search

wasserth / TotalSegmentator

Tool for robust segmentation of >100 important anatomical structures in CT and MR images
Apache License 2.0
1.43k stars 239 forks source link

lung_vessels task - model details #116

Closed kretes closed 1 year ago

kretes commented 1 year ago

Hello,

first of all thanks for the effort on making such an awesome model & package!

I have a question about a subtask lung_vessels - how was the model trained for this task?

In readme you write lung_vessels: lung_vessels (cite [paper](https://www.sciencedirect.com/science/article/pii/S0720048X22001097)), lung_trachea_bronchia but -

  1. the cited paper for lung vessels describes a non-trainable approach to vessel segmentation, and it comes with its own implementation at https://github.com/fsc-mib/travel. I understand what you provide in TotalSegmentator is a model trained somehow on dataset acquired using cited approach?
  2. no reference model for lung_trachea_bronchia

thanks!

wasserth commented 1 year ago

We used the approach described in the paper to create lung vessels and bronchia segmentations on the totalsegmentator dataset. Then we manually checked where those segmentations worked well (it did not work well for all cases). Then we trained a nnunet on those cases (rougly 250 cases). This is the lung_vessels model.

kretes commented 1 year ago

Thanks very much for the clarification!