Why the template-based deformation method performs significantly worse on the SCOT-HEART dataset compared to the ResNet Decoder + Marching, as reported in the paper?

[chaoran4532]

Why the template-based deformation method performs significantly worse on the SCOT-HEART dataset compared to the ResNet Decoder + Marching, as reported in the paper? CorticalFlow has high DSC and CH metrics in cortical reconstruction. Is this performance difference due to the smaller number of points used in your method, or is it related to the inherent characteristics of the dataset itself?

[MalikTeng]

Hi @chaoran4532,

Please accept my sincere apologies for the incredibly late response; I truly appreciate your patience.

To clarify the technical differences, the ResNet Decoder operates in voxel space to generate segmentations, from which the Marching Cubes method extracts the mesh surface. Because this process is voxel-based, it scales significantly better with higher resolutions.

In contrast, template-based methods rely on warping a template to minimise the Chamfer distance to the ground truth. While effective at reducing average error, this approach often lacks the fidelity required for fine geometric details.

I believe the performance gap is due to these inherent limitations in template-based warping. We have invested considerable time into a solution for this, which is presented in our latest research, MorphiNet.

Again, I am very sorry for not responding sooner.

Regards,
@MalikTeng