[Insight-users] 3D Watershed segmentation speed

[Kalle Pahajoki]

Hi

I'm developing a tool based on the watershed segmentation for the 
segmentation of 3D confocal microscopy images, and so far it seems to be 
very suitable for my purposes. The problem with the method is it's 
speed: images I'm working with are typically in range  of512x512x24 - 
1024x1024x24 voxels. For these images, the method is too slow to be usable.

So far I haven't had the patience to run the method on a full sized 
dataset (because I'm working on the software, and not producing any 
meaningful results) but always resorted to scaling the dataset to the 
range 128x128x24. For the actual users, however, I cannot really justify 
this kind of loss of information.

I've usually used the VTK hybrid 2D median for the filtering (since the 
anisotropic diffusion recommended in the ITK book is slow as well 
*grin*). I usually test with Threshold = 0.01 and Level = 0.2. Could 
there be significant speed gains in tweaking the parameters or filtering?

In general, is there any way to speed the method, other than resampling 
the datasets to a smaller size?

Some software used here (e.g. 3D for LSM by Zeiss) offer a similiar 
method with almost interactive speeds (or at least speeds that the users 
can accept). Do they employ a fundamentally different technique?

Kalle