napari Label Propagation


Label propagation through deep registration

License PyPI Python Version tests codecov napari hub

3D semi-automatic segmentation using deep registration-based 2D label propagation

This napari plugin was generated with Cookiecutter using @napari's cookiecutter-napari-plugin template.


See "Semi-automatic muscle segmentation in MR images using deep registration-based label propagation" paper :

[Paper]Paper [PDF]PDF


To install this project :

pip install napari['all']
pip install git+


Download pretrained weights.

Open napari from terminal and start using functions from 'napari-labelprop' plugin (Under Plugins scrolling menu).

Available functions are :

  • Inference : Propagate labels from trained weights (Pytorch checkpoint required)
  • Training : Start training from scratch or from the pretrained weights.

PS : "Unsupervised pretraining" is not yet implemented. See CLI option at LabelProp repository.

Every operation is done in the main thread. So, napari is not responsive during training or inference, but you can still follow the progress in the terminal.


To train a model, reach the plugin in the menu bar :

Plugins > napari-labelprop > Training

Fill the fields with the following information :

  • Image : Select a loaded napari.layers.Image layer to segment
  • Labels : Select a loaded napari.layers.Labels layer with the initial labels
  • hints : Select a loaded napari.layers.Labels layer with scribbled pseudo labels
  • Pretrained checkpoint : Select a pretrained checkpoint from the server-side checkpoint directory
  • Slices shape : Slices are resample to this shape for training and inference, then resampled to original shape. So far, slices must be squares.
  • Propagation axis : Set the axis to use for the propagation dimension
  • Max epochs : Set the maximum number of epochs to train the model
  • Checkpoint output directory
  • Checkpoint name
  • Weighting criteria : Defines the criteria used to weight each direction of propagation ncc = normalized cross correlation (slow but smooth), distance = distance to the nearest label (fast but less accurate)
  • Reduction : When using ncc, defines the reduction to apply to the ncc map mean / local_mean / none. Default is none
  • Use GPU : Set if whether to use the GPU or not. Default is True (GPU). GPU:0 is used by default. To use another GPU, set the CUDA_VISIBLE_DEVICES environment variable before launching napari.

To run inference on a model, reach the plugin in the menu bar :

Plugins > napari-labelprop-remote > Inference

Fill the fields like in the training section. Then, click on the Run button.


Contributions are very welcome. Tests can be run with tox, please ensure the coverage at least stays the same before you submit a pull request.


Distributed under the terms of the BSD-3 license, "napari-labelprop" is free and open source software


If you encounter any problems, please [file an issue] along with a detailed description.


  • 1.0.0

Last updated:

  • 21 June 2023

First released:

  • 19 June 2023


Supported data:

  • Information not submitted

Plugin type:

GitHub activity:

  • Stars: 0
  • Forks: 0
  • Issues + PRs: 0

Python versions supported:

Operating system:


  • deep-labelprop
  • napari-nifti
  • numpy
  • magicgui
  • qtpy