Congenital heart disease (CHD) can alter the connections, geometry, and relative positions of the cardiac chambers and great vessels in ways that differ substantially from one patient to another. In complex defects such as double outlet right ventricle, surgeons cannot rely on a standard anatomical template. Pre-operative planning depends on understanding the individual three-dimensional relationships that determine where blood flows, how vessels connect, and which surgical pathways are technically possible.
CT and cardiovascular MR provide the necessary volumetric information, but the anatomy is still commonly reviewed as stacks of two-dimensional slices. Patient-specific 3D models and virtual-reality visualisation can make hidden spatial relationships easier to inspect from arbitrary viewpoints. The practical bottleneck is segmentation: converting a scan into a reliable model of the chambers and great vessels can require several hours of expert manual work. This limits routine use of detailed 3D planning, particularly for anatomically complex cases.
Surgeons need a patient-specific model that preserves the unusual connections and spatial relationships that matter for an individual operation. A plausible-looking heart is not enough if a chamber boundary, septal defect, or great-vessel connection is wrong. The segmentation therefore has to be accurate enough to support detailed review, available quickly enough to fit pre-operative workflows, and presented in a form that clinicians can inspect before use. Because scans vary across hospitals, scanners, protocols, populations, and imaging modalities, the underlying segmentation method also needs to remain reliable when the data differ from the cohort on which it was developed.
Our current work tests how dependable automated segmentation of the cardiac chambers and great vessels is when the data differ from the cohort a method was developed on. We compare representative approaches across multi-vendor CT and CMR cohorts, holding one out entirely as an out-of-distribution target and repeating every measurement across independent seeds.
Under combined CT+CMR training, nnU-Net achieved the highest in-distribution Dice of 0.77 but fell to 0.51 on ImageCHD. SwinUNETR reached 0.67 OOD Dice, while MAE-pretrained SwinUNETR reached 0.68. With CT-only training, scratch SwinUNETR retained 0.64 Dice. When 11 labelled ImageCHD cases were added for adaptation, all SwinUNETR variants exceeded 0.76 Dice. These experiments address whether segmentation methods intended for patient-specific 3D modelling remain dependable beyond their development cohort; clinical validation of downstream VR models is a separate step.