The scan must be a gray intensity volume, dimension 3 ( HWL ).

Manual annotations must be supplied in an uint8 file of the same size, where each voxel value corresponds to the label class ( 0 as background).

Most MRI scans are isotropic on one plane only, due to the thickness of the slice. Manual annotations must be provided in the isotropic plane. Propagation is therefore performed in the 3rd dimension (to be indicated with z_axis).

Free-form scribbles/indications can also be supplied. This allows the user to annotate certain parts, without having to completely delineate a cut. In addition, hints can be provided in all planes, and outside the annotated top and bottom section, enabling propagation to be extrapolated. The hints file must be of the same type and size as the manual annotations file, with the same class/label correspondences. To specify a hint as part of the background class, voxels must have the label 255.

Pretrained weights can be downloaded here.

Let's consider the following scan scan.nii.gz and a corresponding segmentation file with 3 annotated slices manual_annotation.nii.gz, and some few freehand annotations in hints.nii.gz :

Training and propagation can be done for this single scan as follow :

import nibabel as ni
from labelprop.napari_entry import train_and_infer

scan=ni.load('scan.nii.gz').get_fdata() # Numpy array of dimension (H,W,L)
manual_annotations=ni.load('manual_annotations.nii.gz').get_fdata() # Numpy array of dimension (H,W,L) and dtype uint8
hints=ni.load('hints.nii.gz').get_fdata() # Numpy array of dimension (H,W,L) and dtype uint8
# Train and propagate 
propagations=train_and_infer(
    img=scan,
    mask=manual_annotations,
    pretrained_ckpt='pretrained.ckpt',
    shape=256, # Size of input images for training.
    max_epochs=100,
    z_axis=2, # Propagation axis.
    output_dir='path/to/savedir', 
    name='nameofcheckpoint',
    pretraining=False, # If True, will pretrain the model without using manual_annotations.
    hints=hints, # Optional hints for the propagation. Numpy array of dimension (H,W,L) and dtype uint8
)

propagation_up=propagations[0] # Propagation from the bottom to the top 
propagation_down=propagations[1] # Propagation from the top to the bottom

fused_propagated_annotations=propagations # Fusion of propagation_up and propagation_down. 
# Save results
ni.save(ni.Nifti1Image(fused_propagated_annotations,ni.load('img.nii.gz').affine),'propagated_fused.nii.gz')

Pretraining

Although Labelprop works on a single scan, it is preferable to pre-train the model on a dataset, with or without manual annotations.

Self-supervised

To pretrain the model without using any manual annotations :

$ labelprop pretrain --help
Usage: labelprop.py pretrain [OPTIONS] IMG_LIST

Pretrain the model on a list of images. The images are assumed to be
  greyscale nifti files. IMG_LIST is a text file containing line-separated
  paths to the images.

Options:
  -s, --shape INTEGER         Image size (default: 256)
  -z, --z_axis INTEGER        Axis along which to propagate (default: 2)
  -o, --output_dir DIRECTORY  Output directory for checkpoint
  -n, --name TEXT             Checkpoint name (default : datetime)
  -e, --max_epochs INTEGER    

In this case, the model simply learns to register successive sections with each other, without any anatomical constraints on propagation.

With annotations

Now, to train the model with sparse manual annotations :

$ labelprop train-dataset --help
Usage: labelprop train-dataset [OPTIONS] IMG_MASK_LIST

  Train the model on a full dataset. The images are assumed to be greyscale
  nifti files. Text file containing line-separated paths to greyscale images
  and comma separated associated mask paths

Options:
  -c FILE                     Path to the pretrained checkpoint (.ckpt)
  -s, --shape INTEGER         Image size (default: 256)
  -z, --z_axis INTEGER        Axis along which to propagate (default: 2)
  -o, --output_dir DIRECTORY  Output directory for checkpoint
  -n, --name TEXT             Checkpoint name (default : datetime)
  -e, --max_epochs INTEGER
  --help                      Show this message and exit.

Training

$ labelprop train --help
Usage: labelprop.py train [OPTIONS] IMG_PATH MASK_PATH

Train a model and save the checkpoint and predicted masks. IMG_PATH is a
  greyscale nifti (.nii.gz or .nii) image, while MASKPATH is it related sparse
  segmentation.

Options:
  -h, --hints FILE            Path to the hints image (.nii.gz)
  -s, --shape INTEGER         Image size (default: 256)
  -c, --pretrained_ckpt FILE  Path to the pretrained checkpoint (.ckpt)
  -e, --max_epochs INTEGER
  -z, --z_axis INTEGER        Axis along which to propagate (default: 2)
  -o, --output_dir DIRECTORY  Output directory for checkpoint and predicted
                              masks
  -n, --name TEXT             Prefix for the output files (checkpoint and
                              masks)

Propagating (inference)

$ labelprop propagate --help
Usage: labelprop.py propagate [OPTIONS] IMG_PATH MASK_PATH CHECKPOINT

Propagate labels from sparse segmentation.  IMG_PATH is a greyscale nifti
  (.nii.gz or .nii) image, while MASKPATH is it related sparse segmentation.
  CHECKPOINT is the path to the checkpoint (.ckpt) file.

Options:
  -h, --hints FILE            Path to the hints image (.nii.gz)
  -s, --shape INTEGER         Image size (default: 256)
  -z, --z_axis INTEGER        Axis along which to propagate (default: 2)
  -l, --label INTEGER         Label to propagate (default: 0 = all)
  -o, --output_dir DIRECTORY  Output directory for predicted masks (up, down
                              and fused)
  -n, --name TEXT             Prefix for the output files (masks)

See this repo to use labelprop main functions in Napari (cf. the GIF in the About section).

See also napari-labelprop-remote to run labelprop in a separate process, locally or remotely, which uses the API: