NiftyNet cannot find multiple labels for multiclass segmentation

Question

I'm running a 3-class segmentation problem, with my Label images being "0(many),1(some),2(some)". But I realized when NiftyNet read the images, it only has 0/1 two values. When the label_normalization is on, the ref.txt file says "label from 0,1 to 0,1"; if I turn it off, the prediction also has only 0/1 two values.

Any idea what has gone wrong? Label image format problem?

Many thanks!

Answer 1

Can I ask what the num_classes is in the config file that you are using? When NiftyNet generates the label_normalization, it goes through all the files that it finds. How are you determining the list of files? If it's by searching a path, are you sure the path has all the relevant files on it?