With the dual targets of shifting from episodic to proactive, long-term administration of sufferers with continual illness and decreasing doctor burden, radiologists and IT specialists at Penn Drugs in Philadelphia have developed a synthetic intelligence answer, PennAInsights, that automates the workflow from picture seize by means of AI evaluation to diagnostic reporting. For its potential impression and multidisciplinary nature, Healthcare Innovation acknowledged this undertaking as considered one of its three Innovator Award winners for 2025.
As Charles Kahn, M.D., M.S., professor of radiology on the Hospital of the College of Pennsylvania, explains, conventional radiology workflows are labor-intensive and fail to seize refined early modifications. “With the expansion in process volumes and the smaller development within the variety of radiologists, in some ways, the work outstrips our potential to do it,” he mentioned. “As well as, there may be much more data in a typical imaging research than any human doctor is essentially going to have the ability to extract from it. So the notion is that with AI, there may be lots of data that may very well be extracted from on a regular basis imaging research that may assist help the care of our sufferers.”
The target for creating the PennAInsights platform was to develop an automatic answer that extracts quantitative measures from current imaging research to enhance diagnostic accuracy, scale back doctor burden, and allow early intervention.
Right here is how PennAInsights works: Encrypted photos are transferred from the PACS system to a HIPAA-compliant cloud AI server, the place a collection of validated AI fashions course of the research and return quantitative annotations as DICOM structured experiences straight into the radiologist’s workflow. In pilot purposes, it addresses points corresponding to belly adiposity, liver steatosis, and mind atrophy.
The Penn Drugs radiology and IT execs stress that they had been seeking to create not only a single software however a platform for different AI developments. The answer integrates seamlessly with current PACS and dictation methods by means of a cloud-based, scalable, and safe platform. Its “plug-and-play” modular design permits speedy incorporation of latest AI fashions to handle numerous medical wants.
Walter Witschey, Ph.D., affiliate professor of radiology on the College of Pennsylvania, explains their considering. “Should you have a look at the variety of FDA-approved AI instruments within the subject of radiology, it is huge. However only a few of them are literally applied in medical care as an end-to-end workflow. They’re fairly burdensome to make use of, not interoperable, and never built-in with the well being system. None of those options actually supply what we’re in search of, so far as a totally built-in, interoperable AI answer,” he mentioned. “So we thought we might attempt to construct it ourselves. We’ve the know-how. We’ve been working for a very long time with the technical particulars of photos, for example, the DICOM normal and interoperable connections between units just like the HL7 messaging system. We additionally had some cool AI purposes that we had deployed in analysis that had been very efficient. We’ve deployed them within the Penn Drugs biobank. I believe on the time we began, it appeared fairly dangerous, however as I look again on it now I believe we should not have achieved it every other means.”
One instance of a situation this platform has nice potential to handle is fatty liver illness. “It is essential as a result of individuals who have irregular fats accumulation of their livers are at elevated danger for quite a lot of situations. In extreme instances, individuals get tumors within the liver, or they could want a liver transplant,” defined Kahn who is also vice chair of the Division of Radiology.
There may be an rising prevalence of the illness, and other people do not essentially know that they’ve it, so offering early detection for this situation is a chance to catch the illness early and assist individuals mitigate it and probably save lots of healthcare prices downstream.
“You will discover it on a specific research, however not essentially in its very early kind. That is the place this AI-based detection may help make us extra delicate and be sure that it turns into a routine a part of of the data that is captured,” Kahn added. “So even when you got here in for suspected kidney stones, we will have a look at issues just like the diploma of fats within the liver, or is the liver enlarged? We are able to have a look at bone mineral density to see when you is likely to be osteoporotic. We are able to search for coronary artery calcifications which may point out that you simply’re in danger for coronary artery illness.”
The undertaking started in 2019 however noticed some delays in the course of the pandemic. It went stay in Might 2023. The hassle was multidisciplinary with robust help from IT groups. “From our perspective, it is very important perceive the technical necessities and be sure that it goes by means of the correct governance course of from an infrastructure and safety standpoint and from an information standpoint,” mentioned Anna Schoenbaum, D.N.P., M.S., R.N.-B.C., vice chairman of purposes and digital well being at Penn Drugs. “We accomplice with the suppliers right here to ensure it meets the necessities, to do the testing and to be sure that it’s executed safely.”
Ameena Elahi, IS software supervisor at Penn Drugs, mentioned her workforce regularly seeks modern methods to eradicate workflow delays by automating processes the place potential and maximizing the usage of instruments inside their current infrastructure.
Pilot research outcomes
In a six-month pilot, Penn AInSights processed 2,973 belly CT scans with a median complete turnaround of two.8 minutes. They discovered that 94.9% of research had been accomplished in beneath 5 minutes, making certain that every one AI annotations had been accessible earlier than preliminary report evaluation. Radiology experiences now persistently embody quantitative measures that beforehand appeared in solely 24% of mind MRI experiences addressing neurodegenerative dangers. The pilot workflow operates at roughly $700 per thirty days (lower than $1 per affected person), demonstrating important value effectivity. Early intervention in situations like fatty liver illness and cognitive decline is anticipated to cut back downstream problems and general healthcare prices.
Because the completion of the profitable pilot, the platform is now getting used broadly throughout the hospitals which might be a part of Penn Drugs. The photographs go to a centralized PACS system and get analyzed by the software program. “We’ve analyzed greater than 20,000 research thus far, and we’re within the means of increasing to different areas of the physique, wanting on the chest in addition to some instruments that can combine newer advances in AI,” Witschey defined.
As a subsequent step, Penn Drugs clinicians are exploring integration with giant language fashions (LLMs) to routinely construction unstructured radiology report findings—corresponding to detecting adrenal nodules—to set off medical resolution help in Epic. This extension, at present in a testing surroundings, guarantees to additional streamline reporting and affected person administration.
“Utilizing LLMs to construction unstructured radiology report findings will facilitate real-time medical resolution help, streamlines radiologist workflows, enhances diagnostic consistency, and boosts general effectivity,” Elahi defined.
Kahn mentioned there are even higher prospects. “We’ve this wonderful useful resource referred to as the Penn Drugs Biobank. We’ve the flexibility to hyperlink the varied findings that we derive from imaging research to the genetics and to the opposite healthcare traits — the phenotypes of those sufferers,” he mentioned. “Right here’s the place we will join not simply the day-to-day radiology workflow points, but additionally join imaging data to the genetics and gene expression traits of our sufferers, and join that to their broader well being profile. That is an interesting space.”
Kahn famous that different tutorial medical facilities are additionally engaged on making use of AI to radiology workflow, however he mentioned Penn Drugs has lots of the important thing constructing blocks in place to make this occur. “It’s a must to have strengths all throughout the pipeline,” he mentioned. “It’s a must to have a extremely robust analysis workforce — not solely Walter and his workforce, however the Penn Drugs Biobank and the imaginative and prescient that led to that. It’s a must to have a extremely robust IT workforce, and we’re tremendously grateful for Anna and all of the help that she and her workforce have given us to make this factor potential. After which you need to have individuals on the medical aspect who’re excited by bringing this ahead to assist serve our sufferers.”
A video overview of Penn AInSights is on the market right here.
