Aims: Statistical shape modelling (SSM) enables quantitative analysis of cardiac anatomy beyond conventional scalar measurements. This study evaluated model compactness and reconstruction accuracy in two SSM frameworks applied to biventricular meshes and assessed the sensitivity of both to the choice of reference mesh, to inform framework selection for clinical shape analysis.
Methods: 298 end-diastolic biventricular surface meshes from 119 patients in the TRED-HF and REMIT-DCM trials were processed through Scalismo, which implements Gaussian process morphable models (GPMMs), and Deformetrica, which implements large deformation diffeomorphic metric mapping (LDDMM). Hyperparameters were optimised via grid search in both frameworks. Two reference meshes were tested per framework, a population medoid and a mean-volume mesh. Principal component analysis (PCA) was applied to the shape representations. Compactness was assessed as cumulative variance explained by components each capturing at least 1% of variance. Reconstruction accuracy was measured as the mean symmetric surface distance between each reconstructed and original mesh.
Results: With the medoid reference mesh, Scalismo required 18 principal components (PCs) to capture 79% of shape variation, while Deformetrica required 22 PCs for 61%. Deformetrica achieved a lower mean reconstruction error (0.93 ± 1.09 mm) than Scalismo (1.29 ± 0.68 mm). The first three PCs explained 23%, 9.2%, and 8.3% in Scalismo and 11%, 7.4%, and 4.9% in Deformetrica. Both frameworks performed better with the medoid than the mean-volume reference mesh, with mean errors rising to 1.77 mm and 2.16 mm respectively.
Conclusion: The two frameworks present a trade-off. Scalismo produces a more compact representation suited to statistical testing or cohort stratification. Deformetrica offers greater geometric fidelity where individual reconstruction accuracy is the priority. A geometrically central reference mesh improves performance in both.