Hi Juan,

Thank you for your answer, I kind of do what you propose, although it's not very "elegant":

1) Using util.view_as_windows(), I make chunks with overlay.
2) jolib for parallel processing (I'd like to use skimage.util.apply_parallel but I'm not sure how I can use a fonction with more than 1 arg?)
3) this part is a little "bricolage", I take the overlap of adjacent regions and, using regionprops, I can find the "intensity" (label) of 1 region over the other one. Then I just relabel one of the region.
4) I stitch the block back together, again, it's not pretty, I just add up all the regions into a new array, not sure if there is a better way.

So it works well and it's relatively easy to implement because I only have 4 chunks, I've to see how that will go for a big tile of images.

Best,

Cedric

2017-06-28 3:27 GMT-07:00 Juan Nunez-Iglesias <jni.soma@gmail.com>:
Hi Cedric,

Stitching of separate regions is not a trivial problem… You’ll basically have to:

a) overlap the blocks by a certain amount. You might be able to use skimage.util.apply_parallel more easily here than joblib.
b) compute which labels overlap the most between each neighbouring pair of images
c) relabel as necessary to make those labels match

It’s not super-hard, just a lot of annoying bookkeeping. If you come up with a reasonably robust method to do this, I for one would very much welcome such a contribution into the library!

Here’s a paper from an old colleague of mine who had to deal this on a large-scale problem, and they created a robust method to do the matching:
(page 6)

Juan.

On 27 Jun 2017, 9:20 AM +1000, Cedric Espenel <cedric.espenel@gmail.com>, wrote:
Hi,

I'm working on a project where I have to segment out nucleus of cells in a volume. I use random_walker with define markers for every nuclei. If I try to run it on the all volume the segmentation works well but it's very slow. So I used util.view_as_blocks function from scikit to split my image and markers and I loop over the different block using joblib:

rw_dapi_chunks = Parallel(n_jobs = 4)(delayed(segmentation.random_walker)(chunks[i,j],
                                     chunks_markers[i,j], beta = 3000, mode='cg_mg') for
                                      i in range(2) for j in range(2))

So I end up with a list of 4 images and if I combine these images (btw I'm not sure what is the best way for doing that? I just add them up in an empty array) everything work well, except that I end-up with nuclei with 2 different labels:

 Images intégrées 1
Am I doing something wrong? Is there a way to split your images in block for parallel computing but being able to get the labeling right at the end?

Best,

Cedric
_______________________________________________
scikit-image mailing list
scikit-image@python.org
https://mail.python.org/mailman/listinfo/scikit-image

_______________________________________________
scikit-image mailing list
scikit-image@python.org
https://mail.python.org/mailman/listinfo/scikit-image