B - Demo: The Cookbook Approach
We recommend learning Enhanced CellClassifier using the DEMO mode. We have prepared two biological examples; you can see the program running with full functionality. Please follow the Cookbook approach (below) to see the functionality of the program.
Testing the Demonstration
- Install full version of Matlab
- Download and unzip Enhanced CellClassifier
- Start the program Enhanced CellClassifier in Matlab
[Matlab: >File >Open, navigate to Enhanced CellClassifier.m, and open the program in the Matlab-Editor. Press "run program" (green arrow) to start]
- Select "Demo" in the start menu
- Choose one of the two prepared demonstrations
Note: To start the second demonstration, Enhanced CellClassifier needs to be closed and restarted.
Biological Example 1:
HGF-induced ruffling. After HeLa cells were treated with a hormone (HGF, hepatocyte growth factor) they start to "ruffle", as can be seen in the actin channel. RNAi-mediated knockdown of an important actin regulator (ACTR3), strongly reduces ruffling. Actin is shown in grey, nuclei in green. It is your task to make Enhanced CellClassifier able to recognize ruffling cells, non-ruffling cells and mitotic cells.
Biological example 2:
Docking of Salmonella onto HeLA cells. Salmonella (shown in green) can dock (bind) onto HeLa cells. For reasons currently unclear, Salmonella bind much stronger to mitotic cells than non-mitotic cells. Salmonella are shown in green, nuclei in grey. It is your task to enable Enhanced CellClassifier to distinguish mitotic and non-mitotic cells. With the loaded settings, Enhanced CellClassifier can ultimately differentially analyze Salmonella docking onto mitotic cells, neighbors of mitotic cells and other cells.
This cookbook is meant for first time users of Enhanced CellClassifier. It is not the recommended way to evaluate serious experiments; it is a fast way to see the program working and a demonstration of some functionality.
- Start Matlab and Enhanced CellClassifier as described above
choose Demo after starting
choose 1 of the 2 prepared biological examples
the Main window appears and you see the microscopy images
- Select the "Default" training mode.
In this training mode you can basically train whatever you want.
>>> Click the white box above the "HELP" button and choose "Default"
- Get familiar with the images.
Click on the radio buttons in the lower right of the window which say
for HGF-ruffling (example 1):
"ruffling" "non-ruffling" "ACTR3 HGF" "ACTR3 no HGF"
and for Salmonella docking (example 2):
"bacteria" "no bacteria" "A01" "A02"
You will get an image derived from the wells that fit this description.
Example 1: If you do not know how a ruffle looks like, compare the cells after clicking on "ruffling" or "non-ruffling", the difference is often very obvious, however, in borderline cases it is often a judgment call.
Example 1 + 2: Mitotic cells have bright and small nuclei, have a round, spiky, bright appearance in the actin channel (example 1), or are surrounded by green Salmonella (example 2).
- Select a class
>>> click on the class boxes in the lower right corner of the image. The names of the classes are:
for HGF-ruffling (example 1):
"Ruffling" "Non Ruffling" "Mitotic" ["Garbage"]
and for Salmonella docking (example 2):
"Mitotic" "Non Mitotic" ["Borderline", "Apoptotic", "Garbage" ]
"Deselect" is not a class, it is for removing (deselecting) trained object.
- Start Training
for training click directly onto a cell of the image. The object you click on will get the class you have just selected (point 4). The colored boxes (upper right of the window) show in the field "Current" the number you have trained on this image. Train some 2-10 cells for each class (if on the image).
If you do not know where the borders of the objects are, press
>>> Display >>> Show all cells >>> Show all cells as shadows
- Press "Save Data"
Now, the objects will be memorized. This is the only way to make Enhanced CellClassifier store the objects you just trained, if you select a new image by any other means, the "current" objects will be forgotten.
The colored boxes (upper right of the window) now show in the field "All" the number of all objects memorized during this session.
A new image is selected and you can keep training.
- Train some objects
keep training until about 100-200 objects are trained. Make sure, you have at least some objects from each class (some ruffling, some non-ruffling and some mitotic cells in example 1, mitotic and non-mitotic objects for example 2). You might need to switch wells (step 2).
- SVM Training
open SVM window, press >>> Model >>> SVM
a new window opens.
- Press TRAIN…
[ignore everything else in the window for now]
a window appears, showing the results of the training.
The trained objects have been divided into a "training" and a "validation" set.
The upper panels show the training data; these are the objects which have been used to train the model.
The lower three panels show the validation data. These data have not been used to train the model.
The left graphs show the input, how many objects of each class were present.
The middle graphs show % correct (green) and incorrect predictions (red).
The right graphs show the confusion matrix. The objects trained with class on x-scale are predicted to belong to the class on y-scale.
A good training accuracy is nice; however it is much more important to have good validation accuracy: The model should be able to predict objects it has never encountered before.
Enter a name for the model, after that you can use the model.
- Press RETURN and move back to the main window.
- Change Training mode to "Correction" (see step 2)
For biological example 2 ("docking of Salmonella") applying the predictions can take some time.
- Correct objects that have been incorrectly classified by the model.
Click on cells where you think the predictions are wrong. In the corrections mode the program only remembers the objects you correct.
- TRAIN a new model
After correcting some 50 - 100 objects do SVM-Training again (step 8)
Alternatively press the green TRAIN button (left part of the main window)
Repeat step 9-12 until you are happy with the classification (= until you are sure, the model generalizes well, but for now do not overdo… this is just a demonstration)
- >>> Output >>> AdjustOutput…
this is where the output is defined. Try to understand, how this works. To understand what we mean by "Vector", "Image variable", "Well variable" please read about basic concepts in the Enhanced CellClassifier manual (chapters D and and O).
- >>> Output >>> Generate ExcelFile…
In the white boxes the path is shown were the output data are saved.
press CALCULATE to generate the output or
press RETURN to return to the main window
- Congratulations, you just finished your first Enhanced CellClassifier session
- For enthusiasts:
To get most of the SVM-algorithm… a grid search of the parameters C and gamma is recommended (there is no need for now to understand what C and gamma means, they basically define the details of the mathematical procedure)
In the SVM-window press >>> Model >>> Parameter Optimization
Enter a starting and a final value for both parameters. Please note that every value is given on a log2 scale. The best parameters will be saved (manual chapter N for details).
This link is priceless… we follow exactly this strategy: http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf