Computerized classification method for significant coronary artery stenosis on whole-heart coronary MRA using 3D convolutional neural networks with attention mechanisms.

This study aims to develop a computerized classification method for significant coronary artery stenosis on whole-heart coronary magnetic resonance angiography (WHCMRA) images using a 3D convolutional neural network (3D-CNN) with attention mechanisms. The dataset included 951 segments from WHCMRA images of 75 patients who underwent both WHCMRA and invasive coronary angiography (ICA). Forty-two segments with significant stenosis (luminal diameter reduction $\ge$ 75%) on ICA were annotated on WHCMRA images by an experienced radiologist, whereas 909 segments without it were annotated at representative sites. Volumes of interest (VOIs) of 21 × 21 × 21 voxels centered on annotated points were extracted. The network comprises two feature extractors, two attention mechanisms (for the coronary artery and annotated points), and a classifier. The feature extractors first extracted the feature maps from the VOI. The two attention mechanisms weighted the feature maps of the coronary artery and those the neighborhood of the annotated point, respectively. The classifier finally classified the VOIs into those with and without significant coronary artery stenosis. Using fivefold cross-validation, the classification accuracy, sensitivity, specificity, and AUROC (area under the receiver operating characteristic curve) were 0.875, 0.905, 0.873, and 0.944, respectively. The proposed method showed high classification performance for significant coronary artery stenosis and appears to have a substantial impact on the interpretation of WHCMRA images.