Where the usual graphcut constructs a graph out of the image's voxels, the label graphcut used irregular regions instead. This reduces memory consumption, accelerates the execution and leads mostly to better results.


Original image	Foreground (red) and background (green) marker on original image

Let's assume we want to segment the ventricles in a brain scan (left image). First, we need some markers for foreground (red) and background (green). These can be seen in the right image and are usually either created manually or by another algorithm.

Creating the region image

The label graphcut requires a region image. MedPy supplies the required functionalities. First, we create a gradient magnitude version of our original image with



In [5]:

    
medpy_gradient.py resources/b0.nii.gz output/gradient.nii.gz

Which we then feed to the watershed algorith



In [6]:

    
medpy_watershed.py output/gradient.nii.gz output/watershed.nii.gz --mindist 10

Visually, the result process like this


Original image	Gradient image	Region image	Region image (colored)

Performing the graphcut

Now we just have to supply our images to the script



In [7]:

    
medpy_graphcut_label.py output/gradient.nii.gz output/watershed.nii.gz resources/b0markers.nii.gz output/graphcut_label.nii.gz --boundary=stawiaski

And, after some waiting time, we obtain

, which is a pretty good approximation of the ventricles visible in the brain scan.



In [ ]: