DG surface contains self-intersection

[debinz]

Dear professionals of Hippunfold,

I employed this tool to generate a midthickness hippocampus surface, selecting a 2mm resolution. While examining the results I found that the dentate gyrus surface exhibited self-intersection indicating that an edge of the mesh contacts the surface itself(see accompanying figure). This case would impact the subsequent feature sampling like BOLD signals and other geometric analyses.

I used the latest version of v1.2.1 and run this tool using singularity. Could you kindly offer me some advice to deal with this issue?

Best regards, Debin

[jordandekraker]

Hi Debin, Thanks for pointing this out, its true that the DG surfaces are not always of great quality from HippUnfold using in-vivo MRI. These outputs should maybe be considered more of a proof-of-principle than something to base an experiment around, because of the following consderations:

• DG labels are typically only 1 voxel (0.3mm isotropic) thick, and so in some cases the transform between folded and unfolded space suffers a loss of information due to voxelization/discretization errors (which is probably what led to the misplaced vertices in your case).
• In reality (eg from histology), we know that the DG is far more convoluted than the relatively smooth, simple surfaces here. This isn't readily visible (either to a human or to a NN like the one HippUnfold uses) in typical scans, and instead the labels generated here is based on a group-averaged prior.

Thus the DG surface should be considered only a rough approximation. We thought it was helpful to include it so that people could understand that i) its a discrete surface from the rest of the hippocampus, and ii) it interlocks (or wraps around the edge) of the rest of the hippocampus. Since we're fairly happy with the segmentation quality of the rest of the hippocampus, another way of thinking of how the DG is labelled here is that its labelled by excludion (of the other, more well-defined subfields and SRLM). Hopefully future advances will make it possible to generate better subject-specific DG surfaces, but for now I'm afraid image resolution/quality are a major hurdle.

[akhanf]

I'll also add that these issues are less prominent if a denser surface sampling is used (e.g. 0p5mm), so I would try that first.

[debinz]

I am thankful for your detailed response and clarification. I found your perspectives regarding the restrictions of current DG surfaces generated from HippUnfold to be very beneficial and enlightening. I also deeply appreciate the recommendation to experiment with a denser surface sampling. Although I did produce a 0.5mm resolution DG surface which does exhibit fewer issues, ultimately, I require a 2mm surface to map fMRI signals and ensure good vertex correspondence across individuals.

This raises the following question:

Can the 0.5mm DG surface be downsampled to 2mm to maintain satisfactory vertex correspondence with a template 2mm DG surface (contains 64 vertices)? I am uncertain if merely decreasing the vertex density would correctly align the vertices, given the discretization problems existent.

[jordandekraker]

Happy to help, and yes that makes sense. You can downsample from 0.5mm to 2mm surfaces using this tool: https://github.com/jordandekraker/hippunfold_toolbox/blob/ed615fbb7a8685abfa270aae38d99bacd581c8cd/hippunfold_toolbox/utils.py#L113

However, keep in mind there may still be some partial-voluming of DG vertices with surrounding subfields if the fMRI voxel size is large. You may want to consider combining the DG and CA4 labels, as in this atlas from the same toolbox: https://github.com/jordandekraker/hippunfold_toolbox/tree/main/resources/parc-DeKraker25

[debinz]

Great! Thank you for sharing the tool and the atlas. This will definitely be helpful for downsampling and considering the combination of DG and CA4 labels. Appreciate your response!

[jordandekraker]

Closing, but let me know if you have further issues and I can reopen!