Following a fellowship in artificial intelligence (AI) at the Institute of Imaging and Computer Vision at RWTH Aachen, which only solidified his belief in AI’s great potential in medicine, Daniel Truhn, a Diagnostic and Interventional radiologist at University Hospital Aachen in Germany, started working on AI’s application in medicine, even creating his own working group in the area. Join us as we explore Truhn’s thoughts on the challenges to clinical adoption of AI, what its possible future looks like in radiology, and more.
What is your background/experience with artificial intelligence and what first attracted you to the topic?
By training, I am both a physicist and a physician. After finishing my board certification as a radiologist, I did a fellowship in AI at the Institute of Imaging and Computer Vision at RWTH Aachen. This was approximately when AI started to gain traction in medicine while it had already been used for image processing outside of medicine. When I met the people working there, it was immediately obvious to me that AI had great potential and I soon started working on its application in medicine. I subsequently returned to the hospital and started my own working group.
What are the biggest challenges to AI adoption in clinical practice?
In my view, there are three main challenges: 1) To make AI models clinically useful such that they either make clinicians more efficient or allow one to draw new/better insights based on data – we are starting to get there. 2) To integrate AI into the hospital IT infrastructure. In most hospitals, the IT systems are highly complex in that there are hundreds of different software systems, each with its own user interface. Making real-time clinical data accessible to the AI models is hard – we are getting there by implementing standardized data formats and APIs such as FHIR, but that is a tedious and slow process. 3) To ease the interaction between humans and AI models. Ideally, the human user should have a good feeling about why the AI model draws its conclusion – this touches on the issue of explainability and is a research field that has yet to be explored in large parts.
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?
In three years, we will see the integration of many AI models that support clinicians to become more efficient in dedicated workflows, e.g. reporting on a radiograph. Furthermore, some AI models will emerge which can perform tasks that humans are unable to do, e.g. the stratification of patients for therapy based on biomarkers in their H&E histopathological image; i.e. in three years, we will see an increase in the interaction of humans to AI-models on a one-to-one basis.
In ten years this will change. AI models will communicate with each other and multiple AI agents will collaborate to build a comprehensive assessment of the patient. There will be image analysis agents that speak to agents which analyze the patient’s history and report back to treatment-decision agents with access to guidelines. This means, in ten years, humans will orchestrate a plethora of specialized models, speak to these models in natural language and give orders such as, “Please propose the best treatment for this patient based on all available data”. Therefore, the interaction will be: Human – AI-model-AI-model-…-AI-model.
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?
Dive in! The possibilities are fascinating, but it is hard to stay on top of all existing research. Try to build a solid foundation; for example, take an online course (there are some good courses out there conducted by reputable institutes such as Stanford University). I would also advise not running after every newly published paper – there are too many to keep track of. The established journals (NEJM AI, Radiology AI, etc.) should give you a sufficient overview of what is currently possible in AI in medicine. They are usually lagging behind about 6 months to 1 year because the field is evolving at such a rapid pace, but this will give you a good grasp of what AI is currently being used for; then try to build your own research project.
