It isn't easy to implement radiology artificial intelligence (AI) at scale. In a talk at AuntMinnie.com's Spring 2022 Virtual Conference, Dr. Nina Kottler of Radiology Partners (RP) shared five lessons learned from their AI deployments.
"Artificial intelligence is something that's here to stay," Kottler said. "We should embrace it as a part of our future. And if we embrace it as part of our future, and if we embrace it and learn about it and become the experts, we can drive where this is going to go."
1. Scaling of radiology AI adoption is a nonlinear process
At RP, the pace of implementation of AI moved pretty quickly over the first few years. But then it slowed, according to Kottler.
Processes that worked initially for pilot testing programs now needed to be rewritten, she said. In addition, infrastructure needed to be developed to support high growth and adoption.
"What works for one person or 10 or even 100 doesn't work for 1,000," she said. "And you need to make it functional from a technical standpoint. Once you do that, things can take off again."
2. AI can yield unexpected findings
After deploying computer-aided triage algorithms for critical findings, RP found that the algorithms made their radiologists more efficient.
"What we found is that radiologists were more efficient because they were more efficient on the negative studies," she said. "There's many more negative studies than positive [studies]."
Computer-aided triage algorithms can also make radiologists more sensitive. Although the vast majority of these previously missed results were subtle findings, that doesn't mean they can be ignored, Kottler said.
If outcomes data can eventually show that these detections are worthwhile, those AI findings could potentially be used for triage purposes, helping both patients and the medical system, Kottler said.
AI can also provide some unexpected findings. In her talk, Kottler shared how their rib fracture detection algorithm revealed a pneumothorax in a trauma patient and also bony metastases following detection of a pathologic rib fracture. Similar unexpected findings were also found when applying algorithms for detecting intracranial hemorrhage and pulmonary emboli, for example.
"The AI would find one thing, and I, because I was very attuned to that one thing when I specifically looked at it, I could find even more," she said. "And this is where human plus AI is going to be better together."
3. It's time to invest in radiology AI education.
Radiologists need to become experts in AI, according to Kottler.
"Take your time to transition from being an early learner or an early adopter to an early expert," she said.
But this education can't be a one-and-done event.
"You need to go back and re-educate and make sure there are people actively training the radiologists and reinforcing this information," Kottler said.
4. AI needs to be integrated into the radiology workflow.
There are three viable options currently for integrating AI into the radiology workflow. AI results are sent to the PACS, the PACS "calls" the AI viewer, or the AI viewer is embedded in the PACS, according to Kottler.
"Hopefully in the future there will be standards like maybe the FHIR standard that we're using or other standards that will allow us to integrate and work with the information that we need within our PACS systems," she said. "We're just not there yet."
5. A platform is needed to deploy AI and automate workflows.
A platform needed to utilize AI at scale, Kottler said.
"The platform has to be able to take all of this unstructured data that we have in radiology at massive scales and be able to move it in real-time, and that is not easy," she said.
This platform needs to send imaging data to the AI system and then downstream to the radiologist or another user. That requires two levels of orchestration: one to ensure that the right imaging data is being sent to the right AI system and a second layer to ensure that the AI results get to where they need to be.
This orchestration task can be best handled via a cloud-native platform better than an on-premises server, according to Kottler.
"It's the only way to really scale up and scale down resources as you need them," she said. "It's not based on total volume; it's based on the volume you're sending in any second. In any second you could be sending thousands of instances through at your scale, and you need something that can manage that."
