Skeleton extraction in this code only calls skimage.morphology.skeletonize (2D) directly—no custom 3D processing implemented

[CODE-BULIAO]

Hi, thanks for sharing your code!

3D Skeletonization Not Actually Implemented in Provided Code?

I noticed that in your current skeleton extraction implementation, you are calling skeletonize directly on a 3D array (e.g., bin_seg[0]). However, skimage.morphology.skeletonize only supports 2D arrays, so this code does not actually process the 3D volume as intended. Also, the code doesn't seem to iterate over individual slices (for example, along the z-axis) to do per-slice skeletonization.

Could you clarify:

Is there a reason you didn't use either 3D skeletonization (like skeletonize_3d) or at least loop over each 2D slice for slice-wise processing? If there's an update planned for actual 3D skeletonization support?

How is this implementation different from your previous skeleton extraction approach?

Would love to understand your thoughts on this—thanks!

[mrokuss]

Hi @CODE-BULIAO, thanks for your interest and for taking the time to review the code!

To clarify, the code of skimage.morphology.skeletonize actually does use 3D skeletonization, specifically it will call the appropriate internal function depending on the dimensionality of the input image. If the input is 3D, the function dispatches to the correct method here:

elif image.ndim == 3 or (image.ndim == 2 and method == 'lee'):
    skeleton = _skeletonize_lee(image)

Source: https://github.com/scikit-image/scikit-image/blob/e8a42ba85aaf5fd9322ef9ca51bc21063b22fcae/skimage/morphology/_skeletonize.py#L90

So when a 3D binary volume is passed in (e.g., bin_seg[0]), it correctly performs 3D skeletonization using the Lee et al. method. In our pipeline, we use this "off-the-shelf" 3D skeletonization as a first step to obtain the tubular skeleton. You can see how this is done here:

Skeleton-Recall/nnunetv2/training/dataloading/data_loader_3d_skel.py

Line 9 in f259f63

def calculate_tubed_skeleton(seg_all, do_tube):

So to clarify:

• Yes, 3D skeletonization is supported and used in our code via skimage.morphology.skeletonize.
• There's no need to loop over 2D slices since the function natively handles 3D input.
• This implementation uses a more direct and robust 3D approach compared to any slice-wise alternatives.

Hope this clears things up, and thanks again for the thoughtful questions!