Insights
When Enterprise Imaging Met Artificial Intelligence
Two years ago, Aalborg University Hospital in North Denmark found itself navigating the hazardous seas of backlogged imaging exams. Looking back now, it’s easy to identify the crosswinds that combined to create that perfect storm.
A protracted radiologist shortage. A gathering rise in the number of scanners in service. A steady uptick in imaging orders placed by clinicians practicing defensive medicine. And, maybe most vexing of all, a historic rebound in overall hospital traffic as the COVID-19 pandemic receded.
Since then, the hospital’s imaging volumes have settled into a new, albeit much higher, normal. In fact, big numbers have become the order of the day. From 2020 to this year, for example, the hospital saw CT and MRI scan counts grow 35%and 40%, respectively.
The good news is that, despite the heavy load, imaging operations at Aalborg University Hospital are no longer taking on water.
Referrers are ordering imaging appropriately and confidently. Patients are receiving results without undue delays. And both those developments are solidifying into business as usual. That’s largely because radiologists are getting their reads completed with high accuracy and fast turnaround times.
What tamed the tidal surge? The synergy of artificial intelligence combined with enterprise imaging.
“We will stand down one radiologist until 30% to 35% of examinations are read by one doctor rather than two. At the same time, we will do quality assurance for a time, seeing how the AI performs against a second radiologist reader. We think this will be especially important in mammography.”
– Peter Buss Lasborg, Head of Radiology Department, Aalborg University Hospital, North Denmark
Warding off needless reads
For Peter Buss Lasborg, the head of Aalborg University Hospital’s Radiology Department, those two technologies—AI and EI—have become inseparable from one another. Given the learnings that led him to this view, his department’s experiences building an AI architecture inside an EI infrastructure may help inform other practices looking to adopt both sources of digital transformation at once.
“What we have is a very straightforward EI infrastructure,” Lasborg tells Radiology Business. “It only involves a DICOM object getting transmitted from the modality to the AI software and then on to the PACS. Radiologists can view the AI read as an image overlay or however they prefer to see it.”
In the case of the first indication for which Aalborg University Hospital has tapped AI, orthopedic fracture detection on X-rays, the algorithm separates patients who may have a fracture from those who definitely don’t. Positive or likely positive findings on AI prompt an application to send the images to a radiologist’s work list for expert interpretation. Negative findings automatically let the care team know that the patient can be treated conservatively and released without direct attention from a physician.
The hospital is currently expanding the EI/AI combo from fracture detection to screening mammography. For this use case, negative screens via AI will obviate the need for a radiologist overread and/or additional imaging.
Next will come pneumothorax detection on plain X-rays. Lasborg notes that projected end-users are open to the idea that AI may only be used for residents and other trainees.
“We’re not sure the AI will really benefit experienced body radiologists because they’re going to read all those images either way, positive or negative,” he says. “We will count on our radiologists to have higher sensitivity and specificity than the pneumothorax AI system. However, for younger doctors, especially in the ER, the AI reads should provide a good educational aid.”
On deck and awaiting its turn is MRI prostate AI, along with other indications yet to be named.
Straight line from PACS to EI
At its core, Aalborg University Hospital’s EI infrastructure is a mature PACS that first went into service way back in 2006. The PACS, a reliable workhorse from Canon Medical Informatics, has been serving seven partner hospitals practically all along. That made it a de facto—and pioneering—EI system the first time it powered up.
The transformation has continued as Canon Medical Informatics has built out and refined its PACS line into a leading-edge EI line.
Keeping up with Canon Medical’s PACS and EI updates and upgrades over the past 17 years, Aalborg University Hospital had the EI well ready for AI when the latter was needed during the imaging overload of 2021.
Lasborg suggests one of the top benefits of architecting AI inside a stable EI infrastructure is nimbleness. For example, as new clinical indications come online for AI to help with, the infrastructure will let staff simply “pop in” new AI software packages.
