retinal vasculature segmentation?

[Kenneth Sloan-2]

One of my current projects involves segmenting out the vasculature in images of
human retina (these happen to be autofluorescence images, where the blood vessels
appear as nearly black).

Histograms of even very local pieces of image show two
very overlapping populations of pixels - I have seen recently reported methods
which simply choose a threshold based on these histograms.  
This is “close, but no cigar”, for our purposes.

My current, crude approach uses auto local thresholding plus a bit of mathematical morphology,
followed by manual editing.   My current push is to make the manual editing phase a bit
easier - but that part is under control.

This approach is adequate for my collaborators’ current needs, so I’m now free to explore
an implementation that will be completely automatic.

While I’m happy with our current results, I’d like to compare with (and perhaps learn from)
existing implementations done elsewhere.

If you have an existing (successful) implementation in ImageJ, can you please point
me at results (publications would be outstanding) and/or implementation (ImageJ macros
or plugins) - or even just clues about the approach you have used.

Single image, standard (or at least known) position and pose.  The center of the fovea is known and near the center of the image, the position of the edge of the optic disc is known and near the edge (either to the right, or the left - if it helps the orientation can be trivially standardized).  8-bit gray input; binary output.  Known issues are the overlap in pixel values for the two categories (blood vessel/NOT blood vessel), a noisy, dark avascular zone around the center of the fovea, and some falloff in both the signal AND the optics in the corners (and, to a lesser extent, at the edges).
Ideally, I’d like to find all pixels completely covered by a blood vessel…and no others.  Oh yes…there are also dark patches of image that are due to disease, and not blood vessels - so some shape analysis appears to be necessary.
That’s what we use the manual editing for, now - both to thicken and extend blood vessels that are missed by the automatic procedure AND to erase patches accepted by the automatic procedure which are “obviously” NOT blood vessels.
Unfortunately, this latter category is (of course) the most interesting, and where all the “results” will be found.

--
Kenneth Sloan
[hidden email]

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[Ignacio Arganda-Carreras]

Dear Kenneth,

Maybe you can give a try to the Trainable Weka Segmentation plugin:
http://fiji.sc/Trainable_Weka_Segmentation

In any case, it would help if you post here some images so we can get a
better idea of the segmentation problem.

Best,

ignacio


On Tue, May 20, 2014 at 3:52 AM, Kenneth Sloan <[hidden email]>wrote:

> One of my current projects involves segmenting out the vasculature in
> images of
> human retina (these happen to be autofluorescence images, where the blood
> vessels
> appear as nearly black).
>
> Histograms of even very local pieces of image show two
> very overlapping populations of pixels - I have seen recently reported
> methods
> which simply choose a threshold based on these histograms.
> This is “close, but no cigar”, for our purposes.
>
> My current, crude approach uses auto local thresholding plus a bit of
> mathematical morphology,
> followed by manual editing.   My current push is to make the manual
> editing phase a bit
> easier - but that part is under control.
>
> This approach is adequate for my collaborators’ current needs, so I’m now
> free to explore
> an implementation that will be completely automatic.
>
> While I’m happy with our current results, I’d like to compare with (and
> perhaps learn from)
> existing implementations done elsewhere.
>
> If you have an existing (successful) implementation in ImageJ, can you
> please point
> me at results (publications would be outstanding) and/or implementation
> (ImageJ macros
> or plugins) - or even just clues about the approach you have used.
>
> Single image, standard (or at least known) position and pose.  The center
> of the fovea is known and near the center of the image, the position of the
> edge of the optic disc is known and near the edge (either to the right, or
> the left - if it helps the orientation can be trivially standardized).
>  8-bit gray input; binary output.  Known issues are the overlap in pixel
> values for the two categories (blood vessel/NOT blood vessel), a noisy,
> dark avascular zone around the center of the fovea, and some falloff in
> both the signal AND the optics in the corners (and, to a lesser extent, at
> the edges).
> Ideally, I’d like to find all pixels completely covered by a blood
> vessel…and no others.  Oh yes…there are also dark patches of image that are
> due to disease, and not blood vessels - so some shape analysis appears to
> be necessary.
> That’s what we use the manual editing for, now - both to thicken and
> extend blood vessels that are missed by the automatic procedure AND to
> erase patches accepted by the automatic procedure which are “obviously” NOT
> blood vessels.
> Unfortunately, this latter category is (of course) the most interesting,
> and where all the “results” will be found.
>
> --
> Kenneth Sloan
> [hidden email]
>
> --
> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
>

--
Ignacio Arganda-Carreras, Ph.D.
Seung's lab, 46-5065
Department of Brain and Cognitive Sciences
Massachusetts Institute of Technology
43 Vassar St.
Cambridge, MA 02139
USA

Phone: [phone]
Website: http://bioweb.cnb.csic.es/~iarganda/index_EN.html

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html