Tutorial for cell classification according to stained cytoplasm.

Tutorial for Cell Classification (RFP positive/negative)
Tutorial for cell classification according to stained
cytoplasm.
http://www.nexus.ethz.ch/ -> Software -> TMARKER
Consider the image on the left. We want to classify RFP (cytoplasmic marker) positive and negative
cells (on the right). Note that the cell nuclei are visible bluish with a hematoxylin staining.
PREREQUISITES:
•
•
Java 1.7 (Runtime Environment JRE).
Tissue images of IHC stained tissue (see DemonstrationData.zip, Folder “RFP”)
WORKFLOW
1. Open TMARKER, and drag and drop
the image “RFP.jpg” from the demonstration data into the program.
Task is to count the RFP negative and
positive cells. Cells can be detected by
their blueish nucleus.
Tutorial for Cell Classification (RFP positive/negative)
2. Label some “benign” nuclei (i.e. blue
nuclei with clear cytoplasm) – we abuse
the labeling “benign” and “malignant”
for “clear” and “stained” for now.
Label some “malignant” nuclei (i.e. blue
nuclei with stained cytoplasm)
TIP: Cover typical and also difficult cases
on which the algorithm can learn.
3. Set the nucleus radius to 11, which
best fits the nucleus radius.
11
4. Go to Tools -> Plugins -> Cancer Nucleus Classification.
TIP: If you don’t see any plugins, please
point to the online plugins in the
TMARKER options, and restart.
Set the patch size to 44 (“Auto Size”) and
blurring to 2.
Select the “Segment Nuclei” with
“Graphcut”
and
use
the
“Foregound/Background” Color
Explanation:
- The nuclei are patched with given
patch size, and blurred to reduce
single pixel noise.
- Then, the nuclei in the image
patches are segmented. The “Foreground” are nuclei.
- Since our RFP problem is a colorclassification problem, the Foreground/Background color might
work best for us.
Tutorial for Cell Classification (RFP positive/negative)
5. In the tab “Classifier”, select the Random Forest classifier with 50 trees (default).
In the tab “Process”, select “1-step classification” and click on “Train Classifier
on Labeled Nuclei”.
The classifier is now trained on your labeled nuclei.
6. Now, we want to detect all nuclei.
Go to Tools -> Plugins -> Color Deconvolution
Select following parameters:
Staining Protocol = H DAB.
Tolerance = 11
Blur = 1
T_hema = 146
T_dab = 256 (no nuclei on this channel)
Click on „Estimate“.
TIP: See the Color Deconvolution tutorial
for more information.
The nuclei are displayed in the main window as “Unknown” nuclei (not classified,
yet).
Tutorial for Cell Classification (RFP positive/negative)
7. Now, the detected nuclei are classified
with the trained classifier.
Go back to Tools -> Plugins -> Cancer Nucleus Classification
Click on “Classify Detected Nuclei”.
Now, all nuclei are patched, segmented,
blurred, the feature is extracted and
they are classified.
According to the features you have chosen, this might take a while. The progress is indicated on the bottom of the
main window.
When finished, TMARKER has classified
all nuclei according to your labeled examples.
Note: nuclei on the border of the image
will remain unknown (no feature extraction is possible).
8. (Optional)
When you find misclassified nuclei, you
might want to use the first “Edit” button (on top of TMARKER, beside the
“Background” label button).
Change the class of misclassified nuclei
by multiple clicks on the nucleus.
The nucleus will then be converted to a
“User Label” (i.e. training instance).
After a few changes, you might want to
re-train the classifier and re-classify all
nuclei and control the changes.
So, you can update the classifier until
you are satisfied.
Tutorial for Cell Classification (RFP positive/negative)
9. (Optional)
Find the counts in the TMA List (left side
of TMARKER).
Save the result as HTML report to discuss
it with colleagues.
Save the nuclei as XML file to continue/reproduce the analysis later.