While many people are familiar with AI being used to create fake photos and videos for entertainment purposes, AI technology can also have lifesaving applications.
Doctors are now using AI technology when analyzing CT scans, X-rays and additional imaging.
Dr. Christopher Connor, medical director of radiology for Mercy Hospital Jefferson and physician lead for Mercy Imaging Specialty Council, said Mercy radiologists have been working in tandem with Aidoc to help prioritize urgent cases for about two years.
Connor said the type of AI that radiologists typically use is a triage algorithm, which looks at images and reprioritizes the images it has flagged as having possible positive results to be higher up on doctor’s worklists to be looked at sooner.
If the triage algorithm indicates that there is a positive finding, the radiologist reviews the image to see if they agree with the positive finding, Connor said. If so, the patient may be asked to get further imaging. However, all images, whether flagged for a positive result, are still thoroughly reviewed by a radiologist.
“All the cases still have to be read, but it moves the cases that are positive higher in the list, so they get read sooner, which is important, especially when you’re dealing with critical findings. If we get to them sooner, that’s better because then they get diagnosed sooner, and then they can hopefully come into the hospital before something bad happens.”
Connor said this is especially helpful for routine outpatient cases which doctors may not be expecting emergent findings.
“If someone has an unknown fracture, like a neck fracture, they could get paralyzed if they keep walking around,” Connor said. “Or a pulmonary embolism, which is when you have a blood clot that goes to your lungs, and that can kill people.”
Connor said AI can be good at spotting blood vessel occlusions, brain aneurysms, small fractures, such as cervical spine fractures, and pulmonary embolisms. These issues may be delayed or missed by a radiologist, especially if those issues are unrelated to the treatment the patient is receiving.
“There are definitely some people who have had routine outpatient chest CTs that have had their pulmonary embolism detected sooner,” Connor said.
A lifesaving catch
One such case is a 55-year-old woman who came to Mercy for help with unexplained weight loss. She had a CT scan of her abdomen ordered by her GI specialist to investigate the cause.
At the very top of the image in the thin sliver of lung that appears when the abdomen is scanned, Aidoc flagged a pulmonary embolism. That detail may have slipped past even an experienced radiologist because it was unrelated to the reason for the imaging.
Immediately after her scan finished, Aidoc examined the image and flagged it for priority viewing for the radiologist. Prior to Aidoc, scans at Mercy, like most other hospitals and health systems, were typically read in the order they arrived. A routine outpatient case like hers could have waited a day or longer before it reached a radiologist. Now every scan is immediately examined by Aidoc and evaluated for triage.
A follow-up scan confirmed the potentially deadly pulmonary embolism, which allowed the woman to receive emergency treatment sooner and led to a full recovery.
A news release from Mercy stated that for the most urgent cases, the time from scan to result decreased from about 19.8 hours to just 1.3 hours because the triage algorithm gets those images in front of the radiologists’ eyes first rather than leaving them in a chronological queue.
Connor said he’s understanding of people who have a cautious skepticism about the use of AI in medicine, and he said those concerns should be considered.
“We’re on the cutting edge of a deployment of a pretty new technology, so there will be adjustments we’re going to have to make. It’s important to be responsive to those changes because there’d be nothing worse than to ignore real complaints and concerns that have validity, and then that did negatively impact people. We take that very seriously, and we are evaluating those regularly.”
Connor said it’s important for people to realize that medical AI does not work in the same way as other AI. He said the AI algorithm used in radiology is not the type of AI that learns and improves over time.
“It’s actually created and frozen in a box and then cleared by the FDA. The algorithm cannot change,” Connor said. “If Aidoc wants to change the algorithm, it must go through FDA clearance again.”
Connor said radiology has a history of being proactive with new technology and being early adopters. He said radiologists jumped on board with PACS (Picture Archiving Communication System) for medical imagining when it became available. Before PACS, when medical images like x-rays, CTs, MRIs, ultrasounds or other imaging needed to be captured, stored, reviewed or shared, that meant film, physical folders, storage rooms and couriers.
Connor said the transition to digital was a huge step forward.
“PACS was a huge transition, and it’s very analogous to the transition to practicing without and with AI. The productivity actually increased about 50 percent after the introduction of digital PACS,” Connor said. “I think that’s kind of the long-term solution for what AI is going to do in radiology.”
Connor said Aidoc does not replace any doctors but instead is a tool that goes hand in hand with better results to make sure nothing falls through the cracks and no diagnoses are missed.
Connor said Aidoc does not give patients diagnoses on their medical records or write reports, nor does it recommend treatment plans. Human clinicians are still in charge at every step.
“The physician is the final check, the final safeguard on the process,” Connor said.
He said AI simply reduces the amount of re-checking performed on images by providing a second set of “eyes.”
Connor said radiologists often review images over and over, and having AI coming to the same conclusion can reduce the number of times doctors re-review images.
Time saving is especially crucial in radiology, where staff shortages are extremely common, he said.
About a decade ago, many people in the tech industry were predicting that technology was going to advance so quickly that radiologists would no longer be needed, and many medical students were discouraged from going into radiology, Connor said. As a consequence, hospitals are facing shortages of radiologists, which are still very much needed despite advances in technology.
One 2024 industry estimate showed roughly three open positions for every one available radiologist. At the same time, those doctors are looking at an increasing number of images every year. Mercy Hospital reported imaging volumes were growing 12 to 15 percent annually. This left radiologists unable to keep pace because they were needed to review more and more images every day without the additional time or help to do so. This can create reading backlogs and slower diagnoses and treatment decisions.
Connor said AI has a lot of things it’s very good at, but it has other areas where it’s not as good. He said the algorithm sometimes has “false positives” and highlights images as having potential red flags that are completely normal, such as flagging blood vessels as being nodules or classifying benign brain tumors as brain hemorrhages. AI also has trouble evaluating images where a person has moved in the scanner and may identify hemorrhages that are not there.
“Aidoc makes some calls that are wrong. They are not something any human would think is true. That being said, it does also make some subtle calls that someone might miss,” Connor said.