IHE is formulating AI standards
The use of AI models in the radiological workflow depends largely on standardized interfaces for data exchange. An interest group of IHE Europe is formulating the prerequisites necessary for this purpose. During a Plugathon, it has now made considerable progress.
The vision of personalized medicine sounds promising: From a large number of laboratory values, medical history data, test results and graphs, the physician drafts tailor-made therapy plans. In practice, however, the full potential of integrated diagnostics is still not being used enough. Dr. Marc Kämmerer, radiology specialist and Head of our Innovation Management Unit, also knows the main reason for this: "Diverse software systems are being assembled on the basis of proprietary interfaces and inadequately structured data. But this approach is leading to a dead end. Interoperability can function only when all participants in the process support the respective standards and give data meaningful interoperability on the basis of their level of structuring."
A task force is finding new applications
The AI Interest Group for Imaging (AIGI), which Marc Kämmerer has established, is developing precisely such standards. The group is acting as a task force of IHE Europe. The objective of IHE is to improve the communication between IT systems and medical instruments and to make them usable in clinical workflows. To this end, it is developing so-called integration profiles by means of international standards. The international team of experts of the AIGI is concentrating on the integration of AI solutions in radiological workflows.
During a Plugathon on the occasion of the IHE Connectathons in Trieste in June, the group tested the entire data chain up to the drafting of a radiological diagnostic report for the application of "AI-assisted thoracic and cranial diagnostics". In the process, systems of nine different manufacturers were interlinked. This live testing very quickly revealed the deficits with respect to scalable integration and usability.
One basic for the use of AI in the clinical context is the need for a validation process. On the one hand, this is necessary to ensure that correct results are exchanged efficiently between the systems. On the other hand, the new EU AI Act requires the quality of results of the AI solutions to be monitored. One example of an indicator for this purpose may be the ratio of correct to false results plotted over time. Since validation does not necessarily take place on the same system as that used for further processing of the data (e.g. for writing the diagnosis or for monitoring), an interoperable, cross-system exchange of the results of the verification process is needed.
As the number of AI solutions in use grows, the use of a service discovery will become meaningful. This is a service that, for example, tells a PACS which requirements an AI system places on images to be evaluated. As Marc Kämmerer puts it, we have also made "a giant leap" in delivering AI results from a viewer to a system for reporting the findings. In this regard, the participants in the Plugathon have created the first draft of a data record that already considers the very different requirements of the various manufacturers.
The status quo: a smorgasbord of solutions
However, the task force is still far from reaching its objective. As Marc Kämmerer explains, this has a lot to do with the branch structure of the AI world. "In many cases, the manufacturers come from the startup scene and know little to even nothing about IHE. For this reason, or for lack of standards, they have created their own approaches to solutions, for example as regards communication of status information about the progress of processing or the transfer of results to other systems." Thus a "colorful smorgasbord" of approaches has arisen – just the opposite of what is needed for scalable use of AI in medicine.
In this connection, the IHE has already developed two profiles that take the new workflows into account: AI Results (AIR) defines how the results of medical imaging analyses are reliably stored, retrieved and displayed. And the AI Workflow for Imaging profile (AIW-I) deals with use cases involving the request, management, and performance of AI inference on digital image data. Both profiles are currently in the status of trial implementation by the manufacturers. Kämmerer believes that, with this and the newly formulated use cases, we have now made "an important step" further on the way to integration of AI solutions. At the same time, the Plugathon has shown once again that "We can go further by working together". The requirements for the verification process will be described in the new IHE profile suggestion entitled AI Result Approval for Imaging (AIRAI) around mid-November by the AIGI Group for IHE and, with a little luck, it may be possible to subject them to preliminary testing as early as 2025 in Vienna in a further Plugathon during the IHE Connectathon.

"Integration of AI in the workplace will not be scalable without the use of standards."
Dr. Marc Kämmerer
Head of Innovation Management VISUS