>> Image Navigation >> Alternative Training
Note of caution: This module is still very recent, bugs might still occur.
When making predictions, the SVM algorithm will predict for each object and each class probabilities. The object will be assign to the class with the highest probability. To improve a model, it might be useful to focus only on those objects closest to the decision boundaries of the current module. These objects will have borderline phenotypes. In "Alternative Training", training focuses on exactly these objects. This might help better defining the decision boundaries between classes, thereby improving the model.
In the Enhanced CellClassifier "Main Window", training is done image based. In contrast, in "Alternative Training", 10 objects from a larger sample of images are presented, providing a mixed sample of objects. On the other hand, loading of image data takes more time and might make training slower.
"Alternative Training" uses the currently selected well group from Enhanced CellClassifier and the class currently selected for training. Both can be changed in the "Alternative Training" window.
After starting "Alternative Training" from the Enhanced CellClassifier "Main window", one is
Asked, how many percent of the image data should be loaded. A higher percentage provides a larger set of data from which objects can be selected. On the other hand, loading can be time consuming. Since Matlab has difficulties handling very large variables, selecting a very high number of images takes excessively long, almost freezing the program. Stop the loading and chose a lower percentage value
10 different objects are displayed simultaneously. Each object might originate from a different image.
The 5 objects above are predicted to be the class currently selected. Objects 1+2 belong to the 0-10% objects closest to the decision boundaries.
Objects 3+4 belong to the 5-15% objects closest to the decision boundaries
In other words: Objects 1-4 belong to the current class but have almost been assigned to a different class (difficult decision).
Object 5 belongs to the best 10% of object, far away from the decision boundaries (positive control, certainly belongs to the current class)
The 5 lower objects are predicted not to belong to the class currently selected.
Object 6 belongs to the lowest 10% of objects, far away from the decision boundaries (negative control, almost certain belongs not to this class).
Objects 7+8 are from the 5-15% of objects closest to the decision boundaries,
Objects 9+10 are from the 10% of object closest to the decision.
In other words: Objects 7-10 do not belong to the current class, but almost had been assigned to the current class (difficult decision).
Note: If not enough objects are available (for instance if no object exists which is predicted to belong to the current class), at first the boundaries (0-10% and 5-15%, respectively), will be loosened to 0-50%. In a next step any objects independent of their predicted class will be presented. If even under these loosened conditions no objects can be found, Enhanced CellClassifier complains and suggests loading new objects or selection of a different class.
Note: The color of the panels behind the objects codes for the selected class of each object.
The current objects with labels and data are saved, new objects are displayed
New objects are selected displayed, the current objects are lost.
Shortcut to SVM. A new model will be trained with the settings adjusted while opening the SVM-window the last time.
Last Save Data will be reversed. There is no multiple UNDO option and no REDO option.
Return to the Enhanced CellClassifier main window, keeping or ignoring the objects and settings changes which have been made while using "Alternative Training".
Inverts the current labels. Class 1 will be converted to class 2, class 2 to class 1.
A new well group will be selected (or the current well group again). The data will be (re)loaded, exactly as when starting "Alternative Training".
The class will be changed. Since selection of objects always happens in relation to the current class (see display of objects), new objects will be loaded.
In the lower left corner the number of the objects still available is displayed with their respective class. Note that the objects currently displayed are also included as available
In the lower left corner, the name of the current model is displayed
>> File >> Save object data
The cells currently trained will be saved, as described in Enhanced CellClassifier main window.
>> More >> Plot Histogram
From the assembled data histograms will be generated, as described in Enhanced CellClassifier main window.
>> More >> SVM
Opens the SVM window.
>> More >> Predictions >> Default
>> More >> Predictions >> Model
This option decides on the labels of the objects after loading.
With "Model" selected, the label will be the prediction made by the current model.
With "Default" selected, the label will be the current class for the upper 5 objects and class 2 (negative) for the lower 5. If the current class is 2, the labels of the lower objects will be class 1.
The "Default" mode would prevent a tendency by the user to confirm the current model.
>> More >> Focus exclusion >> Focus exclusion ON
>> More >> Focus exclusion >> Focus exclusion OFF
If focus exclusion is set to "ON", out of focus images, as calculated by the formula adjusted in the FOCUS window will be excluded. Check for explanations there.
>> Display >> Mark objects as shadows
>> Display >> Mark objects with rectangles
>> Display >> Mark objects with outlines
>> Display >> Mark objects OFF
The outlines of the objects might help with training; on the other hand they might hide important information. CC offers 4 ways to mark the objects, as blue shadows (default), as white outlines, as white rectangles or not at all.
Diese Website wird in älteren Versionen von Netscape ohne graphische Elemente dargestellt. Die Funktionalität der Website ist aber trotzdem gewährleistet. Wenn Sie diese Website regelmässig benutzen, empfehlen wir Ihnen, auf Ihrem Computer einen aktuellen Browser zu installieren. Weitere Informationen finden Sie auf
The content in this site is accessible to any browser or Internet device, however, some graphics will display correctly only in the newer versions of Netscape. To get the most out of our site we suggest you upgrade to a newer browser.