label_propagation

Label propagation

Automatic segmentation of organs is performed by propagating labels from the atlas to to either the input moving images or the rigidly registered images.

There are two methods for doing this in LAMA:

1. Inverting the forward transform

After registering moving images towards the population avergae, LAMA can create an inverse transfrom using the method descibed in section 6.1.6 of the [elastix manual[|https://github.com/SuperElastix/elastix/releases/download/5.0.0/elastix-5.0.0-manual.pdf]]. Currently the DisplacementMagnitudePenalty that is used in the inversion in not multi-threaded and so using this step can be a bottleneck in the pipeline.

To use this method, add this line in the lama config. label_propagation = 'invert_transform'

2. Reverse registration

Instead of creating an inverse transform, a second registration is performed where fixed and moving are switched, so that the poulation average is registrered towards the input specimen image. The transform file generatd during this registration is then appplied to the atlas. This method is able to use multi-threaded metric s in elastix and so is much quicker than method one. We have also found that this method produces slighly better segmentation than method 1 with our data.

To use this method, add this line in the lama config. This is the default method. label_propagation = 'reverse_registration'