Comparing Convolutional Neural Networks for Aortic Root Segmentation: A Geometry–oriented Feasibility Study

Background Accurate anatomical characterization of the aortic root is a key prerequisite for transcatheter aortic valve implantation (TAVI) planning. Although deep learning–based segmentation techniques have demonstrated high performance in delineating cardiac structures, the translation of automated segmentation into reliable extraction of clinically relevant anatomical landmarks remains a critical and insufficiently explored challenge. Aim To describe and evaluate a two–stage methodological framework for aortic root analysis, combining automated CNN–based segmentation with exploratory strategies for landmark identification relevant to TAVI planning. Methods In the first stage, manual aortic root segmentations were used to train a U–Net–based model for automatic segmentation of 100 contrast–enhanced cardiac computed tomography scans. Segmentation quality was assessed against expert–defined references to verify the stability and anatomical consistency of the generated masks. In the second stage, manually annotated anatomical landmarks, including cusp nadirs, coronary ostia height, and sinotubular junction, were used as references to explore automated landmark identification approaches operating on the segmented anatomy. This phase focused on assessing geometric consistency, spatial stability, and sources of uncertainty affecting landmark localization rather than definitive performance metrics. Results The CNN–based segmentation model converged stably over the training period and showed a visually adequate definition of the aortic root as an anatomically consistent and accurately observed source between the analyzed scans as a valuable reference for downstream analysis. The automated identification of the target structural landmarks showed varying behaviors based on the type and location of the landmarks. Landmark localization was strongly influenced by local geometrical variability, curvature, and partial volume effects, indicating that there may be a methodological rift between strong segmentation and accurate landmark measurement. Conclusions This study confirms the reliability of model’s aortic root segmentation but limited threshold for the complete automatic extraction of clinically pertinent landmarks in TAVI planning. Clearly delineating segmentation and landmark identification as separate methodological stages may be important for building clinically sensitive AI–based imaging pipelines.