Diagnostic accuracy of a deep learning approach to calculate FFR from coronary CT angiography

The computational fluid dynamics (CFD) approach has been frequently applied to compute the fractional flow reserve (FFR) using computed tomography angiography (CTA). This technique is efficient. We developed the DEEPVESSEL-FFR platform using the emerging deep learning technique to calculate the FFR value out of CTA images in five minutes. This study is to evaluate the DEEPVESSEL-FFR platform using the emerging deep learning technique to calculate the FFR value from CTA images as an efficient method.

A single-center, prospective study was conducted and 63 patients were enrolled for the evaluation of the diagnostic performance of DEEPVESSEL-FFR. Automatic quantification method for the three-dimensional coronary arterial geometry and the deep learning based prediction of FFR were developed to assess the ischemic risk of the stenotic coronary arteries. Diagnostic performance of the DEEPVESSEL-FFR was assessed by using wire-based FFR as reference standard. The primary evaluation factor was defined by using the area under receiver-operation characteristics curve (AUC) analysis.

For per-patient level, taking the cut-off value ≤ 0.8 referring to the FFR measurement, DEEPVESSEL-FFR presented higher diagnostic performance in determining ischemia-related lesions with area under the curve of 0.928 compare to CTA stenotic severity 0.664. DEEPVESSEL-FFR correlated with FFR (R = 0.686, P < 0.001), with a mean difference of -0.006 ± 0.0091 (P = 0.619). The secondary evaluation factors, indicating per vessel accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were 87.3%, 97.14%, 75%, 82.93%, and 95.45%, respectively.

DEEPVESSEL-FFR is a novel method that allows efficient assessment of the functional significance of coronary stenosis.