Enormous potential

Georg Langs teaches at MedUni Vienna’s Department of Radiology and Nuclear Medicine’s Computational Imaging Research Lab. He is also one of the founders of contextflow, an AI start-up from Vienna.

Guido Gebhardt spoke with the scientist about the integration of AI in RIS and PACS.

Professor Langs, how do you assess the current situation of artificial intelligence in radiology?

While there was initially a lot of enthusiasm for solutions that addressed only one specific problem, it soon became clear that radiology practices or clinics cannot and do not want to deal with procuring, integrating and maintaining 30 different algorithms for different application scenarios. At the same time, there is still a need to use AI to improve diagnosis, prognosis and treatment decisions.Very exciting developments are taking place in the field of predictive models and integration of diagnostic data. In practice, users are asking for comprehensive integrated solutions or platforms or marketplaces that make it easier for them to install and integrate several AI algorithms, even from different providers.We can already see that, in addition to the classic providers of AI algorithms for different organs, companies are also establishing themselves that deal with the seamless integration of algorithms in RIS and PACS.My feeling is that we are currently in a start-up phase. At the moment, the first sales are coming in and I am curious to see how the market will look in two years and how the landscape between AI platforms and individual integrations will develop by then. The race for the best AI algorithms is also far from decided – there will still be a lot going on. I assume that comprehensive solutions that cover entire areas and improve diagnostic possibilities in addition to efficiency will prevail.

At the moment, there is talk that there are already well over 200 providers of AI solutions for radiology, and the trend is rising. Will the market consolidate soon?

From the radiologists’ point of view, it is not directly important how many providers there are at the same time. In my opinion, the offer will be reduced to a few in the future. At the same time, their portfolio will increase in order to cover the widest possible range of applications, because the installation and integration costs of individual solutions should not be underestimated.

How do you think the market will continue to develop?

When we talk about artificial intelligence in radiology, we are not just talking about an algorithm that delivers measurements. The added value of AI solutions also depends on how they are integrated into the workflow: from registration to sending findings and their use in interdisciplinary case conferences. This is about interfaces and how the individual measured values of the highly developed algorithms are transferred via the RIS and the PACS to the structured findings and then to the referring physicians. I see enormous development potential in this. The competition will take place in the quality of the AI algorithms and the integration into the workflow, all of which requires joint iterative development work with the radiologists. In order to realise a gain in efficiency and to exploit the diagnostic contribution of imaging to treatment decisions, one has to deal intensively with the user interface in addition to the medical questions and involve all process participants in the workflow integration.

This means that radiologists have to participate more intensively in integration than they are used to from the large units?

Yes, that is what it means. Not only radiologists are affected by data processing, but also referring physicians and other disciplines – the keyword being integrative diagnostics. If all players in the healthcare system are networked in the future, the data flow in radiology will begin with the registration, continue via the modality and the AI-supported reporting into RIS and PACS and from there in the form of a structured report of findings to the referring physicians or the physicians who continue treatment. At the same time, the integration of information from disciplines such as radiology, pathology or laboratory medicine also plays a key role here. The added value of such solutions lies in the seamless processes and the gain in information. All persons involved in the treatment process get immediate access to quantitative values that are decisive for further treatment and the success of the therapy.The challenge is therefore to handle the data correctly and to present the information in such a way that it offers added value. This means that all those involved in the process chain should really engage with artificial intelligence.

As an AI provider, how does contextflow ensure that their data, once it is integrated from the platform into the PACS, is presented correctly, as they would like it to be?

So far, no matter how the integration is carried out, the team at contextflow spends a lot of time with the radiologists to coordinate and optimise the solution – this is a culture that has certainly been taken along from its origins as a spin-off of a medical university and university hospital. It doesn’t matter whether it’s a single integration or a digital marketplace. There is also an ongoing very close cooperation with all integration partners, such as PACS manufacturers, to jointly optimise the integration into the routine. The aim is not only to show the user isolated findings, but to understand exactly where and when which information makes a contribution to the diagnosis and further decisions.

What do you think is the right time to start with AI? Should you start right away or is it better to wait?

In my opinion, now is a good time to enter, as one is still involved in the further development. At the moment it is being decided where the journey will go. Although almost all manufacturers have mature products, the interesting thing at the moment is the fine-tuning and further development: to further improve the solutions, to shape the role of AI and to optimally adapt it to the respective application on site. And that is happening right now.

What about the validation of algorithms in the future? The users definitely want to know why and how the algorithm arrives at its result?

I think the demand for validation of AI solutions is important and right. Algorithms must of course be validated. This takes place within the framework of benchmarks and is proven by numerous scientific studies. At the same time, it is important that the results can be explained. Comparative cases, for example, play an important role here, as they make the reasoning transparent. A key point here is that manufacturers communicate very clearly what the software can and cannot do. Users confirm that they not only work more efficiently with the help of AI, but also achieve better results because, for example, observations become quantifiable.Quite independently of this, all manufacturers must approve their respective systems as medical devices within the framework of the MDR (Medical Device Regulation). Another important aspect is the feedback loops that manufacturers need in order to continuously develop the systems and improve their integration into clinical routine together with the users. AI algorithms become more accurate the more they are trained on real and diverse data. Manufacturers need feedback from confirmed examination results from everyday use to improve the algorithms and adapt them to changing care – there is a lot happening in the field of continuous learning.

Are there any prerequisites that have to be fulfilled if I, as a radiologist, am involved in the integration of AI algorithms?

As already described, one must think about the use of AI beyond one’s own departmental boundaries and include those who collect diagnostic data as well as those who are supplied with data and information by radiology in decision-making processes.Technically, manageable computing power is sufficient to create interfaces between the individual IT systems. The situation is different when it comes to the requirements for automated image analysis. With on-premise solutions, it is advantageous to have a powerful GPU, but cloud solutions can be more efficient overall and also become practicable with the establishment of data protection framework conditions.With the integration of AI, we are in principle already preparing for the next development step: Integrated Diagnostics. This involves intensive, intersectoral and interdisciplinary and probably also international exchange and the integration of images and findings, as we already know from tumour conferences, together with values provided by AI. To achieve this, there is still work to be done within the hospitals and also medicine in terms of digital infrastructure and processes. This joint development at the interface between medicine and AI is, in my opinion, one of the most exciting aspects of the current development. In practice, it is about harmonising technologies and processes to simplify the integration of IT systems. Together with interdisciplinary research, this enables the further development of AI in medicine and to tap into the enormous development potential.