A Strong Linear Baseline for Whole-Heart Cardiac Shape Completion on CT, with an Open Eleven-Structure Statistical Shape Model addresses the challenge of unifying cardiac datasets that contain different anatomical annotations. Because public cardiac cohorts often label different subsets of the heart, researchers cannot easily pool data without shared correspondence. This paper introduces an eleven-structure statistical shape model (SSM) and a completion operator that allows researchers to reconstruct missing cardiac structures from whatever subset of the heart is currently available in their data.
The Problem of Fragmented Cardiac Data
Public cardiac datasets are inconsistent in their anatomical coverage. Some provide only ventricular chambers, while others include various vessels or appendages. Existing models often rely on fixed input and output structures, meaning they cannot adapt to the specific, partial labels provided by different cohorts. Furthermore, previous benchmarks for cardiac shape completion have relied on least-squares projections, which the authors argue are less accurate than the conditional estimators available from a fitted statistical model.
A Unified Eleven-Structure Model
The authors released an SSM built from 383 computed tomography (CT) cases, establishing a shared correspondence across 11,571 vertices. This model covers eleven distinct structures: the left ventricle, myocardium, right ventricle, left atrium, right atrium, aorta, pulmonary artery, left atrial appendage, pulmonary veins, superior vena cava, and inferior vena cava. By using a single template warped into each patient’s space, the researchers created a resource that supports cohort-unification research, allowing for the reconstruction of missing structures based on the ones present in a given scan.
Performance of Completion Estimators
The study compared several methods for reconstructing missing structures, including a closed-form conditional-Gaussian estimator, a mask-conditioned graph variational autoencoder, and nearest-neighbor retrieval. On a held-out set of 76 cases, the closed-form conditional-Gaussian estimator achieved a mean per-vertex error of 3.717 mm. This outperformed the graph variational autoencoder (5.248 mm) and nearest-neighbor retrieval (8.931 mm). The authors note that the closed-form estimator consistently performed better across different test scenarios, including external datasets where expert manual labels were available for comparison.
Limitations and Intended Use
The authors emphasize that this work is intended to support cohort-unification research on aligned CT data and is not for clinical use. Several limitations are noted:
Cardiac Phase: The dataset includes non-gated acquisitions, meaning the cardiac phase is sampled arbitrarily and treated as unstructured variation.
Data Source: The internal model was built using automatically generated labels from TotalSegmentator v2, which the authors categorize as "silver-standard." * Validation: While the model was tested against external expert-labeled cohorts, the reference-closeness rule meant that only a subset of the eleven structures could be scored against expert manual labels.
Scope: The model does not represent total myocardial mass, as it only includes the left-ventricular wall and blood-pool surfaces for the other ten structures.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!