In radiology, we have now moved beyond the phase of ‘just experimenting a bit with AI’. The question is no longer whether AI has a role, but how we can effectively implement it. Not as a hobby project of a single enthusiastic physician, but at scale, across the entire department, with clear governance and preferably a solid business case.
Increasingly, hospitals are choosing not to purchase and implement individual AI solutions separately, but to invest in a more holistic infrastructure: an AI platform. AI platforms promise to make the implementation and integration of AI easier. Instead of having to go through a separate procurement process and technical integration for each AI product, a platform provides a single central access point: one contract party, one technical integration, one place for monitoring, and a marketplace from which hospitals can choose among multiple AI solutions. This enables quicker implementation of new AI applications, clearer management, and simpler evaluation and switching between different products.
Various companies offer such AI platforms. There are PACS vendors like Sectra and Agfa, independent platform providers like Incepto and deepc, major AI suppliers likeAidoc, OEM companies like Siemens and Philips, and post-processing vendors like TeraRecon. So, there's a lot to choose from, but what should you pay attention to?
Spaarne Gasthuis recently decided to purchase an AI platform, and Amsterdam UMCintends to do the same. Kicky van Leeuwen, AI implementation specialist at Romion Health, discusses the motivations and purchasing process with Inez Verpalen, a radiology resident at both hospitals and postdoc researcher at Amsterdam UMC. Due to her strong interest in AI and her research in this area, she is closely involved in the platform selection process.
We were looking for a suitable way to integrate multiple AI models simultaneously. We wanted a streamlined and structured approach that would also be future-proof. Without a platform, integrating an AI model is a complex process that takes a lot of time, both in decision-making about which AI tool to use and in the technical integration. This entire process is currently very costly, especially since we often don’t know in advance whether an AI model will provide a real clinical benefit.
Nowadays, for a single AI application, take AI for fracture detection, for example, there are many different AI products available. These products can differ in effectiveness and in how they affect workflow. To be able to test them both retrospectively and prospectively, we looked for an AI platform that would allow such evaluations, so we could determine which product is most suitable for our localsetting.Some AI platforms, for example, offer dashboards to monitor AI results, helping departments gain insight into the impact of AI implementation. A platform also harmonizes the technical integration of different AI tools and supports scaling AI models within the department.An additional potential benefit is that AI platform providers often have contracts with multiple AI vendors, which can reduce the price of a product through market competition. A similar effect could arise in the future if hospitals run multiple platforms in parallel within one department, creating competition between platforms. In Norway, for instance, they’re running three platforms simultaneously.
Currently, there is no reimbursement from health insurers for an AI platform. Hospitals or medical partnerships must therefore invest themselves if they want to purchase a platform. This also applies to individual AI applications, a budget needs to be allocated for those as well.In most cases, the purchasing process begins with an orientation on different AI platforms and drafting a business case. It is important to involve a broader team in defining the requirements, including PACS management and the IT team. Final decisions are made higher up, at the departmental board, procurement department,and hospital executive board.
We wanted to be able to compare different vendors for each AI application, both in retrospective validation settings and in prospective ones. So the platform needed toenable comparisons of models within our own patient population. Additionally, some platforms offer functionality to integrate AI results, such as measurements, directly into the radiology report. This prevents radiologists from having to dictate them manually.
Yes, definitely. Due to its focus on research, it is important for Amsterdam UMC to be able to implement and validate self-developed algorithms through the platform. On the other hand, general hospitals are more likely to finalize a business case quickly, as they benefit more directly from current AI models that help reduce workload or decrease after-hours service burdens. The PACS and EHR systems also vary between hospitals, not necessarily between academic and general hospitals, but this can still influence the choice of AI platform.
Of course, you need a legal contract. That is mostly handled outside the radiology team but is crucial for ensuring patient privacy through pseudonymization and for securing the cloud infrastructure. Often, you can also choose to run the analyses on-premises (locally), depending on the local computing capacity and the hospital’s preferences.
Once that is in place, technical integration must follow, with new connections to the PACS and EHR. I don’t yet have a full overview of that complex IT integration, and close collaboration with the IT department is essential. It will be a lengthy process, all the more reason to get it right the first time.
I mainly hope that choosing a platform allows us to keep up quickly with the latest developments and greatly enhances the scalability of AI in radiology.
I expect significant gains in reporting quality, in terms of grammar, spelling, and direct integration of AI analyses into the report. Maybe even leading to fully automated reports linked to AI results.
Additionally, I believe the current generation of AI tools is just the beginning. So much more is coming, and we want to be prepared for that. For CT and MRI, the number of AI applications is still relatively limited, but that is going to change. In five years, today’s AI apps will have evolved or been replaced by better tools. By then, you want your platform integration to be solidly in place so that you no longer need to worry about technical implementation and can truly focus on demonstrating clinical benefit and making a real impact with AI.
