http://imagej.273.s1.nabble.com/Cell-counting-with-imageJ-tp3690375p3690377.html
You're right, I never thought of that. For the sample images, users could mark the cells' locations with the ImageJ manual cell counter plugin. Then it runs the tests. If there are multiple particles close to the marker (like when cell is fragmented into 10 pieces), those are scored as a duplicate count. If a particle is not close to a marker, that is scored as a false count. Closeness is defined as 'inside the user-set cell radius'. So the tests should measure # of particles counted, number of duplicate counts, number of false counts. It would choose the image processing algorithm that optimizes all three. The user would not have extra work since he has to do some manual counting anyway.
> Date: Sat, 21 Nov 2009 09:17:39 +0000
> From:
[hidden email]
> Subject: Re: Cell counting with imageJ
> To:
[hidden email]
>
> On Friday 20 November 2009, Lukas Hoffmann wrote:
> > My basic idea is, you count a few images by hand and tell the program what
> > the count "should" be. Then the program changes the way it does the
> > thresholding & image processing in many different ways, and
> > compares all of its test counts with the correct count. When the
> > program finds one type of image processing that has a low error rate
> > compared to the manual counting, it uses that type of processing for
> > the rest of your images.
>
> The problem with that approach is that it is counting objects without
> knowing whether they are the same in the gold-standard count and in the test
> count.
>
> If you have 10 objects (each one is a cell) in the gold standard and only 1 of
> those cells is recognised in the test image, but the cell is fragmented by the
> segmentation procedure in 10 pieces, you would be assuming that the program is
> counting perfectly when it is not.
>
> Counts alone is not enough. You need to know that the segmented blobs somewhat
> match (in position, size, shape) the gold standard blobs, only then you could
> count them, but otherwise the assumption that they are the same in both images
> is not satisfied.
>
> Regards
>
> G.