search

neuronets / nobrainer

A framework for developing neural network models for 3D image processing.
Other
158 stars 45 forks source link

Multi-channel volumes #192

Open wanderine opened 2 years ago

wanderine commented 2 years ago

I am trying to synthesize multi-channel volumes, using 3D progressive GAN, where one or several channels contain the segmentation annotations. I tried this a bit with the old code (2020) but with "nobrainer convert" it does not seem to be possible to create multi-channel volumes? Because the help text says

--multi-resolution / --no-multi-resolution Create tfrecords for multiple resolutions. if set, labels in csv need to be scalar. [default: no-multi-resolution]

satra commented 2 years ago

@wanderine - general multichannel volumes was added by @richford - it's possible that the multiresolution support for multichannel wasn't added. for the multi-resolution case, i suspect the labels would still be scalar, but the inputs could have multiple associated volumes. if you send in a quick PR, happy to review and merge.

perhaps we can also consider both multi-resolution and multi-channel for both input and output. things like vox2vox or autoencoders may benefit from this. since we didn't have any model for this the multi-resolution was left as a mapping to a scalar.