Abstract
BACKGROUND: Automated analysis of cardiac computed tomography (CCT) studies may help in personalized management and outcome prediction of patients undergoing transcatheter aortic valve replacement (TAVR). The current methods are often preceded by a manual selection of the region of interest. To address this limitation, this study aims to develop an object-oriented aortic root detection pipeline. METHODS: All consecutive patients who underwent CCT for TAVR procedure, from January to July 2023 at our center, were retrospectively collected. Patients with previous prosthesis or permanent pacemaker were excluded. Baseline bounding box annotations were performed by a single expert, and tilt angle measurements were performed by 2 for interobserver comparison. A pretrained convolutional neural network was used for aortic root detection, and its performance was evaluated by recall, precision, F1, average precision at an intersection over union overlap of 50% and mean average precision (mAP) 50%-95% on 100 unseen test set. For tilt alignment, intensity thresholding, connected component, and principal component analyses were proposed. Results were evaluated by Bland-Altman comparison. RESULTS: Of the 228 TAVR patients with preprocedural CCT, 179 were eligible, and their axial contrast-enhanced CCTs could be retrieved successfully; 100 CCTs were assigned to the test set, and the remaining to the training and validation using a 4:1 split. The model detected the aortic root with recall, precision, and F1 score of 99.0%, for all 3; mAP50 of 99.5%; and mAP50-95 of 60.4%. The tilt prediction algorithm had a mean error of 7.9° (Q1-Q3, -5.3° to 21.1°) compared with 3.3° (Q1-Q3, -6.7° to 13.4°) interobserver difference. CONCLUSIONS: This study demonstrates the robust performance of a fully automated pipeline for aortic root detection and analysis of key features in pre-TAVR CCTs. Further prospective studies are required for clinical developments.