By Dan Harvey
Vol. 18 No. 9 P. 10
Artificial intelligence helps radiologists identify cardiac risk factors.
Artificial intelligence (AI), the branch of computer science that enables technology to perform human cognitive functions such as learning and problem solving, has established a strong foothold in many sectors of industry. In the past decade, AI and, more specifically, machine learning—a type of AI—made inroads into the health care sector. New tools have been developed and are now being used in the clinical setting to assess the extent of coronary artery disease (CAD). Combining AI algorithms with CT and coronary CT angiography (CCTA), researchers and product manufacturers have designed software tools to help physicians better diagnose and treat patients.
So far, two companies—HeartFlow of Redwood City, California, and Israel-based Zebra Medical Vision—have commercialized such technology. Their offerings provide information not only about coronary artery anatomy but also about physiology, such as arterial blood flow. As a result, physicians can more accurately diagnose patients with suspected CAD and deliver a definitive personalized treatment plan for each patient. With these tools, AI essentially serves clinicians as an automated assistant.
The tools' developers expect that they will have a significant impact on heart health. In the United States, more than 16 million individuals suffer from CAD, which causes coronary arteries to narrow, reducing blood flow and potentially leading to lethal heart attacks. Clinicians need to know as quickly and accurately as possible whether a patient has a significant blockage.
Estimating Fractional Flow Reserve
HeartFlow's noninvasive fractional flow reserve (FFR) technology, known as FFRct, was designed to provide clinicians with more knowledge about the impact of CAD on blood flow to the heart. Charles Taylor, PhD, cofounded HeartFlow with Christopher Zarins, MD, 10 years ago. The company's founding was a natural progression from Taylor's research at Stanford University, where he worked on developing image processing methods to extract anatomy and create 3D computational models to simulate blood flow.
"We focused the company on creating patient-specific models, particularly for coronary anatomy to simulate blood flow and provide a noninvasive estimation of fractional flow reserve, which was becoming the gold standard for invasive assessments of obstructive coronary disease," Taylor recalls.
FFR values demonstrate blood pressure differences around lesions to determine the functional impact of the lesion and the likelihood of reduced blood flow to the heart. Whether obtained invasively or simulated noninvasively, these values help physicians determine the right course of action. Taylor says that, at the time of the company's founding, no machine learning methods were available to apply to image segmentation problems, extract anatomy from image data, or build a computational model for a patient's specific blood flow analysis.
"This was important because small errors in the anatomy can translate into larger errors in the physiology," Taylor says. "So, five years ago, before most people had heard of it, we started applying machine learning methods to extracting anatomy."
In recent years, HeartFlow has focused on applying deep learning, a type of machine learning. The company's current product is trained on segmentation analysis from thousands of patients. The numbers will increase over time and lead to continued improvement in the quality of image segmentation.
To date, more than 13,000 patients have received a HeartFlow FFRct. During this process, data from a patient's noninvasive CCTA are uploaded from a facility's system into the cloud. A personalized 3D model of the patient's arteries is constructed with deep learning algorithms. Millions of complex equations simulate blood flow in the model and assess the impact of blockages. The results are provided to the patient's clinician via a secure web interface.
In 2015, FFRct was the subject of an important study—the multicenter, controlled, prospective PLATFORM (Prospective Longitudinal Trial of FFRct: Outcome and Resource Impacts) trial. This study, led by Duke University School of Medicine researchers, included more than 580 patients and revealed that FFRct decreased unnecessary invasive diagnostic testing for CAD by 83%. The study compared standard diagnostic strategies with the FFRct noninvasive technology. According to the study authors, specific findings include the following:
• Among patients with a planned invasive coronary angiography (ICA), 73% had an ICA that showed no significant blockage or obstruction. Only 12% of patients evaluated using an FFRct-guided strategy went on to have an ICA that showed no significant obstruction (83% reduction). In 61% of patients, the use of an FFRct-guided strategy resulted in the cancelation of a planned invasive test.
• Despite the difference in the number of patients who required ICA, the rate of revascularization procedures, such as coronary stenting or bypass surgery, was similar between patients assigned to standard testing strategies (31.6%) and those assigned to FFRct-guided strategies (28.5%).
• No clinical adverse events were reported at the 90-day follow-up in any of the 117 patients whose physicians chose to cancel an ICA based on FFRct-guided strategy.
In their published study, the researchers concluded that use of FFRct enabled physicians to efficiently triage patients to the most appropriate care and to dramatically reduce the use of invasive testing. While AI and machine learning demonstrate value as well as great potential, Taylor is quick to point out that his company and its technology are not going to replace any health care professional.
"We aren't making tools designed to replace the radiologist," he says. "There has been so much talk about AI, but that is one thing that many people misunderstand."
The technology should only help the radiologist perform better. "Coronary CT data can be very challenging to read in some patients, especially those with increasing disease burden," Taylor says. "With our technology and the additional data we make available, we believe we can help make a radiologist a super-radiologist. We see the machine learning capability and the radiologist working together to be able to extract the best available information for the patient and enable radiologists to read at much higher levels. Some people think that we shouldn't train more radiologists because they will be replaced by machines. That's hubris and is dead wrong. With our tools, the radiologist is armed with information to provide better patient service. We're empowering, not replacing, the radiologist."
In July, the American Medical Association issued a set of new Category III CPT codes for HeartFlow FFRct. In June, Blue Cross Blue Shield Association's Evidence Street issued a positive health care evidence review of FFRct. The technology received European CE mark approval in 2011 and FDA clearance in November 2014.
