U-Net Benchmark

Revisiting model scaling with a U-net benchmark for 3D medical image segmentation
Scientific Reports (2025)

Paper: https://www.nature.com/articles/s41598-025-15617-1

Overview

The script add_sdw_variants_to_nnunetv2_plans.py adds 18 custom U-Net configurations to nnU-Net v2 plans files. These configurations systematically vary three structural hyperparameters:

• S ∈ {4, 5, 6} → number of resolution stages
• D ∈ {2, 3} → convolution layers per stage
• W ∈ {16, 32, 64} → initial channel width

This yields 18 variants (3 × 2 × 3) named as: 3d_fullres_S{S}D{D}W{W}

Environment Setup

This script is based on nnU-Net v2.4.1 (release link), which is the latest version as of August 2025.

Note: For other nnU-Net versions (e.g., nnU-Net v1, v2.2), the nnUNetPlans.json structure may differ slightly, requiring minor code adjustments.

Ensure the nnUNet_preprocessed environment variable is properly configured before running the script.

Usage

1. Plan and preprocess your dataset (standard nnU-Net workflow):

   nnUNetv2_plan_and_preprocess -d DATASET_ID

2. Add S/D/W variants to the plans:

   python add_sdw_variants_to_nnunetv2_plans.py -d DATASET_ID

3. Train models with the new configurations:

   nnUNetv2_train DATASET_ID 3d_fullres_S{S}D{D}W{W} FOLD
   # e.g., nnUNetv2_train 5 3d_fullres_S4D2W16 0

Citation

@article{huang2025unetbenchmark,
  title   = {Revisiting model scaling with a U-net benchmark for 3D medical image segmentation},
  author  = {Huang, Ziyan and Ye, Jin and Wang, Haoyu and Deng, Zhongying and Yang, Zhikai and Su, Yanzhou and Liu, Jie and Li, Tianbin and Gu, Yun and Zhang, Shaoting and Qiao, Yu and Gu, Lixu and He, Junjun},
  journal = {Scientific Reports},
  volume  = {15},
  number  = {29795},
  year    = {2025},
  doi     = {10.1038/s41598-025-15617-1}
}

Acknowledgements

We are grateful to:

• The nnU-Net team for the open-source codebase this project builds upon.
• The contributors of the 42 public datasets that enable open and reproducible research.
• Members of the GMAI team who contributed to this project.