Development and validation of a deep learning-based fully automated algorithm for pre-TAVR CT assessment of the aortic valvular complex and detection of anatomical risk factors: a retrospective, multicentre study

Abstract

Background: Pre-procedural computed tomography (CT) imaging assessment of the aortic valvular complex (AVC) is essential for the success of transcatheter aortic valve replacement (TAVR). However, pre-TAVR assessment is a time-intensive process, and the visual assessment of anatomical structures at the AVC shows interobserver variability. This study aimed to develop and validate a deep learning-based algorithm for pre-TAVR CT assessment and anatomical risk factor detection.

Methods: This retrospective, multicentre study used AVC CT scans to develop a deep learning-based, fully automated algorithm, which was then internally and externally validated. After loading CT scans into the algorithm, it automatically assessed the essential anatomical structure data required for TAVR planning. CT scans of 1252 TAVR candidates continuously enrolled from Fuwai Hospital were used to establish training and internal validation datasets, while CT scans of 100 patients with aortic valve disease across 19 Chinese hospitals served as an external validation dataset. The validation focused on segmentation performance, localisation and measurement accuracy of key anatomical structures, detection ability of specific anatomical risk factors, and improvement in assessment efficiency.

Findings: Relative to senior observers, our algorithm achieved significant consistent performance with remarkable accuracy, efficiency and ease in segmentation, localisation, and the assessment of the aortic annulus perimeter-derived diameter, and other basic planes, coronary ostia height, calcification volume, and aortic angle. The intraclass correlation coefficient values for the algorithm in the internal and external validation datasets were up to 0.998 (95% confidence interval 0.998-0.998), respectively. Furthermore, the algorithm demonstrated high alignment in detecting specific anatomical risk factors, with accuracy, sensitivity, and specificity up to 0.989 (95% CI 0.973-0.996), 0.979 (95% CI 0.936-0.995), 0.986 (95% CI 0.945-0.998), respectively.

Interpretation: Our algorithm efficiently performs pre-TAVR assessments by using AVC CT imaging with accuracy comparable to senior observers, potentially improving TAVR planning in clinical practice.

Funding: National Key R&D Program of China (2020YFC2008100), CAMS Innovation Fund for Medical Sciences (2022-I2M-C&T-B-044).

Keywords: AI; CT; Deep learning; TAVI; TAVR.