Posted by
Herbie on
Jan 07, 2018; 1:12pm
URL: http://imagej.273.s1.nabble.com/Weka-Trainable-Segmentation-alternative-tp5019833p5019839.html
Good day Adrián Villalba Felipe,
yesterday I took a closer look at your sample image data and, as you've
already conceded, the staining is not specific enough.
I found that a reasonable contrast of the relevant islets is obtained by
the magenta channel after transforming from RGB to CMYK colour space. In
general, your sample images contain very little colour information, i.e.
if you find a suitable colour space, the information is mostly contained
in a single channel. In order to find the optimum colour space (CMYK
isn't bad but perhaps not optimum) I recommend to experiment with the
"Colour_Deconvolution"-plugin.
Using the magenta channel I wasn't able to find an automatic
thresholding scheme that worked with all of your five sample images. A
threshold-based segmentation of sample image "wEKA_INSULITIS0005.tif"
was impossible even with non-automatic thresholding.
WEKA-classification based on colour may fail due the reasons mentioned
above. Based on structural features, a segmentation may be possible if a
large and adequate training-set is used.
All of the above holds for segmentation of the islets from the surround,
not for the classification of the types of islets which is much more
involved and I doubt that the latter is possible with reasonable success
for the present staining.
In short: A more specific staining appears to be required.
Regards
Herbie
:::::::::::::::::::::::::::::::::::::::::::::
Am 06.01.18 um 23:24 schrieb Adrián Villalba:
> Dear Ignacio,
>
> This is an example image with the different types of islets and insulitis.
> The legend is the following:
> A = grade 0, no insulitis; B = grade 1, peri-insular; C = grade 2, mild
> insulitis (<25%
> of the islet infiltrated); D = grade 3, moderate insulitis (25–75% of the
> islet
> infiltrated); E = grade 4, severe insulitis (>75% islet infiltration).
>
> I would like to compare my template images against a new image
> automatically, so i do not have to do it manually in the microscope. But
> Weka has not been useful, first of all because Hematoxilin/Eosin staining
> is not very different between islets and the rest of tissue. Moreover, the
> different score of the 5 types of islets (scored from 0 to 4) is also
> tricky for the Weka algorithm. Do you think i can use a different approach?
>
>
> Thank you very much for your time and attention,
>
> <
https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
> Libre
> de virus. www.avast.com
> <
https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
> <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
>
> 2018-01-06 0:39 GMT+01:00 Adrián Villalba <
[hidden email]>:
>
>> Dear all,
>>
>> I am trying to use the Weka Trainable Segmentation plugin in order to
>> classify islets of Langerhans in Hematoxilin/Eosin tissues. There are four
>> different types of islets depending on insulitis (immune cell infiltration
>> within the islet) as i show you in an attached JPG-picture (just to show
>> you the expected result, not to manipulate).
>>
>> My goal is to do it automatically in imageJ, rather tan manual scoring of
>> pictures. So i thought it would be a good idea to use the Trainable Weka
>> Segmentation plugin just to train the algorithm to do it automatically but
>> it fails. (I cannot attach the classifier.model archive because it is
>> rejected by the mailing list conditions).
>>
>> I think that maybe it is not a proble for the Weka, instead being a
>> conceptual problem and that maybe you could know another imageJ tool to
>> pursuit that goal.
>>
>> Thank you very much in advance!
--
ImageJ mailing list:
http://imagej.nih.gov/ij/list.html