Finding rings is indeed harder and less common than finding blobs. It can be easier when the rings are very regularly shaped, but you seem to have some variety here. Have you tried inverting your image and then looking for blobs? I'm not saying it will work fantastically well, but given that each cell is a dark circle surrounded by a brighter ring, it could work together with some filtering steps.

Did this in my thesis (long ago:-). Your cells look like those I hd to work with. If you like, check out

Article A neural network architecture for automatic segmentation of ...

Did this in my thesis (long ago:-). Your cells look like those I hd to work with. If you like, check out ...

Article Human vs. Machine: Evaluation of Fluorescence Micrographs

If you just want to make a statistics from many IF images, not the exact accurate number (that should not be possible), you can use the stereological method. See attached literatures. Measure the total area of your cells in one IF image and then get the count by dividing the averaged area of a single cell. Hope it works.

have a look at the plugin Object J for the program imageJ both are free ware:

http://simon.bio.uva.nl/objectj/

Sorry for forgetting uploading the attached literature in my last response.

The latest version of Imaje J software has a function that make counting alot easier try it and see.

Dear Mikael Niku,

You can contact my Ph.D.student, Fausy Dias, who developed an instrument to count bacteria colonies, at [email protected].

Best regards,

Marcelo S. Alencar

Wow, so many great suggestions in so little time! Thanks for all the ideas an the literature. I did try the inversion trick, but it seems that the typical segmentation algorithms still expect objects with a darker core and lighter borders, so couldn't make it work. The area substraction/averaging method might work well enough, and I will also try the other tools and methods suggested. Tim, are any neural network classifiers available as ready-to-use software or how does it work in practice? Sanusi, which function do you mean? Adam, I'm sorry but I unfortunately have no experience in Matlab.

We have applied different supervised learning architectures, like SVMs etc. Those should be easy to get in open source code or in MATLAB module or via R. The features were computed using PCA or ICA, both are available in many ways. Start with PCA. If you have more then one fluorescence channel you may also have a look at ...

Article Integrating semantic annotation and information visualizatio...

Hi Mikael,

It sounds like you need a solution that does not require computational expertise to employ? I played around with your image in CellProfiler and came up with the attached results. I'm not claiming it solves your problem but perhaps it gets your farther down the path and at least it's in a software package that is free and you can tweak yourself.

It does decently on the cells that are not in the big clump at the bottom middle and I'm sure you could improve on those results given more effort; this was just a quick try. As far as the clump goes, I'm not sure you can expect lots of improvement with this kind of pipeline, even with tweaking. There are regions that are pretty difficult to count by eye within the clump. Perhaps someone else's idea of measuring the area occupied by cells and dividing by typical cell size is the best one - it's simple anyway! You can use the MeasureAreaOccupied module of CellProfiler to do that. You can ask questions about CellProfiler at its online forum at www.cellprofiler.org

(BTW, my idea of inverting was a bad one - there were imperceptible variations in the background staining that result in 'dark holes' everywhere in the image, so trying to identify cells via dark holes didn't work well at all).

Attached is the CellProfiler pipeline that achieved those results; you can download the software at www.cellprofiler.org

Thanks Anne! That looks already better than what I achieved in Profiler after quite some time. May I ask two details here: how do you make the objects round like that? And is it somehow possible to tell the software that the objects need to be of certain circularity (like in the ImageJ particle detection)? I found only the size-based criteria, which seems rather limited to me.

Glad I could help! I can't devote more time right now, but if you post to the CellProfiler Forum, someone else on the team might be able to help you further along; you could just link to this post to introduce the topic. (I'm not sure what you mean about making them round - the algorithm smooths the image then looks for peaks in the intensity in order to divide clumped objects. This tends to produce round objects when the cells look like yours). Regarding your other question, I'm not certain whether it's possible to implement a circularity criterion... I suspect there might be a clever way to do it in the software already, but if not that's a great suggestion to make on the forum as well.

Hi Mikael

Do you have any idea of how many cells you got in the example image (manual counting)?

I think it is always good to have some kind of truth when you looking for a segmentaion and counting algorithms.

I did a rough estimation by removing the background and integrate the intensity of the image and a template cell. Dividing the sum of the image with the sum of the template cell gave me that the image contains 1000-1500 cells. The number depending of how well you pick a template cell and estimating of the background. I would say that the cluster of cells is hard to count with segmentation methods.

Thanks Anders, yes we sure are counting a selection of sample images manually to see if the "automatic" methods are correct at all. For the example image, I got about 580 cells manually, so in this case the averaging seems to give a somewhat too large estimate. But of course that depends a lot on how you set the threshold... it seems to be quite difficult as the signal intensity varies a lot between positive cells.

Dear Mikael Niku,

You can read this paper (on segmentation of DAPI and FITC immunofluorescence images). For me, a good segmentation is not enought. You must use a pattern for count the cells on each aggregate or overlap cells group.

Article Segmentation of complex nucleus configurations in biological images

Thanks for the interesting paper. That's exactly what I started thinking when tinkering with my images! Unfortunately I lack a suitably ready-to-use tool for the job, as I'm not able to do any real programming at least in the time available. Haven't found any usable plugins like this for ImageJ or CellProfiler at least.

Honestly the image given is not very good (JPEG artifacts and so on), so it is hard to judge. A simple method in your case could be to completely ignore cell boundaries, instead focusing on their average intensity. Assuming cells are fairly transparent, their fluorescence luminosity adds up, so if you can work out an intensity for individual cells based on the few that are already individualized at the top of your image, then by considering the total intensity of the image, dividing by the average intensity yields a likely total number of cells. This would have to be validated of course but this is a very simple method.