Further reading recommended by Daniel Forsberg:
The present state of AI in radiology imaging
As we begin the conversation, Daniel explains that the hype surrounding AI in medical imaging has obviously peaked. “This is exciting to me as we can finally move beyond the hype to a point where all parties, including healthcare organizations and staff, researchers, and vendors, are ready to have open and honest conversations about AI and how radiology can benefit from it, not just based on what we have seen in slick demos or eye-catching presentations,” he says.
Nynke agrees with Daniel but also emphasizes the significance of properly setting expectations for those who begin adopting AI: “People underestimate the fact that AI is still in an early stage. If you look at the rise of PACS or the first CT scan, it took many years before they were widely used and approved.” She continues, “Having said that, AI currently has several useful applications for use cases that are already generating value in clinical practice today. For example, AI can be used to find a lesion, quantify a tumor, or find an acute brain bleed. These are narrow use cases, but if you integrate them properly into your existing clinical workflow, they can generate substantial value in your clinical practice. I think that’s where the industry is today.”
This is where we are now: with every application and every study carried out, we are building the foundation of AI in radiology imaging. We are creating the stepping-stones for the broad-scale adoption that we know will happen.
According to Daniel, AI in medical imaging is currently facing some really tough questions, such as the issue of “model generalizability (i.e., how well a model trained on one dataset would perform when applied to another dataset), tangible and measurable values, and building business cases to show the return on investment (ROI) and justify the investment.” In an attempt to answer these questions, several interesting research projects/studies are now under way in a prospective setting that will help shed light on some of the questions raised. When asked to name a few, Daniel mentions two studies that are investigating the use of AI for breast screening: Artificial Intelligence in Large-scale Breast Cancer Screening (ScreenTrustCAD) and Mammography Screening With Artificial Intelligence (MASAI).
To conclude, Nynke shares a metaphor that, according to her, reflects the current state of AI in the medical imaging industry: “AI development and adoption is best explained by comparing it to a construction site. You see the frame but definitely not the whole building yet. This is where we are now: with every application and every study carried out, we are building the foundation of AI in radiology imaging. We are creating the stepping-stones for the broad scale adoption that we know will happen.”
Opportunities and benefits
When asked what benefits AI can provide to radiologists, Daniel sees many potential benefits but believes two of the most pressing questions in the radiology AI business right now are determining what value will be created and how to measure it. Before we can honestly say anything about the greatest potential value, both of these questions must be answered. However, there is one area that is rather straightforward and has a lot of potential—the use of AI as the first reader for breast screening (considering the European setting of double reading) for exams with a very low risk of suspicious findings. “This could be a game-changer in countries where there is a shortage of breast radiologists,” Daniel points out. Similar kinds of use cases where AI can cut the workload significantly include lung screening and stroke cases.
I believe that the real opportunity with AI is that if adoption is carried out successfully and once you’ve made an impact, you should look at it from a workflow and organizational perspective rather than a purely technological one.
For a long time, the output supplied by AI applications, or any image analysis tool, has primarily focused on detection and characterization of findings to assist the radiologist during image review. AI certainly has a role to play here, but what Daniel finds interesting is how the output from AI applications can be used throughout the radiologist’s entire workflow.
Nynke agrees: “I believe that the real opportunity with AI is that, if adoption is carried out successfully and once you’ve made an impact, you should look at it from a workflow and organizational perspective rather than a purely technological one,” she says.
Looking at AI integration from an end-to-end workflow perspective rather than a strictly technical or clinical technology viewpoint, there are a number of things you can do. “And of course, it all still depends on the purpose of the application. However, AI could help with your prioritization of work, telling you what exam you need to look at first. AI could also help with the allocation of resources, like delegating complex cases to more experienced professionals and simpler cases to juniors. It could help you create more consistent reporting in your departments since everyone would receive the same output from the AI and, therefore, the overall quality of the reporting would improve, making it less important to referring physicians which radiologist has reported a particular case.”
“Another added value if you start with AI today,” Nynke continues, “is that you can start learning what it means for the organization to adopt technology like AI, and you can get your department AI ready. There are a lot of things you will run into once you start adopting AI in clinical practice. New roles will emerge, and new processes will need to be developed. Even though you might not see the full value today, it could still be relevant to implement this type of project just to understand what it means for your department to start working with this technology. Alternatively, you can wait until this technology matures, of course. But if you wait a few years, then you will still have to start by figuring out what implications it has for your department, so you need to make sure you prepare for the widespread adoption we all know will happen. Innovation adoption—that is a skill you need to learn,” Nynke concludes.
