FFR-CT algorithm guides management of revascularization

A deep-learning algorithm that analyzes fractional flow reserve CT (FFR-CT) scans was able to determine almost as well as invasive FFR which patients suspected of having coronary artery disease should be referred for revascularization. The finding could lead to better, less-invasive care, according to a presentation at RSNA 2017.

Researchers measured FFR by applying a deep-learning algorithm to coronary CT angiography (CCTA) scans of patients suspected of having coronary artery disease (CAD). They found that it was a good match compared to recommendations suggested by analyzing conventional FFR, which requires invasive coronary angiography. The group believes the algorithm could be more effective than other CT-based tools for assessing FFR.

"FFR-CT performed very well in determining the appropriate treatment strategy, with an accuracy of almost 100%," Dr. Philipp von Knebel Doeberitz from the Medical University of South Carolina told session attendees. "FFR-CT may be a noninvasive one-stop shop to aid the diagnostic protocol and guide therapeutic decision-making in patients with suspected or known coronary artery disease."

Guiding therapeutic strategies

Dr. Philipp von Knebel Doeberitz from the Medical University of South Carolina.

FFR-CT using specialized software has already proved to be a reliable method for assessing coronary stenosis, and large prospective trials have shown the technique to be capable of reducing rates of unnecessary invasive coronary angiography, according to von Knebel Doeberitz.

"As the diagnostic abilities of FFR-CT have already been shown, we wanted to investigate how an FFR-CT-guided approach for therapeutic decision-making would have performed," he told AuntMinnie.com.

Von Knebel Doeberitz led a team that went a step further by applying a machine-learning investigational algorithm (syngo.via Frontier, Siemens Healthineers) to FFR-CT to see if the algorithm could triage patients for revascularization, a decision that's typically made by a physician. Citing early work on the algorithm published in 2016 by Itu et al, von Knebel Doeberitz affirmed that the machine-learning algorithm offers benefits comparable to currently available techniques that use computational fluid dynamics to calculate FFR from CT scans.

The added advantage of the deep-learning algorithm is that it performs FFR computation in less than an hour, compared with the several hours it takes with other FFR-CT methods. Itu et al used a fully connected deep neural network consisting of four hidden layers to train the algorithm on a database of 12,000 synthetically generated coronary vessel trees reflecting various anatomies of patients with coronary artery disease.

In a retrospective study presented at RSNA 2017, von Knebel Doeberitz and colleagues evaluated the data of 74 patients who had at least one coronary stenosis with a diameter of at least 50% and who underwent both coronary CT angiography and invasive coronary angiography (ICA). They used a prototype machine-learning algorithm to collect FFR-CT values based on the patients' CCTA scans and then analyze the data to determine the appropriate therapeutic strategy -- medical therapy or revascularization.

By analyzing the FFR-CT scans, the algorithm identified obstructive coronary artery disease requiring revascularization (FFR < 0.8) in 35 out of 74 patients, and it directed the remaining 39 patients to medical therapy. The reference standard, invasive FFR, would have sent only one more patient (36) to revascularization. Among the 35 patients for whom the algorithm suggested revascularization, it selected 32 for percutaneous coronary intervention and three for coronary artery bypass grafting. This designation mirrored the therapeutic strategy indicated by invasive FFR.

The FFR-CT algorithm turned out to be nearly as accurate as invasive FFR when determining the appropriate revascularization procedure. It had an accuracy of 99%, a sensitivity of 97%, a specificity of 100%, a positive predictive value of 100%, and a negative predictive value of 97%. The average age of the patient population was 62.2 years; 62% were male.

Shifting diagnostic workflows

The researchers recognized the need for a larger prospective clinical trial to corroborate their findings. They also acknowledged a couple of study limitations: the small patient cohort, as well as potential selection bias because the researchers did not include patients with previous myocardial infarction or revascularization.

Drawing FFR measurements from CCTA scans in an all-inclusive test has the potential to shift diagnostic workflows when managing patients with coronary artery disease, von Knebel Doeberitz said. It can adequately inform therapeutic decision-making without requiring invasive procedures.

"Ultimately, we believe that [adding] FFR-CT in the clinical workup of patients with known or suspected coronary artery disease will be beneficial for patients," he said.