Convert TotalSegmentator dataset into the nnUNet format

TotalSegmentator does not provide the scripts for preprocessing the dataset for nnUNet training (wasserth/TotalSegmentator#124). This repo provides a way to do that. Supports nnUNet v2 only.

---

Step 1 - Create the required environment

conda create -n ts_to_nnunet pandas jupyter tqdm simpleitk=2.0.2 -c simpleitk

SimpleITK v2.0.2 is required as explained here due to an ITK issue.

Step 2 - Download the TotalSegmentator dataset

Download the TotalSegmentator dataset from Zenodo.

Step 3 - Convert the TotalSegmentator into the nnUNet format

Use the examples.ipynb either as a demo on how to convert the data or to actually convert your own. Make sure to use the environment created in the Step 1.

Step 4 - Preprocess the dataset and specify the nnUNet's cross-validation folds

Run nnUNetv2_plan_and_preprocess -c 3d_fullres -d DATASET_ID --verify_dataset_integrity to preprocess the data. TotalSegmentator trains only 3d_fullres models, so here we specify that we need the data for it only via -c 3d_fullres.

Once its preprocessed, copy the splits_final.json into the dataset's preprocessed directory inside of your nnUNet_preprocessed dir, overwriting the generated splits_final.json. The copied splits_final.json follows TotalSegmentator's train/val split and specifies only fold 0.

Step 5 - Train only on fold 0

Train on the fold 0 only, just like the TotalSegmentator team did: nnUNetv2_train DATASET_ID 3d_fullres 0. Other folds do not exist.

Note that TotalSegmentator's models were trained using nnUNet v1.