Deep-learning algorithm automates FFR-CT estimation

Sunday, November 25 | 11:35 a.m.-11:45 a.m. | SSA03-06 | Room S404AB

Researchers from Japan have developed a deep-learning model capable of automatically estimating minimum fractional flow reserve CT (FFR-CT) with an accuracy approaching that of invasive FFR.

The group, led by Dr. Kanako Kumamaru, PhD, from Juntendo University in Tokyo, devised the model by combining a series of deep-learning algorithms, including a 3D convolutional ladder network, a conditional generative adversarial network, and two independent neural networks.

Current use of FFR-CT requires clinicians to define coronary artery tree anatomy manually before analysis, which is time-consuming and also subject to variability and imaging artifacts that can compromise the accuracy of the results, Kumamaru told AuntMinnie.com.

"Unlike other FFR-CT technologies or previously reported machine learning-based FFR-CT, our model enables us to be completely free from time-consuming vessel segmentation," he said.

Rather than identifying every individual stenosis, the deep-learning approach determines whether a patient has functional stenosis based on the FFR measurements and, in turn, whether the patient will require invasive coronary angiography.

The researchers tested their approach on 1,052 patients who underwent coronary CT angiography. They found that the automated deep-learning model was highly accurate in calculating FFR.

"We anticipate that our model will contribute to improving clinical workflow when deciding on candidates suitable for revascularization, after further improvement of the accuracy of the model with a larger training dataset," Kumamaru said.