Nowadays a discussion about artificial intelligence (AI) in radiology triggers neither exaggerated nor apocalyptic feelings. That is good. Because now that tempers are no longer flaring and the debates have become less emotional, it is finally possible to discuss the true potentials and limitations of AI. For example, with Prof. Dr. Elmar Kotter, President of the European Society of Medical Imaging Informatics and Chair of the "eHealth and Informatics Subcommittee" of the European Society of Radiology.
VIEW: Prof. Kotter, what do you understand by AI in radiology?
Elmar Kotter: This is one of the most difficult questions in the entire range of issues: I think the answer to it is in a constant state of development. In general, I would distinguish between a very narrowly framed definition and a broad interpretation of the term: "narrow" AI for specialized tasks such as identification of a fracture and "broad" or "general" AI. In radiology, we currently find ourselves in the narrowly bounded frame. This means that the systems are trained to handle a specific task, for example the recognition of pathologies or of pulmonary nodules. I like to use the term "Augmented Intelligence" for radiologists. AI also helps in taking control of the mass of images with which we are confronted.
What does the broad definition include?
A decisive step that AI in radiology must take in the coming years is the incorporation of clinical information. When we incorporate laboratory values or further disease-related data into AI in addition to the images and relate them to the images, the tasks can rapidly become more complex. But "Out-of-the-Box" networks do not yet exist for building appropriate infrastructures with non-image information. This is why manufacturers and users are concentrating for the moment on the purely image-related AI applications, where results can be obtained and published rapidly.
But AI is already able today to do a little more than just image evaluation in radiology ...
Of course, software is now able to accomplish prioritization of the worklists. There are systems that support us in differential diagnosis or optimize the protocols on the instruments. But for the most part they involve applications such as fracture review, determination of bone age or recognition and measurement of neuroradiological parameters. Thus the potential of AI is certainly not exhausted, although it still faces great challenges. One of these is the incorporation of clinical information. Another is the integration of AI in the radiological workflow. Institutions such as IHE are indeed working in the background on standardized procedures, but at present they lack a uniform recipe for the purpose.
What we should not overlook in the entire discussion is the trust that the radiologists have for technology. We must understand how AI functions. What it can do, and above all what it can't.
In my opinion, however, the greatest problem with a view to AI applications is financing.
Can't time be saved and the quality of healthcare be improved by the use of AI, so that the revenue structure is optimized?
AI indeed helps us at the very least to master the flood of data produced by the increasing numbers of cases and the increasing number of images per examination. Moreover, it frees us from mindless tasks such as recognizing and counting pulmonary nodules. At the same time, however, we can't do without radiologists and generate direct savings. And so far we can't charge differently for a diagnosis that is more reliable or has been made more quickly by means of AI. Thus a financing model for the use of AI is lacking at present. In the meantime, things are a little different in the USA.
What exactly do you want and hope for from AI in your discipline?
A very important task of AI will exist in weaving a kind of safety net for the radiologists. Already today we have an enormous workload, which will continue to increase – and with it the risk of errors. By taking over repetitive and monotonous tasks, AI can unburden us radiologists and improve the quality of healthcare.
I see a further advantage of AI in making certain examinations more easily quantifiable and thus more objective. Take the example once more of pulmonary nodules: From a large number of foci, the radiologist will single out only a few for measurement and assessment for the purpose of monitoring progress. AI can determine all nodules and thus replace the currently subjective picture by a comprehensively objective one.
And finally, the current situation is still such that we don't evaluate, in any kind of structured manner, much of the information that we now collect every single day. This is a real grave for data. We are concentrating on certain pathologies, and we are simply ignoring other data. For example, a systematic measurement of aorta diameter or of bone or liver density. These values would represent a good early warning system for diseases, but we are completely unable to compile the values in structured manner. AI could do this.
How much will the working procedures of radiologists be changed by the use of AI as you would like?
In principle, I'm not worried that radiology will run out of work. If software can take care of a task, then so be it. In this way the daily work in radiology will presumably look different, because with AI we have a potential partner on our side, which – if we understand it correctly and know its strengths and weaknesses – can supplement our work very meaningfully and compensate for our weaknesses. To ensure that this promise will be honored, we must develop realistic expectation attitudes and not overload AI with promises of salvation. If we succeed in that, we will have a golden opportunity to establish radiology as a leader in clinical information management in the clinics. Among medical professionals, we are those who know best how to handle information. And we have the most experience with AI. That's why I see good opportunities for our discipline to achieve still greater relevance and to position it at the center of medicine.
"So AI is helping to manage the mass of images we face."
Prof. Dr. Elmar Kotter
President of European Society of Medical Imaging Informatics and Chair des „eHealth and Informatics Subcommittee“ der European Society of Radiology