Pediatric Segmentation Approaches via Adult Augmentations and Transfer Learning

Table of Contents

• Overview
• Features
• Citation
• Checkpoints & Pretrained Models
• Quickstart
• Usage
• Dependencies
• Documentation
• Running Tests
• Contributing
• License

Overview

PSAT addresses pediatric segmentation challenges by leveraging adult, pediatric, and mixed datasets, advanced augmentation strategies, and transfer learning. It is designed for researchers and practitioners working on medical image segmentation, especially in pediatric contexts.

Features

• Flexible Training Plans: Use adult, pediatric, or mixed data ($Pa$, $Pp$, $P_m$)
• Customizable Learning Sets: Adult-only, pediatric-only, or mixed ($Sa$, $Sp$, $S_m$)
• Augmentation Strategies: Default ($Ad$) and contraction-based ($Ac$)
• Transfer Learning: Direct inference ($To$), fine-tuning ($Tp$), continual learning ($T_m$)
• Pretrained Models: Ready-to-use checkpoints for nnU-Net
• Evaluation Scripts: For fast metrics computation

Citation

If you use this code, please cite our paper:

@inproceedings{kirscher:hal-05137403, TITLE = {{PSAT: Pediatric Segmentation Approaches via Adult Augmentations and Transfer Learning}}, AUTHOR = {Kirscher, Tristan and Faisan, Sylvain and Coubez, Xavier and Barrier, Loris and Meyer, Philippe}, URL = {https://hal.science/hal-05137403}, BOOKTITLE = {{28th INTERNATIONAL CONFERENCE ON MEDICAL IMAGE COMPUTING AND COMPUTER ASSISTED INTERVENTION}}, ADDRESS = {Daejon, South Korea}, YEAR = {2025}, MONTH = Sep, KEYWORDS = {Pediatric Segmentation Age Bias Domain Adaptation Transfer Learning ; Pediatric Segmentation ; Age Bias ; Domain Adaptation ; Transfer Learning}, PDF = {https://hal.science/hal-05137403v1/file/main.pdf}, HAL_ID = {hal-05137403}, HAL_VERSION = {v1}, }

This repository includes a CITATION.cff file for standardized citation metadata. You can also use the "Cite this repository" button on GitHub to obtain citation formats automatically.

Checkpoints & Pretrained Models

We provide two model checkpoints for nnU-Net: - mixedmodelcontinual_learning.zip - purepediatricmodel.zip

Installing Pretrained Models

1. Download Pretrained Weights:

   • Go to the GitHub Releases page.
   • Download mixed_model_continual_learning.zip and pure_pediatric_model.zip.
   • Place them in resources/checkpoints/.

2. Install the Checkpoint Using nnU-Net: bash nnUNetv2_install_pretrained_model_from_zip resources/checkpoints/mixed_model_continual_learning.zip nnUNetv2_install_pretrained_model_from_zip resources/checkpoints/pure_pediatric_model.zip

3. Run Inference: After installing a checkpoint, run inference on your images: bash nnUNetv2_predict -i <input_images_dir> -o <output_dir> -d <dataset_id> -c <trainer_name> -f 0 -tr <task_name> Replace <input_images_dir>, <output_dir>, <dataset_id>, <trainer_name>, and <task_name> as appropriate. See nnUNet documentation for details.

For more details, see the Resources section.

Quickstart

1. Install dependencies: bash pip install -r requirements.txt

2. Evaluate Metrics (Example): bash python scripts/compute_metrics.py <ground_truth_dir> <predictions_dir> Replace <ground_truth_dir> and <predictions_dir> with your folder paths containing NIfTI files.

Usage

• Preprocessing, Training, and Inference:
  • See the nnUNet documentation and nnUNet/nnUNet.md for details on running full pipelines.
• Scripts:
  • Utility scripts are in the scripts/ directory. See scripts/scripts.md for usage.

Dependencies

• nibabel
• numpy
• pandas
• p_tqdm
• scipy
• surface-distance

(See requirements.txt for the full list.)

Documentation

• nnUNet
• Dataloading
• Preprocessing
• Training
• Resources
• Scripts

Running Tests

Install dependencies listed in requirements.txt and run:

bash pytest -q

Contributing

Contributions are welcome! Please open issues or pull requests for bug fixes, improvements, or new features.

License

This project is licensed under the MIT License. See LICENSE for details.