Smooth 3D Cardiac Geometry Reconstruction from 2D CT Slice Segmentations: Wall-Thickness Agreement with a Manual...

Myocardial wall thickness is a key measurement in cardiac imaging: it guides the diagnosis of hypertrophic cardiomyopathy, informs risk stratification for sudden cardiac death, and tracks ventricular remodeling in heart failure. Obtaining this measurement reliably requires accurate three-dimensional geometry of the heart. When two-dimensional segmentation masks are stacked to form a 3D shape, the coarse through-plane spacing typical of clinical cardiac CT produces jagged, stair-step surfaces that distort the very thickness measurements clinicians need. This work addresses the pipeline from 2D segmentation to smooth 3D cardiac geometry, culminating in wall thickness quantification benchmarked against single-expert manual reference measurements in a single-center cohort. To address both the segmentation and the reconstruction challenges identified above, we developed an end-to-end pipeline with two stages. In the first stage, a U-Net enhanced with attention gating and deep supervision segments cardiac structures from CT slices; to incorporate limited through-plane context, the model receives a small stack of adjacent slices as additional input channels (a 2.5D strategy). We trained on a manually annotated clinical cohort alongside a larger public dataset and conducted systematic ablations over view orientation, number of adjacent slices, dataset composition, and augmentation strategy, all with strict subject-wise splits to prevent data leakage. In the second stage, to eliminate the stair-step artifacts that compromise thickness quantification, we reconstruct smooth 3D surfaces using a spacing-aware approach based on signed-distance-field interpolation and Marching Cubes and compare it against a contour-based lofting alternative. Finally, wall thickness measurements are compared against a manual-segmentation-derived reference. We evaluated the pipeline across segmentation, reconstruction, and wall-thickness agreement. For segmentation, axial-plane models outperformed sagittal and coronal alternatives, reflecting the denser sampling and better anatomical alignment of the acquisition plane. Combined training on manual and public datasets achieved the highest ventricular Dice scores (LV, 89.1%; RV, 88.0%). Multi-slice 2.5D input improved over pure 2D across all cardiac structures, with peak performance at 3-5 adjacent slices (best LV, 90.9%; best RV, 89.3%; best EPI, 88.7%). Data augmentation further improved accuracy across all structures, with the largest gain observed for the RV (4.2 percentage points). These segmentation comparisons are reported descriptively; the held-out test set is too small to support significance testing. For reconstruction, the spacing-aware Marching Cubes method outperformed contour lofting in surface fidelity (LV ASSD, 0.795 vs. 1.223 mm) and measurable LV surface (75.6% to 88.3% of endocardial vertices yielding a valid thickness). Automated wall thickness agreed with the manual reference to within 0.4 mm mean absolute error in every region (LV free wall, 9.63 vs. 9.77 mm; RV free wall, 6.24 vs. 5.99 mm), with regional values in the range reported in the cardiac imaging literature. This work demonstrates that 2D segmentation, when paired with spacing-aware 3D reconstruction, can produce cardiac geometry whose derived wall-thickness measurements agree closely with single-expert manual reference measurements in a single-center cohort. The primary finding is endpoint-level agreement: automated thickness measurements show sub-millimeter mean error against the manual reference, a promising proof of concept that motivates larger multi-site, multi-vendor validation with independent expert measurements before clinical use.