Left atrium segmentation in 4D CT imaging using cycle-consistent registration

Ilse Meijer, Louis van Harten, Laura Carneiro Penetra Pires, Irene van der Snoek Fidalgo, Marije M. Vis, Nils Planken, Ivana Isgum, Joris R. de Groot, Nerea Arrarte Terreros, Henk Marquering

Keywords: 4DCT, segmentation, deformable registration, implicit neural representations, left atrium, atrial fibrillation

Abstract: Motion analysis of the left atrium (LA) using 4D cardiac computed tomography (CT) is valuable in atrial fibrillation management and thrombus risk assessment. Challenges in accurate time-resolved segmentation arise from patient-specific pulmonary vein and left atrial appendage anatomy, complex LA motion, and contrast variability. This study introduces and evaluates the use of cycle-consistent deformable image registration, an implementation of implicit neural representations (INRs), for LA segmentation across the cardiac cycle. Ten 4D cardiac CT images were analysed. Manual segmentations were used as the reference standard. Cycle-consistent segmentation demonstrated superior performance in LA volume and Hausdorff distance compared to an nnU-Net model with consistent transfer of manual post-processing across timeframes. These findings highlight the potential of deformable image registration with INRs as a robust approach for 4D cardiac image segmentation.

Submission Number: 65