great experience for me has also been my involvement in the activities of the European Society of Medical Informatics (EuSoMII), which is the leading subspecialty society under the ESR dealing with AI. I have also been a member of the Trainee Editorial Board of the RSNA journal “Radiology: Artificial Intelligence” where I participated in the steering committee of CLAIM guidelines for medical imaging AI. For many years, I have collaborated closely with computer scientists and engineers who helped me see medical imaging from an AI perspective, understanding that AI will be deeply incorporated in our future as radiologists, and we need to embrace it to be able to progress.
What are the biggest challenges to AI adoption in clinical practice?
There are several challenges hindering AI adoption in clinical practice. To my mind, one of the most important is the lack of training of radiologists on AI methods. A lack of deep knowledge on the topic develops into an unreasonable fear of the use of such methods, which slows the adoption of AI in our practice. The second most important challenge is the lack of large prospective trials and lack of robust evidence supporting the clinical use of AI. The field needs to progress from small-scale retrospective studies to large multi-center prospective trials that offer indisputable evidence about the use of AI to address clinically relevant problems. In addition, regulations need to be developed to support radiologists who use AI for everyday clinical tasks, addressing cases where medical errors happen due to incorrect model predictions. Finally, transparency in research reporting will address reproducibility issues and increase trust in research results, allowing the wider adoption of novel AI algorithms.
Give us an example (or an educated guess) of what you think AI will be able to do in 3 years. What about in 10 years?
The field of medical imaging AI evolves rapidly and accurate predictions with a 10-year horizon are not possible. As a personal educated guess, I believe that foundation/multi-modal models will be incorporated into our practice and routine tasks such as measurements and exam protocolling will be largely performed with the assistance of AI. I strongly believe that the preparation of reports will be assisted by large language models and that AI tools will be widely incorporated in PACS systems to assist diagnostic decisions in distinct everyday clinical dilemmas. Nevertheless, expert (radiologist) oversight will be necessary to filter out false predictions by AI, even though the need for human interpretation will be significantly reduced, especially for screening tasks, since AI systems will be in place to flag potentially significant findings. Finally, the use of AI-assisted image reconstruction will lead to a significant rise in exam output, increasing the need to train more radiologists.
As a radiologist, what is your advice for younger colleagues wanting to dive into the topic of Artificial Intelligence, Machine/Deep Learning, and/or Radiomics? What are your tricks for staying up to date in this fast-evolving field?
I believe that the most important piece of advice is that younger colleagues need to delve into these topics to the same extent as they become educated in traditional diagnostic and interventional radiology topics. A good start would be the courses and webinars offered by the ESR/ESOR and subspecialty societies such as EuSoMII. The “Masterclass on AI” offered by ESR in collaboration with EuSoMII is an excellent resource to start with, providing all the basic knowledge for the majority of radiologists. Many people believe that they need to start by learning how to code, which may be appealing to those not interested in AI research and computer science. This is a common misconception which stops younger colleagues from getting acquainted with these topics. However, since AI will be incorporated into every aspect of our radiological routine, we all need to be familiar with at least the basic concepts and terminology of AI, to be able to understand how these tools work and to be able to critically assess the results obtained by their use. Some tricks for staying up to date in this field include the attendance of relevant lectures at radiology conferences such as the ECR, reading important review papers on key topics published in ESR and RSNA journals, and subscribing to newsletters which provide a weekly/monthly digest of advancements in the field, which are simplified for radiologists without a computer science background.