Challenge info:
The challenge deals with lymph node quantification (segmentation) but has a catch: Not all Lymph nodes are annotated within a patient. Out of e.g. 5 Lymphnodes (LN) only a single one is annotated.
to recreate the submitted model:
-
Download data
-
Preprocess data -calculate bodypart regression scores using the run_inference_on_bodypartregression.py (alternatively if only using the challenge data you can skip this stem and use the precomputed bodypartregression values provided) -Dataset preprocessing can be found in Dataset918_LNQ.py
-
convert the additional dataset using the convert_additional_dataset.py script
-
copy the converted file into the raw nnU-Net dataset (make sure to modify the dataset.json as well to account for the added cases)
-
Run the plan function of nnU-Net using the resenc planner (nUNetv2_plan_experiment -d -c 3d_fullres -pl ResEncUNetPlanner -overwrite_plans_name resenc_planner)
-
Modify the resulting plans file to incorporate: "batch_size": 4, "patch_size": [ 128, 192, 192 ]
-
Run nnU-Net preprocessing (nnUNetv2_preprocess -d -plans_name resenc_planner -c 3d_fullres)
-
Train nnU-Net using the trainer: nnUNetTrainerDA5
-
Inference can be run via the nnUNetv2_predict function
-
For containerisation see the code provided in Algorithm
-
your nnU-Net results folder needs to be copied into the results folder there (the model cant be provided here since the files are to big to be committed)
Installation help for the used tools can be found here: https://github.com/MIC-DKFZ/nnUNet https://github.com/MIC-DKFZ/BodyPartRegression https://github.com/wasserth/TotalSegmentator