SeqSeg, From New Dataset to Training to Inference

A tutorial on setting up SeqSeg for medical image segmentation

1. Data Preparation and Structure

The first step in using SeqSeg is to preprocess your data.

We require the following data:

• A directory containing the images named images
• A directory containing the masks named truths
• A directory containing the centerlines named centerlines
  • as .vtp files

A few things to note:

• The images and masks and centerlines should have the same name
• If you need to extract centerlines from masks, you can use the SeqSeg/centerlines.py script, or use VMTK or other tools
  • Centerlines must contain radius information in the .vtp file
• Make sure that the masks, images and centerlines align correctly, for example open and view together in an image viewer e.g. Paraview
• Make sure the images contain origin, spacing and direction information in the metadata
  • This is important for correct alignment before centerline extraction

2. Data Preprocessing (for Training)

The next step is to preprocess the data for training. SeqSeg requires a model trained on local patches, so we need to extract patches from the images and masks based on centerlines.

The repository for this is BloodVesselML3D and the script is gather_sampling_data_parallel.py. This requires the following arguments:

• config - the configuration file for the dataset, which you must change to match your data

Next, we must change the naming structure to match nnU-Net. This is done with the dataset_dirs/create_nnunet.py script. The new data can be output anywhere, but we recommend directly into the nnU-Net Raw directory.

3. Training

The next step is to train the model. This is done with the specific nnU-Net commands, which are detailed in the nnU-Net repository.

This requires two commands (see documentation for more details):

• prepocessing command
• train command

You can train:

• 2D models
• 3D low resolution models
• 3D full resolution models

4. Inference

The final step is to run SeqSeg inference on new data. This is done with the SeqSeg/seqseg.py script. You need direct access to the directory containing the images and seed points, and another containing the trained model weights.