AI inches into the real world and delivers benefits in radiology
Improving efficiency and coordination offer some immediate applications, with clinical potential emerging as well.
Artificial intelligence in healthcare still seems to exude some sort of future world, “it’ll happen someday” vibe.
That’s far from the truth in radiology. Notably, there was a “the future is now” undercurrent at the recent annual conference of the Radiological Society of North America in Chicago.
AI is already delivering benefits for healthcare organizations, with more expected to come, especially in tandem with other technological advancements occurring in imaging. In sum, the hope for AI is to increase radiologists’ efficiency, relieve work burdens, and enable the use of technology to achieve proactive and preventive care.
AI as more than an add-on
In the past, radiologists have had to navigate between different technology stacks, running the gamut from wide varieties of imaging devices, picture archiving and communications systems (PACS), radiology information systems, and more.
And formerly, AI for radiology has been applied at different stack levels, not always fitting neatly into workflows. At this year’s RSNA show, there was an emphasis on designing systems that engaged AI throughout the imaging cycle, increasing accuracy and efficiency to better backstop and support radiologists.
Radiology suffers from inefficiency and inconsistency that has downstream effects on quality, notes Sham Sokka, chief operating officer and technology officer for digital health at RadNet. “There’s an urgent need to navigate clinical, financial and operational challenges,” he said. “It results in disconnected patient engagement, a strained workforce and inconsistent clinical outcomes.”
Two years ago, Los Angeles-based RadNet, which provides outpatient diagnostic imaging services, announced the launch of DeepHealth, a product line it designed to use AI to “drive efficiency and transform radiology’s role in healthcare.”
DeepHealth offers TechLive, an AI-powered remote scanning platform that enables “collaborative multi-modality imaging operations,” which is vendor agnostic and enables a radiologist to direct the work of technicians to remotely do imaging studies.
Sokka says the company’s use of AI helps to create “an integrated solution on a cloud-based structure. The worklist, reporting, viewer, advanced visualization and analytics … can be synced together to enable continual improvement and build trust in AI.”
Coalescing around efficiency
Radiologists have been reluctant to trust AI with the clinical aspects of their jobs, but because of workload volumes, declines in workforces and burnout concerns, are now open to using the technology to boost efficiency, says David Niewolny, director of business development for healthcare and medical at NVIDIA.
The company has had a dedicated healthcare practice for five years, but NVIDIA is engaging with companies and organizations to go deeper in several verticals, Niewolny says. For example, it’s offering MONAI, an open-source framework that enables deep learning in healthcare imaging.
As an example of the potential, Aidoc noted that it is entering a partnership with NVIDIA MONAI in an agreement announced at RSNA25. Through the effort, Aidoc aims to “expand its clinical AI scope to help health systems deploy imaging AI models … into routine clinical use at scale.”
The Aidoc initiative is emerging as "demand for imaging AI continues to climb,” the company notes. “Health systems are adopting more vendor solutions, and leading hospitals and academic centers are increasingly developing their own models tailored to local populations and priorities. Yet the biggest barrier remains the last mile: not training models but deploying them effectively in clinical workflows and scaling them across sites and modalities.” Aidoc’s approach, developed through the partnership with NVIDIA and Quibim, is intended to make “imaging AI deployable and scalable in clinical practice.”
Achieving clinical results
Despite latent reticence using AI for clinical purposes, some evidence is emerging that it can make a difference in specific use cases.
For example, a study recently published in Nature Health was based on data from DeepHealth’s Breast Suite applications in what RadNet calls the largest real-world analysis of AI-powered breast cancer screening in the U.S. The data, from mammograms from more than 579,000 women across 100-plus community-based imaging sites, demonstrated a 21 percent increase in the rate of breast cancer detection.
“Furthermore, the technology has been proven to raise the performance of generalist radiologists to the level of specialists, expanding access to high-quality breast care in regions where experienced readers may be limited,” the company contends.
In another application of imaging supported by AI, RapidAI marries the technologies to better diagnose aneurysms that result in strokes, detecting them and then visualizing them, through the use of its Lumina 3D technology, in three-dimensional renderings. Already, technological advances are enabling faster diagnoses that enable time-critical treatment to begin in six hours vs. 24 hours only a few years ago, says Karim Karti, CEO of the San Mateo, Calif.-based company.
Treatment with clot-busting drugs is crucial because ruptured aneurysms “can be very lethal; you want to be proactive, and you want to be able to monitor it and use advanced technology to know when to intervene,” Karti says.
The company hopes to further develop the technology to provide insights for different kinds of strokes and other conditions, such as pulmonary embolisms, where speed in response is critical.
“Integration in workflow is critical,” Karti says. “Radiology needs to interact seamlessly with a PACS and not slow down the reads. “The second integration point is with clinicians. One of the things we’ve continued to expand is the quality of reporting, reducing cognitive burden, and enabling clinicians to make decisions faster.
“Radiologists don’t need to spend a lot of energy doing mundane things,” he concludes. “We’re now specializing in detection, but ultimately we’re getting into the predictive realm. The earlier you intervene, the better the outcomes and the lower the cost.”