maarzt
commented
10 months ago Ok, interesting. In Labkit I call a collection of ROIs a "labeling". So if I understand you correctly you would like to start Labkit with a pre-initialized "labeling". That is definitely possible. Here is an example for you:
import java.awt.event.WindowAdapter;
import java.awt.event.WindowEvent;
import java.util.Arrays;
import javax.swing.JFrame;
import javax.swing.WindowConstants;
import ij.ImagePlus;
import net.imagej.ImgPlus;
import net.imglib2.FinalInterval;
import net.imglib2.RandomAccess;
import net.imglib2.img.VirtualStackAdapter;
import net.imglib2.img.display.imagej.ImageJFunctions;
import net.imglib2.roi.labeling.LabelingType;
import net.imglib2.type.numeric.integer.UnsignedByteType;
import net.imglib2.type.numeric.real.FloatType;
import org.scijava.Context;
import sc.fiji.labkit.ui.SegmentationComponent;
import sc.fiji.labkit.ui.inputimage.DatasetInputImage;
import sc.fiji.labkit.ui.labeling.Label;
import sc.fiji.labkit.ui.labeling.Labeling;
import sc.fiji.labkit.ui.models.DefaultSegmentationModel;
import sc.fiji.labkit.ui.segmentation.SegmentationTool;
import sc.fiji.labkit.ui.segmentation.Segmenter;
/**
* An example how one can start a "custom" Labkit with pre-initialized labeling.
*/
public class LabkitExample {
public static void main(String... args) {
// open an image and create a segmentation model
Context context = new Context();
ImagePlus imagePlus = new ImagePlus("https://imagej.nih.gov/ij/images/t1-head.zip");
ImgPlus<?> image = VirtualStackAdapter.wrap(imagePlus);
image.setChannelMinimum(0, 0 ); // set brightness and contrast for displaying the given image
image.setChannelMaximum(0, 1000 );
DefaultSegmentationModel segmentationModel =
new DefaultSegmentationModel(context, new DatasetInputImage(image));
// Initialize an empty labeling of correct size
FinalInterval imageSize = new FinalInterval(image);
Labeling labeling = Labeling.createEmpty(Arrays.asList("S1", "S2", "background"), imageSize);
// add one straight line
RandomAccess<LabelingType<Label>> randomAccess = labeling.randomAccess();
LabelingType<Label> color = randomAccess.get().createVariable();
color.add(labeling.getLabel("S1"));
for (int i = 0; i < 200; i++) {
randomAccess.setPosition(i, 0); // set X
randomAccess.setPosition(i, 1); // set Y
randomAccess.setPosition(i, 2); // set Z
randomAccess.get().set( color );
}
// Create the gui components
segmentationModel.imageLabelingModel().labeling().set(labeling);
JFrame frame = new JFrame();
SegmentationComponent segmentationComponent =
new SegmentationComponent(frame, segmentationModel, false);
frame.add(segmentationComponent);
frame.setSize(1000, 600);
frame.setDefaultCloseOperation( WindowConstants.DISPOSE_ON_CLOSE );
frame.addWindowListener(new WindowAdapter() {
@Override
public void windowClosed(WindowEvent e) {
showResults( segmentationModel, image );
segmentationComponent.close(); // don't forget to close the segmentation component if you are done.
}
});
frame.setVisible(true);
}
private static void showResults(DefaultSegmentationModel segmentationModel, ImgPlus<?> image) {
// Compute and show the complete segmentation and probability map.
Segmenter segmenter = segmentationModel.segmenterList().segmenters().get().get(0);
SegmentationTool segmentationTool = new SegmentationTool(segmenter);
ImgPlus<UnsignedByteType> segmentation = segmentationTool.segment(image);
ImageJFunctions.show(segmentation).setDisplayRange(0, 3);
ImgPlus<FloatType> probabilityMap = segmentationTool.probabilityMap(image);
ImageJFunctions.show(probabilityMap).setDisplayRange(0, 1);
}
}The example start's Labkit on a given image and also shows a labeling that is pre-initialized with a straight line.
Additionally if you train a pixel classifier and close the window. The results will be shown in two new windows.
@tferr I hope this example helps you!
tferr
@maarzt ,
We've been hacking ways to integrate both Labkit and TWS into SNT. Right now we can import and train models from paths, but a more useful command would to to "Send neurites" to Labkit, so that the traced paths of selected neurites could be pre-loaded into a Labkit instance as labels.
This would allow for pixel-perfect labels along thin neurites, which - in our experience - significantly improves the training relatively to freehand annotations.
In TWS, we convert Paths into ROIs and feed those to its GUI at startup. It is a bit clunky because we need to use reflection, but it works quite well (will probably open a PR there to formalize this feature). Labkit can hold much larger datasets and would be a pity to be restricted to TWS.
Do you have pointers on how to attempt this in Labkit? I've looked at
LabelBrushControllerbriefly, but did not find it useful. Did I miss something? I am convinced that a method that would allow to 'paint' programmatically would work. A method that converts poly/freehand-lines into labels would be best.NB: We can convert selected paths into a digitized skeleton mask and train a model with it, but that is not interactive, and defeats a bit what we are trying to achieve.