UCSF

Patient-specific three-dimensional (3D) models can support surgical planning for complex congenital heart disease (CHD), but manual segmentation of cardiac computed tomography (CT) is time intensive. This thesis evaluated nnU-Net v2 for multi-structure CHD CT segmentation in three phases. Phase 1 screened nnU-Net configurations on the public CHD68 dataset and selected a 1,250-epoch model with learning rate 0.01. The resulting CHD68 model (M0) achieved mean case-level Dice 0.846 on 14 validation cases. Phase 2 applied M0 without additional training to 41 UCSF CT cases to establish a local baseline. Across six shared blood-pool structures, mean Dice decreased from 0.845 on CHD68 to 0.389 on UCSF, with substantial case heterogeneity and several complete structure omissions. Phase 3 trained a UCSF blood-pool model (M1), a UCSF + CHD68 blood-pool model (M2), and a ten-label UCSF hollow model (M3). On the same nine held-out UCSF blood-pool cases, mean Dice was 0.488 for M0, 0.754 for M1, and 0.755 for M2; M2 also largely retained CHD68 performance. M3 achieved mean Dice 0.469 across ten structures and 14 validation cases. A secondary ground-truth-bounded analysis increased mean Dice to 0.484 (+0.015), with the largest changes in the SVC, aorta, IVC, and pulmonary veins. Qualitative review showed that several nonfailure hollow predictions preserved major cardiac geometry despite modest overlap. Hollow and blood-pool Dice are not directly comparable because the targets differ in thickness, topology, branch extent, and truncation. The findings support continued development of hollow segmentation as an expert-editable starting point for the CA3D+ 3D-modeling workflow, with annotation refinement, cohort expansion, and direct evaluation of correction burden and printability as next steps.