Training segmentation and objects besides mitochondria

[Rick Giuly]

Good questions,

One note: You'll probably want to switch to using settings2.py in the
command line. It tells the system to count any nonzero pixels in the
training segmentation as the object you want to segment.

For training segmentation you just want to paint all the mitochondria in
solid. (If you outline them in IMOD, imodmop fills them in when it
creates the raster images.) You don't need to mark anything else. You
might see some other objects in the example training data, but those
were used for different tests.

Easiest way to detect other objects is to just create another training
set for the other object and run another process to handle it. Let me
know if you want to try this, and I can give some advice.


-Rick

kurt weiss wrote:
> Rick,
> A few more fundamental questions have arose:
>
> The segmentation training set looks like 'all' membranes are outlined
> (cell membranes and mito membranes, but not vesicles):
> 1. If this is done by hand, why not outline only mitochondria? (is it to
> avoid false negatives during the inital pixel classification, allowing
> you to eliminate false positives in each subsequent contour analysis step)
> 2. Can cytoseg differentiate by more than just two types (mito and
> no-mito)? I'm assuming yes, but you'd have to change some parameters
> (3.where would these files be?), so it will only differentiate two (mito
> and non) classes at a time.
>
> Regards,
> Kurt
>
> On Wed, Oct 31, 2012 at 2:53 PM, Rick Giuly <rgi...@gmail.com
> <mailto:rgi...@gmail.com>> wrote:
>
>
> Thanks! I'll add that to the wiki docs
>
> -Rick
>
>
> kurt weiss wrote:
>
> Rick,
>
> You might consider posting this and drawing some attention to it:
> http://manojbits.wordpress.__com/2012/09/25/installing-__orange-in-ubuntu-12-04/
> <http://manojbits.wordpress.com/2012/09/25/installing-orange-in-ubuntu-12-04/>
> Installing Orange on Ubuntu 12.04 was not straightforward. I was
> happy
> to finally find this. I imagine this will be a stumbling block for
> many of us.
>
> Regards,
> Kurt
>
> --
> .
>
>
>
>
>
> --
> .

[Rick Giuly]

Hi Kurt,

More good questions

> Is it necessary to go to this level of detail on each and every
> mitochondrion in the image? In other words, am I better off with a
> smaller image so I can do this detail for every mito in the image, or
> can I use a larger image, and only highlight a few of them perfectly -
> leaving the rest untouched?
>

You'll typically want to use an image smaller than the full data set.
Cytoseg will use the unmarked areas as "non-mitochondria" examples. So,
if you only mark a few in the stack, it won't be able to learn properly.
I usually make training data like the example shows, with about
700x700x20 at 5nm XY resolution. Our full images are larger, thousands
by thousands.

> Your training segmentation set looks like it was much easier to
> construct - simply highlighting all the dark membrane-like areas - yet
> cytoseg was still able to pick out blobs(mitochondria) - I assume this
> is partly due to to you using settings2.py in the example run. Can
> you breifly explain what the different settings files specialize in?
> and if I am better off following a path similiar to what you did in
> the example , or the steps outlined in Q1 above?

If you check the actual pixel values in the training segmentation,
you'll find that they are different for the mitochondria than for the
outer cell membranes. It's not obvious looking at it as an image. So,
the mito were actually labelled separately. That training data was
created in Amira, and the technique used was thresholding the data as a
starting point and then fixing it up by hand. The settings file tells
the system what values in the training segmentation should be considered
"salient" (mito pixels in this case). If you take a look at the file,
you'll see some examples.

Hope that helps

-Rick