suppl/5_splits.md

Splits

After labeling, the dataset was divided into six parts: one for testing and five for use as folds in a cross-validation training scenarios. The split was stratified based on

• Primary Stratification: Data source (TotalSegmentator vs. VerSe) to ensure that each split reflects the original datasets.
• Secondary Stratification: Vertebrae present in each scan, to maintain a balanced distribution of anatomical structures across the splits.

Additionally, CT scans that were originally assigned to the test sets (73 from TotalSegmentator and 105 from VerSe) were preferentially included in the test set of this dataset. To address the rarity of certain labels, we prioritized an equal distribution of these rare labels over the more common vertebrae.

For detailed reference, we provide the following files:

• data.csv: This file provides an overview of the vertebrae present in each case, the original assignment sets (set_original, either training or testing), the newly assigned set (set, either training or testing), and the specific fold (fold, -1 for testing and 0-4 for individual training folds) to which each case was assigned.
• splits_train-test.json: This file contains the split into training and test sets.
• splits_folds.json: This file details the split into training and cross-validation sets. It is equivalent to the splits-final.json file used in nnU-Net.