AI frees up valuable time for radiologists in a Swedish healthcare region

Can AI help us free up time and reduce mounting pressure?

Breast cancer is the most common type of cancer in the world and its incidence is increasing. At the same time, there is a growing shortage of breast radiologists in many countries, including Sweden, leading to mounting pressure on existing physicians. The mammography operation in Region Värmland is no exception, and in 2020 there were internal discussions about AI being a key piece of the puzzle in terms of better meeting these challenges. In Värmland, around 30,000 women undergo mammograms every year, and the images are double-read by two breast radiologists independently of each other—a procedure that is highly time-consuming. The need for effective medical IT solutions has never been greater.

“The single biggest driver behind our decision to start working with AI-based diagnostic tools at all was definitely the genuine shortage of radiologists in the field of mammography,” says Jonas Söderberg, Deputy Head of Imaging and Functional Diagnostics and County Medical Officer at Region Värmland. “We saw challenges with understaffing in our organization that we thought could be addressed using new technology, viewing this as an opportunity to free up valuable working hours for our extremely busy radiologists.”

Once the clinical use case had been identified and the region had formulated the problem they were trying to solve, it was time to find a suitable tool. They investigated the options available in the market, including at a visit to the major European breast radiology trade fair EUSOBI, to come up with a shortlist prior to evaluating a number of different algorithms. Ultimately, they chose the Transpara® AI application, developed by the Dutch company ScreenPoint Medical, supplied via Sectra Amplifier Services. One of the decisive factors was the advanced integration of the application with Sectra PACS—seamless integration is essential for both efficiency and security.

A structured implementation project

Getting the green light within the organization turned out to be a long process, as the “structured implementation” approach was used—as is always the case when new medical technology is to be introduced. In hindsight, this is seen as very positive by those involved, particularly when it comes to ethical challenges linked to the use of AI, such as the possibility that the AI application could produce inaccurate results and lead to incorrect treatment.

“When we were told internally that we would have to go through the process of structured implementation, it felt like an uphill struggle because we knew it would take time. But in retrospect, really getting to the bottom of the ethical issues, for example, has been positive,” says Jonas.

“It’s actually the same process, no matter what new device or technology you’re introducing in Swedish healthcare,” adds Mikael Fredholm, a Medical Device Engineer in the region. “Machines can make mistakes, but there must be processes in place to ensure that you get consistent, high-quality results over time.”

Because of the process, the region also conducted its own analysis during the implementation project.

“We were required to demonstrate that the algorithm was at least as good as a physician,” says Mikael. “We had looked at studies from Oslo[1] and elsewhere, but we also wanted to test it out on our own population. So, we conducted a retrospective study at the time of implementation looking at three months of production and the correlation between detected pathological findings and a low score from the algorithm. The study showed that less than 5% of the findings had what is known as a ‘Transpara score’ below five, which is the threshold for cases to be single-read.”

He continues: “We intend to redo the corresponding study on a recurring basis to see if we can fine-tune parameters in our implementation and to see if the doctors’ assessment is influenced by the tool. It would be good if there was an accepted method to repeatedly ensure that the algorithm continues to do what it is supposed to and maintain high quality, in the same way that we check other equipment regularly. Such a method may be developed by an AI supplier or a public authority.”

The importance of an overall assessment

However, getting the technical side of the solution in place went more quickly. As previously mentioned, the decision was made to purchase and implement the AI application via the cloud-based service Sectra Amplifier Services, which Region Värmland was the first to do in Scandinavia. The application is closely integrated with the existing Sectra PACS and existing workflows, providing a seamless experience for the user, who can work in a single interface with no switching or extra logins.

“We want to minimize the number of ad hoc solutions throughout the region. It is quite easy to acquire a new device or application for a minimal area of application. With this new approach, we have a solution that is well integrated into the business system from Sectra that we already use. Once we got the green light, the project went very well and Sectra was very much on the ball regarding their part of the implementation,” says Jonas.

