Fully automated CT analysis with a deep-learning based algorithm for pre-procedural planning in transcatheter aortic valve implantation

Abstract

Abstract Background Pre-procedural cardiac computed tomography analysis has emerged to an essential tool in transcatheter aortic valve implantation (TAVI). It facilitates the valve choice and allows for optimal vascular access planning. For this purpose, software programs have been specifically developed to assist a fast and precise analysis of the generated data set in high quality. However, these software programs only assist the evaluator and are not able to perform a complete analysis on their own. Purpose We report the performance of a fully-automated, deep-learning based algorithm for pre-procedural CT analysis as compared to the current clinical standard. Methods Ninety-eight patients with symptomatic severe aortic stenosis undergoing TAVI were retrospectively enrolled. The pre-procedural data set was analyzed in two ways: - conventionally with a commercially available TAVI CT-analysis software by experiences users . - by a fully-automated CT-analysis platform with a deep-learning based algorithm. The current beta-version of the platform allows the determination of the annulus diameter, -area, -perimeter and the distance measurements of the coronary arteries to the aortic annulus. Hence, analyzable parameters were limited to the aforementioned, to ensure comparability. Statistical analysis: Mean, standard deviation, mean of difference and the interclass-correlation coefficient (ICC) were calculated and Bland-Altman plots created. Results The mean annulus diameter was 24.5±2.5mm (conventional) vs. 24.1±2.4 mm (deep learning). The mean of difference was 0.42±0.82mm (Bland-Altmann diagram. Figure 1). The ICC was 0.964 (0.925; 0.980), showing an excellent correlation between the two values. The mean annulus perimeter was 77.7±7.4mm (conventional) vs. 75.6±9.3mm (deep learning), showing a very good correlation of the means [mean of difference: 2.11±5.75mm; ICC: 0.852 (0.759; 0.906)]. The distance from the annulus to the left coronary artery was 14.0±3.2mm (conventional) vs. 12.6±2.8mm (deep learning). We depicted a good correlation between the two values [mean of difference: 1.45±2.52mm; ICC 0.736 (IQR 0.483; 0.850)]. The distance from the annulus to the right coronary artery was 17.1±2.7mm (conventional) vs. 16.5±3.2mm (deep learning), respectively. A very good correlation between the two values could be shown [mean of difference 0.57±2.2mm; ICC 0.822 (0.730; 0.882)], as well. Additionally, both methods visualize the measurements and allow manual adjustments (Figure 2). Conclusion In this retrospective analysis, deep-learning based analysis of peri-procedural derived CT data sets showed an very good correlation with conventional assessment. If results can be sustained in a bigger patient cohort and extended to more complex measurements, this artificial intelligence based fully-automated CT-analysis could emerge to a valuable alternative to conventional CT assessment in the pre-procedural planning for TAVI.Figure 1 - Bland-Altman plotFigure 2- Measurements by both methods