Data Preparation
Microscopy Image Browser is the recommended tool for create and viewing data.
Fiji can be useful for doing post-segmentation processing and analysis.
8bit or 16bit greyscale 3D tif images, preferably normalized (contrast enhanced).
Can have isotropic or anisotropic resolution. But xy dimension should have the same resolution.
There is no upper size limit on the training image, as the tool work by feeding the network with small randomly cropped patches of each training image.
Example (only showing the top most slice):
Left: Semantic segmentation. Right: Instance segmentation.
8bit or 16bit greyscale 3D tif image, have the same spatial size as the corresponding image file. The file name should start with "Labels_".
Value of 0 is background while value of 1 is foreground.
If you are using Microscopy Image Browser, then the tif file generated by "save model as..." option should be just the thing you need.
Example (only showing the top most slice):
Same as above. But value of 0 is unlabeled, value of 1 is foreground, and value of 2 is background.
Example (only showing the top most slice):
Same as Fully Labelled Semantic Segmentation. But except that each value represents a different foreground object. Value of 0 is still background.
Example (only showing the top most slice):
In default, the tool automatically uses the data from the "Dataset" folder in the repository root directory. Although you can set it to somewhere else.
Place training images and they labels to "train" folder, validation images and they labels to "val" folder. Place the images that you want the network to predict to "predict" folder. You will find your result at "result" folder.