There’s a bit more to it than that for both IT staff and end-users, of course, but the complementarity of the two technologies means there’s little to no need for new hardware apart from storage solutions. Even then, the cloud may suffice as a labor-unintensive option.
Aalborg University Hospital is early in its AI/EI journey, with only radiology and its referrers aboard so far. But cardiology will soon join the journey. Then will follow, slowly but surely, imaging-intensive clinical departments using non-DICOM data protocols—dermatology, ophthalmology, wound care, endoscopy and point-of-care ultrasound.
For now the plan is to let AI relieve one radiologist of reading duties per imaging exam.
“We will stand down one radiologist until 30% to 35% of examinations are read by one doctor rather than two,” Lasborg says. “At the same time, we will do quality assurance for a time, seeing how the AI performs against a second radiologist reader. We think this will be especially important in mammography.”
“One of the biggest differentiators between enterprise imaging and standard PACS, at least for end-users, is fast and secure collaboration between radiologists and referrers. Without EI you have silos. You have a bunch of departmental systems housing their own copies of data and propagating their own data subsets.”
– Tim Dawson, Chief Technology Officer, Canon Medical Informatics
Best yet to come
Looking ahead, Lasborg foresees a future with, for one thing, fewer imaging orders. He hopes the success of the EI/AI combination will embolden referrers to ease off the “just in case” impulse. Instead, in a best-case scenario, most will return to trusting their diagnostic instincts: They’ll order imaging to look for specific ills suspected from patients’ presenting symptoms and spoken complaints.
Lasborg also wants to increasingly tailor Aalborg University Hospital’s AI models to fit North Denmark’s own population.
“Most of the AI systems available today have one input-output mechanism, and that’s all you can get,” he says. “A lot of these models are good enough as starters, but they’re trained on patient data from outside our geographical area. It will be more highly beneficial when we have many more deep learning algorithms trained on—and continuously refined with—datasets from the patients we actually serve.”
High tech, high touch
At Canon Medical Informatics, where Aalborg University Hospital’s EI system was born, Chief Technology Officer Tim Dawson offers insights into the hospital’s success merging new AI architecture into an established EI infrastructure.
“One of the biggest differentiators between enterprise imaging and standard PACS, at least for end-users, is fast and secure collaboration between radiologists and referrers,” says Dawson, who’s been with Canon Medical since 2010. “Without EI you have silos. You have a bunch of departmental systems housing their own copies of data and propagating their own data subsets.”
This presents both clinical and data-security challenges for IT staff. The trick is making sure collaborating physicians can move data without losing any important clinical context or overriding any security measures, Dawson points out.
The key to making it all work is building both EI and AI on new web-based protocols rather than the older DICOM protocols. “This enables us, even on demand, to cache data at the edge rather than centrally,” Dawson explains. “Central caching leaves you open to a major failure if, for example, a natural disaster knocks out local or regional power for an extended period.”
Unstinting focus on care
Dawson has much more to say about the technical ins and outs of EI and AI, but he’s just as keen on the softer skills that, he says, Canon Medical Informatics instills in its people.
“Not long ago I interviewed a technically skilled job candidate for a fairly high-level IT position,” he explains. “I asked why he was looking to stay in healthcare rather than move to another industry. He answered by telling me he’d recently brought in his son for care and had a chance to view the imaging.” When the care team showed him the scans, he saw they were using a PACS he’d had some hand in designing and developing.
The candidate convincingly conveyed to Dawson his excitement over the experience.
Dawson offers: “He’s exactly the kind of person we want to employ at Canon Medical Informatics.”
Lasborg is similarly forward-facing. “We’re always striving to improve care for the patients, families and communities we serve,” he says. “Technology is only part of our continuous improvement mission, and EI and AI are only part of our total technology picture. But they’re a big reason behind our confidence as a leading provider of healthcare services in Denmark.”
The imaging seas are calm. The radiology ship is righted. Aalborg University Hospital is trimmed to maneuver apace for years to come.