Lynne Hurwitz Koweek, MD, an associate professor of radiology at Duke University School of Medicine, where FFRct has been used following the PLATFORM study, says the technology provides a highly accurate tool. "Our physicians have been able to accurately evaluate CAD and avoid unnecessary invasive procedures," Koweek says. She adds that the FFRct-enabled reduction in unnecessary, inconclusive, and inaccurate tests can reduce costs and make a staff much more productive—as long as the adopting sites have available 64-slice (or greater) CCTA.
The University Hospitals of Cleveland is another health system that has adopted the technology. "We started using it in 2015," says Robert C. Gilkeson MD, vice chairman of research and a professor of radiology at University Hospitals. "We were the first health system to adopt the use of HeartFlow in patients with chest pain. The successful implementation was enabled through a close partnership between radiology and cardiology. As an interventional cardiologist, the director of our Heart and Vascular Institute, Dan Simon, was acutely aware of the high rate of negative invasive coronary catheterizations. We were committed to a model that would decrease that number and save costs to the health system."
Beyond the clinical benefits, Gilkeson also sees economic advantages. "When we look at our projected reduction in catheterizations, volumes, and the cost of the volumes, we stand to realize significant cost savings."
Like Taylor, Gilkeson stresses that FFRct is not a replacement technology. He adds that in a truly successful implementation, the human and the computer-aided device need to work together. The radiologists haven't been replaced; their diagnostic abilities are enhanced. The radiologist will interpret the findings.
"Though AI is involved, it is important to interpret the CCTA and the data from HeartFlow together. Because of the improvement in sensitivity that is evident, an integrative approach—with human and computer—is required."
Both Koweek and Gilkeson state that the FFRct technology is not difficult to master. "The learning curve is not that steep," Koweek says. "The clinicians who operate it already have the requisite skills, learning, and experience to be able to readily and fairly easily use the technology."
Gilkeson says it only takes several hours to learn how to use FFRct. "But with the small learning curve, the user needs the understanding of the CT world. Close adherence to optimal cardiac CT techniques is crucial to successful FFRct. So, there is that kind of investment at the start." He adds that to best understand the technology, the reader needs to communicate with angiographic or cardiology colleagues.
The time needed by HeartFlow to analyze the data from the CT and provide results to the clinician is minimal. "HeartFlow has speeded up the implementation such that most analyses can now be accomplished within 24 hours," Gilkeson says.
Knowing the Score
As with HeartFlow's product, Zebra Medical Vision's Deep Learning Imaging Analytics Engine provides early detection of people at high risk for severe cardiovascular events. In this case, AI is used to automatically calculate coronary calcium scores.
The quantification of coronary calcification is a strong predictor for cardiovascular events such as heart attack or strokes. Conventional coronary calcium scoring requires dedicated cardiac, ECG-gated CT performed with and without contrast. With the Zebra algorithm, the score can be obtained from low-dose noncontrast chest CT data.
"There has always been the computer-intensive and algorithm-intensive element in medical imaging, so when AI and machine learning came about, one of the most relevant places in health care for implementation turned out to be imaging," says Elad Benjamin, Zebra's cofounder and CEO.
Benjamin describes how machine learning comes into play. "It is a different approach to teaching software how to identify features in images," he explains.
"There can be hundreds of thousands of features in an image. And software doesn't learn a visual fingerprint on its own. For instance, if you feed it an image of broken rib, and it says, 'This is a broken jaw,' you've got to correct it and tell it, 'This is a broken rib.' It will calibrate itself so it won't make the same mistake again."
But Benjamin echoes Taylor's thought about the relationship between user and technology. "AI will not do the diagnosis, and it won't replace physicians. It will only enhance their skills."
Zebra's analytic engine can be deployed in both cloud and on-premise configurations, and its insights help radiologists provide more comprehensive and accurate reports. The company's coronary calcium scoring algorithm can run a broader spectrum of cases involving people who aren't necessarily suspected of having cardiovascular disease.
"This opens up the opportunity to provide coronary calcium scores to a much broader population than a standard chest CT," Benjamin says. "When you can do that, potentially you can catch people who are at high risk that you didn't know about before because they didn't have external symptoms."
Benjamin says Zebra has received positive user feedback about its coronary calcium scoring tool. Also, as with the HeartFlow product, Zebra's analytic engine can be integrated into PACS, RIS, reporting, and EMR systems.
Zebra has been granted CE approval and subsequent release of its Deep Learning Imaging Analytics Engine in Europe, as well as regulatory clearance and product release in Australia and New Zealand. Earlier this year, US-based Intermountain Healthcare partnered with Zebra to integrate machine learning in medical imaging analysis, with the goal of providing better patient care. Intermountain is interested because Zebra's analytics engine is not just applicable to cardiac care; it receives imaging data and analyzes findings indicative of metabolic and bone health as well. Intermountain will use the results to identify patients at risk. The Zebra analytic engine currently analyzes CT data for signs of fatty liver, excess coronary calcium, emphysema, low bone density, and vertebral compression fractures.
In addition, Zebra has entered into another partnership—a reseller arrangement—with Carestream Health. This partnership is intended to provide imaging analytics that deliver automated population health insights and decision support tools. The software will analyze past and current imaging studies within Carestream's radiology diagnostic client to identify chronic conditions that may require preventive treatment, according to the company.
Also, Siemens Healthineers and GE Healthcare have entered into separate global collaboration agreements with HeartFlow; this will increase HeartFlow FFRct's clinical availability and adoption.
Looking ahead, Taylor sees AI in imaging as something that will become more pervasive in health care. "Five to 10 years ago, people were saying it would be a long time before tools like this would be available in radiology," he says. "That turned out to be not true. It's amazing how well these tools work and how fast they're improving."
— Dan Harvey is a freelance writer based in Wilmington, Delaware.