Mikael adds: “We made an overall assessment of the entire solution, and as we have a data processor agreement in place—which encompasses analysis of anonymized cases, encryption between the parties involved, and requirements that the images must be deleted after analysis and that it must be impossible to identify the woman in the images—the region is OK with this solution. If we chose to go with another AI application, we would make a new analysis, even though the Sectra platform as such has been analyzed.”

“Another important aspect for us was the financing model,” Jonas continues. “When implementing new products or services, large upfront payments are inconvenient. We appreciate the model of paying per exam. It makes us feel secure about testing new technologies and increases our willingness to adopt new solutions. We’re fine with paying for something we use a lot, but if it’s not well received by users and not used that frequently, we don’t want to be stuck with a large annual fee or a long contract.”

Human and machine in symbiosis

Katalin Szögyi is a Chief Physician at the Breast Center at Karlstad Central Hospital and one of the breast radiologists who has been involved in the introduction of Transpara in Värmland. With the AI tool seamlessly integrated with the Sectra diagnostic imaging application, Katalin and the other physicians get help in prioritizing examinations that require their immediate attention when time is short.

“The AI tool can help us pre-sort all the examinations. Using a ten-point scale—a ‘Transpara score’—they are classified according to cancer risk, allowing us to focus on high-risk cases first if there are time constraints,” Katalin explains.

In addition to presorting, they use Transpara to support them in making decisions when suspected findings and their corresponding “AI points” are flagged using the function for computer-aided detection (CAD) in Sectra’s PACS. Human perception is then combined with computerized image analysis—human and machine in symbiosis.

“We’re very satisfied with this new way of working,” continues Katalin. “It’s a little hard to see why the algorithm sometimes gives a low score to something we physicians have trouble examining, such as a dense breast, and gives a high score to something that we, experienced physicians, can immediately see is healthy. But in general, I find that having the algorithm as a tool gives me a great deal of confidence. It never gets tired or distracted, and always delivers a consistent, high-quality result. We also often use the CAD function and take an extra look at the area that is marked as having the highest risk of findings.”

She says that it was easy for them to get started with the new procedure and that she feels the working environment has improved for the radiologists overall. “This way of working has now become completely natural for us and I would personally be sad if we no longer had access to the tool, because it really makes things easier and less stressful, and it allows me to feel more confident in my decisions.”

In the future, Katalin hopes that AI applications like Transpara will also take historical data into account, which she believes could reduce the number of “false positives” from the algorithm. “That’s often how we work to see if there really is a change, or if what may look abnormal is normal for that particular woman,” she concludes.

The way forward for Värmland and for AI in diagnostic imaging

What does the future look like for the use of AI specifically in diagnostic imaging in Region Värmland? Jonas says that currently they are mainly interested in AI algorithms that solve problems they either come across frequently or that arise in a stressful environment.

“Different types of lung screening tools are therefore welcome, because of the sheer quantity,” he explains. “Where ‘saving the brain’ in acute cases of stroke is concerned, for example, the situation is time-critical and new tools are also of interest in this area. Another area is prostate MRI, where the staff shortage combined with an increased volume of examinations would make an AI tool a valuable addition. We hope that we won’t have to do all the work on market research in these areas ourselves, but instead that Sectra can help us by pointing out the appropriate algorithms that are available.”

Finally, we ask Mikael and Jonas to share some practical advice or recommendations for other healthcare regions that want to explore this area.

“It doesn’t really matter what you’re introducing, the most important thing is to get the organization on board with the project,” says Mikael. “The physicians have always been positive in our implementation project, but we’ve also adapted the integration in Sectra PACS to their specific wishes.”

Jonas particularly stresses the importance of starting the internal process early—because it takes time. “It’s also a challenge in that everyone always must do a feasibility study before really getting started. Can we standardize this between regions? In this case, Värmland was first, but I think we need to be able to trust each other more to avoid all regions having to make the same long journey,” he concludes.

References

1. Artificial Intelligence Evaluation of 122 969 Mammography Examinations from a Population-based Screening Program