Recursive multiresolution convolutional neural networks for 3D aortic valve annulus planimetry.

Abstract

Transcatheter aortic valve replacement (TAVR) is the standard of care in a large population of patients with severe symptomatic aortic valve stenosis. The sizing of TAVR devices is done from ECG-gated CT angiographic image volumes. The most crucial step of the analysis is the determination of the aortic valve annular plane. In this paper, we present a fully tridimensional recursive multiresolution convolutional neural network (CNN) to infer the location and orientation of the aortic valve annular plane. We manually labeled 1007 ECG-gated CT volumes from 94 patients with severe degenerative aortic valve stenosis. The algorithm was implemented and trained using the TensorFlow framework (Google LLC, USA). We performed K-fold cross-validation with K = 9 groups such that CT volumes from a given patient are assigned to only one group. We achieved an average out-of-plane localization error of (0.7 ± 0.6) mm for the training dataset and of (0.9 ± 0.8) mm for the evaluation dataset, which is on par with other published methods and clinically insignificant. The angular orientation error was (3.9 ± 2.3)° for the training dataset and (6.4 ± 4.0)° for the evaluation dataset. For the evaluation dataset, 84.6% of evaluation image volumes had a better than 10° angular error, which is similar to expert-level accuracy. When measured in the inferred annular plane, the relative measurement error was (4.73±5.32)% for the annular area and (2.46±2.94)% for the annular perimeter. The proposed algorithm is the first application of CNN to aortic valve planimetry and achieves an accuracy on par with proposed automated methods for localization and approaches an expert-level accuracy for orientation. The method relies on no heuristic specific to the aortic valve and may be generalizable to other anatomical features.