Dr. Piotr Radojewsk: In response, it is worth looking in more detail at evidence and its classification into different levels: The lowest level is technical evidence that merely says that an AI system functions. At the second level is diagnostic accuracy, at the third level is the ability to think diagnostically, and at the fourth level is therapeutic evidence. Steps five and six ultimately describe the primary aims, namely the outcome for patients and finally the effects on society. Most studies are available for the second level, that is, the diagnostic efficiency; beyond that, the evidence gets rather thin. This is also due to the fact that the requirements from authorities have also not stipulated higher levels of evidence—a situation that is currently changing. This is in part because legislation is being adapted. For another, hospitals and clinics are demanding it.
A good catchword: How do you decide whether AI is used in your hospital and for your work in radiology?
Dr. Piotr Radojewsk: The relevant societies have since developed “To buy or not to buy” guidelines for hospitals. These are checklists intended to make it easier for hospitals during the procurement process to identify solutions that cover their actual needs. Ultimately, when using AI it is also essential that the solution fits into a hospital’s workflows. For example, we evaluated an excellent solution for stroke diagnostics. Unfortunately, the software was not developed or adapted for the Swiss healthcare system. That meant that it simply did not fit into the hospital’s workflows.
We initially discussed the parts of the body where AI in radiology is used. What are the procedural steps that AI currently supports?
Dr. Piotr Radojewsk: In principle, AI can be used along the entire radiological workflow: in planning, image acquisition, evaluation or analysis, and reporting. Image acquisition in particular is a major issue at the moment.
Where exactly does the use of AI fit in image acquisition?
Dr. Piotr Radojewsk: Acceleration. All renowned manufacturers of MRI instruments are currently working on reducing imaging times. Examination times in MRI can now be reduced by 50 to 90 percent. We are now reaching the examination times needed for CT. Underlying this acceleration is undersampling: fewer data points are collected and the AI fills in the missing information. But even here we have to ask about the evidence: Does an accelerated image reconstructed with AI correspond to reality? Are we possibly missing important information in the image or are there artifacts?
The strongest domain for AI along the radiological workflow is currently analysis with quantification and validation. This is an area with the greatest potential to save time for radiologists in their daily work. So that this can actually be implemented, hospitals have to ask themselves exactly what they want and need so that they benefit: should AI act as a background triage system? Should it support decision making or quantify and validate available findings? These questions need to be answered.
You sketch isolated solutions. It would be interesting if the entire clinical process could be optimized using AI.
Dr. Piotr Radojewsk: Yes, and there are good approaches to achieve this. One example comes from the treatment for ischemic stroke. If the AI system detects an occlusion in the images, the necessary treatment chain is triggered in the background. The AI then takes over the organizational tasks, forwarding information and notifying departments when a patient will be arriving. In highly fragmented healthcare systems, this can greatly reduce the “door-to-needle time.” There is enormous potential but success depends greatly on adapting the tool to the requirements of the particular healthcare system.
Let’s talk again about the foundation of a good AI solution: the data …
Dr. Piotr Radojewsk: … a critical point. Posters with the call to “Donate data—detect diseases earlier” can be seen around Switzerland at the moment. And that’s what it’s about. We need lots of data, we need valid data. For all AI applications, not just those that detect diseases earlier. Unfortunately, at the moment we do not have enough data, which is a major limitation of AI. Another challenge are what is known as data mismatches. This means that models are trained on data that do not reflect subsequent clinical reality, for example, the relevant patient population. This is really a key point. Some AI experts postulate that the data and not the algorithms are the key to success in AI. I agree with this opinion in principle.
In the context of data quality, the need to structure data so that they can be linked also plays a role. We must also get to the point where different data sources—laboratory data, imaging data, text data—can be linked together. This assumes standardization and in turn this requires structuring.
Let’s have another look at the future: What do you hope to get from AI in five years?
Dr. Piotr Radojewsk: I hope that we have a much better body of evidence so that we can more clearly understand which solutions are really useful. The second important development concerns the use of large language models. An initial triage could be performed using an AI chatbot, for example. We’ll see whether that eventuates. However, the most important message in my opinion is that AI will not replace radiologists. AI will be a partner for employees in healthcare. It will confirm my diagnoses, for example, or confirm to the nurse that a medication has been given—AI will become a daily helper, of that I am convinced.
Many thanks for the interview.
