Disambiguate Gaussian-model limitations from data and architecture limitations

Determine whether the observed superiority of the regularised conditional-Gaussian estimator over the tested nonlinear completion models is caused by insufficient sample size, model misspecification, or the selected registration strategy and loss, and establish whether a better nonlinear model, registration strategy, or loss can improve eleven-structure cardiac CT shape completion.

Background

The paper finds that the regularised conditional-Gaussian estimator outperforms the selected graph beta-variational autoencoder and the other tested comparators on the evaluated aligned-CT completion tasks. However, the mixture models, architecture variations, pooling hierarchies, alternative objectives, and larger datasets were not sufficiently explored to determine why the linear estimator performs better or whether the result reflects the limitations of the available data and experimental design.

The authors therefore leave unresolved whether the observed performance gap reflects inadequate sample size, misspecified nonlinear models, or choices in registration and optimisation, and whether a substantially better nonlinear completion system could outperform the conditional-Gaussian baseline.