Building a future-proof AI portfolio
Many health systems have trialed one or two AI applications in some way. “I think that a common experience for them has been the challenges associated with deploying and integrating the AI applications,” says Daniel. He feels that experience from working with individual applications has been “quite cumbersome and many find the whole aspect of scaling the number of AI applications used quite daunting.” This is exactly where the Sectra Amplifier Marketplace comes in. The purpose of Sectra Amplifier Marketplace is to offer a single platform, and one point of integration to gain access to a multitude of AI applications. The platform handles deployment and scaling of computing resources, which makes the whole process much easier.
“Sectra Amplifier Marketplace is the starting point to making AI more accessible and it’s very much in line with where we think the industry is right now. We will continue to elaborate and extend that portfolio, working together with customers and partners and pursuing validation in clinical practice,” says Nynke. However, the ROI needs to be shown to create a business case. “The health economic evidence to validate the impact of AI on clinical practice, whether from efficiency or an economic standpoint, is still in the early stages. Combined with some AI vendors’ price points, I believe this is what makes people hesitant to even try. Because sometimes, all you want to do is test something and see whether it works. We are establishing Sectra Marketplace to help with this since it allows you to try different applications without making a major investment,” says Nynke. Sectra Amplifier Marketplace can help you save both time and money by allowing you to try different apps without needing to set up a new contract, a new server, or new integrations. “One of our main aims is to make sure the cost of switching is low and to show that you don’t need to become locked into an individual AI vendor. The AI applications that are beneficial today might not be good in five years’ time. To avoid this, we provide our customers with the flexibility to test out and change between applications so that they can find what suits them best. And that gives you confidence in your ability to build a future-proof AI portfolio for your department,” says Nynke.
AI will help with more monitoring and with repetitive tasks. I believe it will even get to the point where AI will take on the role of first reader or autonomous reader in some cases, and that will put even more pressure on healthcare to have high-quality governance and organizational embedding in place.
How to prepare for the future
“I think that radiologists definitely will not be out of work. Some people say that that’s the way it will go, but I don’t think so. I believe they will be empowered by this technology,” says Nynke. “I have high hopes for AI in medical imaging. I always think about the 80-20 rule, so AI will assist in 80% of the radiology business and that’s a lot, right? AI will help with more monitoring and with repetitive tasks. I believe it will even get to the point where AI will take on the role of first reader or autonomous reader in some cases, and that will put even more pressure on healthcare to have high-quality governance and organizational embedding in place. I believe that radiologists will perform more specialized tasks, the ones that require more attention. But AI will surely make an impact on large volume workflows and in that sense, I think it will be a big support.” She emphasizes the importance of having your department tailored for AI adoption. “This is always the case with technology adoption: we focus on technology and money, which are important, but never focus on the soft aspects such as people. I believe there was once a change management consultant who said culture eats strategy for breakfast.”
When planning AI adoption, it is important to think about workflow, motivating stakeholders as well as specifying the end result and how to track it. “Cultural organizational aspects are just as much a part of the adoption process. In early-stage technology, you need to be able to say, ‘This didn’t work.’ It doesn’t mean that nothing works, it just means this isn’t working so we need to do something else,” says Nynke.
When we ask Nynke what the most important success factor is for the acceptance of AI in radiology, she replies, “Enthusiasm and trust. Since this is an early technology, you need to have a certain amount of enthusiasm and trust in the technology and trust that this innovation will deliver value in the long term. Because, in the end, the human aspect will be the driving force behind the change. It will be up to the human factor—the people involved—to advocate for change and to convince their peers and stakeholders that this is the correct course of action. Sectra can assist in the sense that we can say, ‘These are applications that will create value, so you can trust them, and we can assist with the technical integration and the adoption of AI applications for you to try out.’ However, a spark in the healthcare institution is still required to fuel this shift.”
AI in medical imaging is here to stay. As with all innovative technologies, it comes with its own set of issues and necessitates the involvement of some early adopters who are willing to incorporate solutions without knowing all the answers. But one thing we are sure of is that the implementation of AI applications offers great potential to help radiologists with their workflow and to prevent a crisis in digital medical imaging. Along with an opportunity to analyze the actual ROI, AI is highly likely to revolutionize the workflow of digital medical imaging in radiology.