Automated computation and comparison of three-dimensional left atrial shape descriptors from deep learning-assisted CT segmentation in atrial fibrillation: a retrospective methodological imaging study

Background Three-dimensional left atrial (LA) shape descriptors are increasingly used to characterize atrial remodeling in atrial fibrillation (AF). However, several metrics have been reported under overlapping terminology, especially “sphericity”, and their relationships with each other, LA volume, and segmentation-boundary uncertainty remain insufficiently defined. Methods This retrospective single-centre methodological imaging study included 186 patients with AF who underwent pre-ablation contrast-enhanced cardiac CT angiography. A MedNeXt-based workflow generated standardized LA-body meshes. Six global descriptors were computed: Wadell sphericity (ΨW), two coordinate-covariance eigenvalue-ratio indices (E and F), convex hull ratio (CHR), and two diameter-normalized volume sphericity indices based on the maximum diameter (DNVSImax) and the mean of three orthogonal diameters (DNVSImean). A stratified 40-case subset underwent expert-guided correction, and predefined voxel, mesh, and ± 1-voxel truncation-plane perturbations assessed workflow sensitivity. Duplicate execution quantified implementation consistency. Pearson correlations with Spearman sensitivity, descriptive exploratory LA-volume-adjusted subtype models, and a component-based allometric sensitivity analysis were performed. The AF-subtype × LA-volume interaction was not evaluated. Results In the 40-case expert-guided correction subset, Dice was 0.956 ± 0.015, HD95 was 1.62 ± 0.55 mm, and descriptor ICCs between automated and corrected outputs ranged from 0.871 to 0.947. These estimates quantify agreement with expert-guided correction, not independent manual validation. Duplicate-run ICCs were ≥ 0.996, indicating high implementation consistency. ΨW was essentially unrelated to DNVSImax and DNVSImean, whereas the two DNVSI variants were strongly correlated. E, F, CHR, DNVSImax, and DNVSImean were associated with LA volume, and the component analysis showed direction-dependent, non-isometric scaling. The multivariable CHR association was modest and nominal (standardized β = −0.23, 95% CI − 0.42 to − 0.04), and CHR showed the greatest tested boundary sensitivity (95th-percentile change, 4.8%). Conclusions The six CT-derived descriptors represent distinct geometric constructs. Expert-guided correction and boundary perturbation characterized their measurement sensitivity. Volume adjustment does not by itself establish volume-independent shape effects. The modest nominal CHR association, together with its sensitivity to boundary placement, remains hypothesis-generating. Clinical translation will require explicit formulas, standardized anatomical boundaries, appropriate size modeling, independent manual validation, scan–rescan evaluation, and external validation.