Automatic aortic root morphology assessment algorithm for Transcatheter Aortic Valve Implantation planning

Accurate planning for Transcatheter Aortic Valve Implantation (TAVI) is crucial for reducing complications, and this requires an anatomical assessment of the aortic root (AR). However, there is currently no cost-effective, automated algorithm for this process. We trained a self-supervised 3D V-Net convolutional neural network on 132 Computed Tomography (CT) scans and 100 Computed Tomography Coronary Angiography (CTCA) scans to segment both the AR and the left ventricle (LV). Based on the segmentation results, we propose an iterative generation algorithm based on tri-positioning to locate the annulus plane and integrate it into a complete workflow for AR morphology assessment. The results were validated against manual measurements from 30 TAVI candidates. The trained CNN effectively segmented the AR and the LV, achieving an average Dice score of 0.953 for the AR and 0.928 for the LV. The automatic measurements closely matched the manual annotations, with the maximum diameter of the annulus of the aortic valve differing by -2.07 [-4.11, -0.04] mm (bias and 95% limits of agreement for the manual subtraction algorithm), the mean diameter of the annulus of the aortic valve differing by -0.58 [-1.48, 0.32] mm. The differences in the maximum diameters of the sinus of the sinotubular junction (STJ) and the minimum diameters of the aortic sinus calculated by the automatic method were -1.33 [-3.01, 0.36] mm and -3.06 [-6.71, 0.58] mm, respectively. For the height of the aortic sinus and coronary artery, the bias and the limits of agreement were -0.08 [-3.74, 3.57] mm and -1.26 [-7.95, 5.44] mm. The proposed algorithm is a fully automated solution for the quantitative assessment of the morphology of the AR for pre-TAVI planning. The method was validated against clinical expert manual annotations and demonstrated high speed, efficiency, and consistency in AR anatomical assessment, offering potential for time and cost savings while uncovering more possible parameters.