A variation of this article named 'AI-platform: handig, maar waarop te letten bij de aanschaf?' was previously published in Dutch in the Memorad.
Increasingly, hospitals are choosing not to purchase and implement individual AI solutions separately, but to invest in a more holistic infrastructure: an AI platform. AI platforms promise to make the implementation and integration of AI easier. Instead of having to go through a separate procurement process and technical integration for each AI product, a platform provides a single central access point: one contract party, one technical integration, one place for monitoring, and a marketplace from which hospitals can choose among multiple AI solutions. This enables quicker implementation of new AI applications, clearer management, and simpler evaluation and switching between different products.
Various companies offer such AI platforms. There are PACS vendors like Sectra and Agfa, independent platform providers like Incepto and deepc, major AI suppliers likeAidoc, OEM companies like Siemens and Philips, and post-processing vendors like TeraRecon. So, there's a lot to choose from, but what should you pay attention to?
Spaarne Gasthuis recently decided to purchase an AI platform, and Amsterdam UMCintends to do the same. Kicky van Leeuwen, AI implementation specialist at Romion Health, discusses the motivations and purchasing process with Inez Verpalen, a radiology resident at both hospitals and postdoc researcher at Amsterdam UMC. Due to her strong interest in AI and her research in this area, she is closely involved in the platform selection process.
We were looking for a suitable way to integrate multiple AI models simultaneously. We wanted a streamlined and structured approach that would also be future-proof. Without a platform, integrating an AI model is a complex process that takes a lot of time, both in decision-making about which AI tool to use and in the technical integration. This entire process is currently very costly, especially since we often don’t know in advance whether an AI model will provide a real clinical benefit.
Nowadays, for a single AI application, take AI for fracture detection, for example, there are many different AI products available. These products can differ in effectiveness and in how they affect workflow. To be able to test them both retrospectively and prospectively, we looked for an AI platform that would allow such evaluations, so we could determine which product is most suitable for our localsetting.Some AI platforms, for example, offer dashboards to monitor AI results, helping departments gain insight into the impact of AI implementation. A platform also harmonizes the technical integration of different AI tools and supports scaling AI models within the department.An additional potential benefit is that AI platform providers often have contracts with multiple AI vendors, which can reduce the price of a product through market competition. A similar effect could arise in the future if hospitals run multiple platforms in parallel within one department, creating competition between platforms. In Norway, for instance, they’re running three platforms simultaneously.
Currently, there is no reimbursement from health insurers for an AI platform. Hospitals or medical partnerships must therefore invest themselves if they want to purchase a platform. This also applies to individual AI applications, a budget needs to be allocated for those as well.In most cases, the purchasing process begins with an orientation on different AI platforms and drafting a business case. It is important to involve a broader team in defining the requirements, including PACS management and the IT team. Final decisions are made higher up, at the departmental board, procurement department,and hospital executive board.
We wanted to be able to compare different vendors for each AI application, both in retrospective validation settings and in prospective ones. So the platform needed toenable comparisons of models within our own patient population. Additionally, some platforms offer functionality to integrate AI results, such as measurements, directly into the radiology report. This prevents radiologists from having to dictate them manually.
Yes, definitely. Due to its focus on research, it is important for Amsterdam UMC to be able to implement and validate self-developed algorithms through the platform. On the other hand, general hospitals are more likely to finalize a business case quickly, as they benefit more directly from current AI models that help reduce workload or decrease after-hours service burdens. The PACS and EHR systems also vary between hospitals, not necessarily between academic and general hospitals, but this can still influence the choice of AI platform.
Of course, you need a legal contract. That is mostly handled outside the radiology team but is crucial for ensuring patient privacy through pseudonymization and for securing the cloud infrastructure. Often, you can also choose to run the analyses on-premises (locally), depending on the local computing capacity and the hospital’s preferences.
Once that is in place, technical integration must follow, with new connections to the PACS and EHR. I don’t yet have a full overview of that complex IT integration, and close collaboration with the IT department is essential. It will be a lengthy process, all the more reason to get it right the first time.
I mainly hope that choosing a platform allows us to keep up quickly with the latest developments and greatly enhances the scalability of AI in radiology.
I expect significant gains in reporting quality, in terms of grammar, spelling, and direct integration of AI analyses into the report. Maybe even leading to fully automated reports linked to AI results.
Additionally, I believe the current generation of AI tools is just the beginning. So much more is coming, and we want to be prepared for that. For CT and MRI, the number of AI applications is still relatively limited, but that is going to change. In five years, today’s AI apps will have evolved or been replaced by better tools. By then, you want your platform integration to be solidly in place so that you no longer need to worry about technical implementation and can truly focus on demonstrating clinical benefit and making a real impact with AI.
A variation of this article named 'AI-platform: handig, maar waarop te letten bij de aanschaf?' was previously published in Dutch in the Memorad.
