Segmentation of granular "stuff"

[Jolien Connor]

Hi,

I am trying to do something similar to what many try
to do with thresholding -- find a way for the software
to pick out my pixels of interest away from the
background.  In my case, it is granular "stuff" that
are intracellular inclusions in the brain.  We have
many image montages taken of mouse brain sections and
we are trying to quantitate the percentage of specific
brain regions occupied by the inclusions (percent
volume).  Some of the monochrome images are here:

www.waisman.wisc.edu/cmn/mutant high load.tif
www.waisman.wisc.edu/cmn/mutant high load and
faint.tif
www.waisman.wisc.edu/cmn/mutant low load and faint.tif
www.waisman.wisc.edu/cmn/control.tif

They are very large since they are montages; contact
me at [hidden email] if you want them and
can't get them.  

If you look at the pictures of the mutants compared
with the control, you can see what they are -- the
bright granular stuff all over the place.  Cells that
contain the inclusions sometimes have processes that
show up, and we only want to pick out the brightest,
most granular stuff, which are the inclusions.  We
want a method that can tell the difference between
mutants and controls, as well as between high load
mutants and low load mutants (lots of them versus just
a few).  

The biggest problem right now seems to be that some of
our pictures are faint, possibly because the stain
weakened, our mercury bulb Atto Arc controller is
dying, or some other reason.  As humans, we can still
pick them out, but we don't have a thresholding
algorithm that can handle this problems.  

We have been using the Subtract Background function,
then thresholding the image, and can't get consistent
results.  We use a brightfield image that lines up
with each monochrome image to trace our regions, then
have ImageJ calculate the % Area occupied by pixels
above threshold.  There's no threshold we could pick
that would correctly differentiate between controls,
low load mutants and high load mutants.  Any ideas?

Thanks,

Jolien Connor
University of Wisconsin

[Michael Schmid]

Hi Jolien,

first concerning the uploaded images - the blanks in the filenames
should be written "%20" in URLs. So, e.g., it is

www.waisman.wisc.edu/cmn/mutant%20high%20load.tif

Concerning faint images: There are plugins that may solve the
problem, e.g. the Stack NOrmalizer
   http://rsb.info.nih.gov/ij/plugins/normalizer.html
and the Background Subtraction and Image Normalization plugin
   http://rsb.info.nih.gov/ij/plugins/normalize.html

Another way to do it might be saturating a small percentage
of the pixels (Brightness&Contrast/Auto) and applying this as
new grayscale (8-bit images only). The Macro Recorder will show
you how this translates into macro code.

If some images don't have the bright granular "stuff" at all,
you may have to analyze at the histogram, e.g. determine the
"mode" (grayvalue of the highest peak, usually some kind of
background) and draw your conclusions from that (maybe divide
by some value proportional to the "mode")

Michael
________________________________________________________________

On 6 Jul 2007, at 20:42, Jolien Connor wrote:

> Hi,
>
> I am trying to do something similar to what many try
> to do with thresholding -- find a way for the software
> to pick out my pixels of interest away from the
> background.  In my case, it is granular "stuff" that
> are intracellular inclusions in the brain.  We have
> many image montages taken of mouse brain sections and
> we are trying to quantitate the percentage of specific
> brain regions occupied by the inclusions (percent
> volume).  Some of the monochrome images are here:
>
> www.waisman.wisc.edu/cmn/mutant high load.tif
> www.waisman.wisc.edu/cmn/mutant high load and
> faint.tif
> www.waisman.wisc.edu/cmn/mutant low load and faint.tif
> www.waisman.wisc.edu/cmn/control.tif
>
> They are very large since they are montages; contact
> me at [hidden email] if you want them and
> can't get them.
>
> If you look at the pictures of the mutants compared
> with the control, you can see what they are -- the
> bright granular stuff all over the place.  Cells that
> contain the inclusions sometimes have processes that
> show up, and we only want to pick out the brightest,
> most granular stuff, which are the inclusions.  We
> want a method that can tell the difference between
> mutants and controls, as well as between high load
> mutants and low load mutants (lots of them versus just
> a few).
>
> The biggest problem right now seems to be that some of
> our pictures are faint, possibly because the stain
> weakened, our mercury bulb Atto Arc controller is
> dying, or some other reason.  As humans, we can still
> pick them out, but we don't have a thresholding
> algorithm that can handle this problems.
>
> We have been using the Subtract Background function,
> then thresholding the image, and can't get consistent
> results.  We use a brightfield image that lines up
> with each monochrome image to trace our regions, then
> have ImageJ calculate the % Area occupied by pixels
> above threshold.  There's no threshold we could pick
> that would correctly differentiate between controls,
> low load mutants and high load mutants.  Any ideas?
>
> Thanks,
>
> Jolien Connor
> University of Wisconsin

[Dimiter Prodanov]

Hi,

If you are not specifically interested in counts you can do granulometry
or granulometric filtering to isolate the "particles" of specific sizes.
Then you can threshold and obtain the area fraction within the ROI of
interest.

Greetings

Dimiter Prodanov

> On 6 Jul 2007, at 20:42, Jolien Connor wrote:
>
>> Hi,
>>
>> I am trying to do something similar to what many try
>> to do with thresholding -- find a way for the software
>> to pick out my pixels of interest away from the
>> background.  In my case, it is granular "stuff" that
>> are intracellular inclusions in the brain.  We have
>> many image montages taken of mouse brain sections and
>> we are trying to quantitate the percentage of specific
>> brain regions occupied by the inclusions (percent
>> volume).  Some of the monochrome images are here:
>>
>> www.waisman.wisc.edu/cmn/mutant high load.tif
>> www.waisman.wisc.edu/cmn/mutant high load and
>> faint.tif
>> www.waisman.wisc.edu/cmn/mutant low load and faint.tif
>> www.waisman.wisc.edu/cmn/control.tif
>>
>> They are very large since they are montages; contact
>> me at [hidden email] if you want them and
>> can't get them.
>>
>> If you look at the pictures of the mutants compared
>> with the control, you can see what they are -- the
>> bright granular stuff all over the place.  Cells that
>> contain the inclusions sometimes have processes that
>> show up, and we only want to pick out the brightest,
>> most granular stuff, which are the inclusions.  We
>> want a method that can tell the difference between
>> mutants and controls, as well as between high load
>> mutants and low load mutants (lots of them versus just
>> a few).
>>
>> The biggest problem right now seems to be that some of
>> our pictures are faint, possibly because the stain
>> weakened, our mercury bulb Atto Arc controller is
>> dying, or some other reason.  As humans, we can still
>> pick them out, but we don't have a thresholding
>> algorithm that can handle this problems.
>>
>> We have been using the Subtract Background function,
>> then thresholding the image, and can't get consistent
>> results.  We use a brightfield image that lines up
>> with each monochrome image to trace our regions, then
>> have ImageJ calculate the % Area occupied by pixels
>> above threshold.  There's no threshold we could pick
>> that would correctly differentiate between controls,
>> low load mutants and high load mutants.  Any ideas?
>>
>> Thanks,
>>
>> Jolien Connor
>> University of Wisconsin
>
> ------------------------------
>