http://imagej.273.s1.nabble.com/Segmentation-of-DIC-or-Hoffman-Modulation-Contrast-images-help-please-tp5012661p5012666.html
I would consider an approach that is one level higher than raw image processing. DIC images produce distinctive doublets (a bright region adjacent to a darker region - always (for the same setup) at the same angle.
Once you can identify a cell from it’s doublet, and have a reasonable guess as to its size, finding the boundary and estimating area should become easier.
My first try might be to design a custom convolution kernel that responds preferentially to doublets at a fixed angle. There is still the problem of estimating the angle (if it’s not stable) but that should be doable based either on meta-information or user input.
I second the idea of using a texture measurement to eliminate regions of the image that are out of focus (variance is one - I would perhaps go a step further and use coefficient of variation - stdDev/mean - but there are others)
> On May 1, 2015, at 08:02 , Aryeh Weiss <
[hidden email]> wrote:
>
> On 5/1/15 8:48 AM, Jacqui Ross wrote:
>> Hi Aryeh,
>>
>> Thanks for your reply. I did try using a variance filter (the built-in one under Process - Filters - Variance) with different radii but I was unable to achieve a good result. The resultant circles were often incomplete so that when I then converted to binary, I had to do a lot of additional processing (Closing, filling holes, etc.) and then the outlines weren't very accurate.
> Yes -- these methods are better at marking objects than getting accurate boundaries. You might be able to use the inaccurate segmentation that produces as a mask against the original variance image, which produces reasonable arcs around your in-focus cells.
>> I know that you presented on the Trainable Weka Segmentation at the ImageJ conference that I attended a couple of years ago. My notes on that weren't fantastic (I pulled them out!) but in your case you also had some fluorescence labelling to help inform the segmentation.
> That problem was easier than yours (isn't it always that way?). The texture was much better defined, and I did not have a continuum of out-of-focus or partially out-of-focus cells. There was a DAPI channel which I used to exclude artitfacts which were not cells (since they did not contain nuclei). However, that will not help you, since your out of focus cells are still cells.
>
> You might have an easier time here if you can acquire a z-stack (even though it is wide-field) and use the EDF plugin or similar to have all of your cells in-focus. That would be easier to segment. Alternatively, if you choose to be very strict in your segmentation (ie, take only the cells that are really sharp and textured), then I think you will be able to variance and it relatives to get an accurate segmentation and boundary.
>
> Best regards
> --aryeh
>
>>
>> -----Original Message-----
>> From: Aryeh Weiss [mailto:
[hidden email]] On Behalf Of Aryeh Weiss
>> Sent: Friday, 1 May 2015 5:07 p.m.
>> To: Jacqui Ross
>> Subject: Re: Segmentation of DIC or Hoffman Modulation Contrast images - help please
>>
>>
>> Try converting to 8-bit and running a variance filter
>> (Process>Filters>Variance...) followed by thresholding. This will enhance the cells in focus due to their texture.
>>
>> --aryeh
>>
>> On 5/1/15 7:37 AM, Jacqui Ross wrote:
>>> Hi All,
>>>
>>> I'm helping a PhD student with analysing some Hoffman modulation contrast images of cells. She's primarily interested in changes in diameter. The cells are embedded in a 3D matrix and compression is being applied.
>>>
>>> In the images, there are nice cells in focus with clear boundaries, plus others which are out of focus which we don't want to measure as any measurements won't be accurate.
>>> These kind of images are really tricky to segment as anyone who has tried, already knows. I've tried lots of different filters (edge, etc.) , FFT filtering and the Trainable Weka Segmentation but have been unable to achieve good enough results to be able to then threshold the cells automatically.
>>>
>>> I've come to the end of the line for now so am asking for your expert help in case anyone has some suggestions:). I note that in 2006 Monique Vasseur offered some DIC images to a PhD student called Daniel Mauch in Germany but I'm not sure if anything came of that project. There are a few papers out there (some mention Hilbert Transform, FFT) but I haven't been successful in implementing anything from those papers as yet.
>>>
>>> In the meantime, my solution is to use the Pseudo flat field correction plugin from Jan Brocher's BioVoxxel Toolbox (Thanks Jan!) with a very small radius (5) to flatten the background and out of focus cells while preserving the in focus cell outlines. We can then use the Cell Magic Wand (Thanks Theo!) to create selections that can be loaded into the ROI Manager and then measured. This works really well but requires that the cells be selected manually.
>>>
>>> The Cell Magic Wand Tool works on the colour or grayscale so we can also split the channels from the colour image if needed and use the channel image with the most contrast.
>>>
>>> I've attached an image in case anyone has any ideas. The image has been cropped out of a larger image so that it's not too big and it's pink because there's cell culture medium there (in case anyone was wondering..).
>>>
>>> Look forward to hearing any suggestions!
>>>
>>> Kind regards,
>>>
>>> Jacqui
>>> Jacqueline Ross
>>> Biomedical Imaging Microscopist
>>> Biomedical Imaging Research Unit
>>> School of Medical Sciences
>>> Faculty of Medical & Health Sciences
>>> The University of Auckland
>>> Private Bag 92019
>>> Auckland 1142, NEW ZEALAND
>>>
>>> Tel: 64 9 923 7438
>>> Fax: 64 9 373 7484
>>>
>>>
http://www.fmhs.auckland.ac.nz/sms/biru/>>>
>>>
>>> --
>>> ImageJ mailing list:
http://imagej.nih.gov/ij/list.html>>
>> --
>> Aryeh Weiss
>> Faculty of Engineering
>> Bar Ilan University
>> Ramat Gan 52900 Israel
>>
>> Ph: 972-3-5317638
>> FAX: 972-3-7384051
>>
>>
>> .
>>
>
>
> --
> Aryeh Weiss
> Faculty of Engineering
> Bar Ilan University
> Ramat Gan 52900 Israel
>
> Ph: 972-3-5317638
> FAX: 972-3-7384051
>
> --
> ImageJ mailing list:
http://imagej.nih.gov/ij/list.html