Dose Guidance for Radiotherapy-oriented Deep Learning Segmentation (DOSELO)

This repository contains the code for the paper "Dose Guidance for Radiotherapy-oriented Deep Learning Segmentation (DOSELO)" which is accepted for the MICCAI 2023 in Vancouver.

Paper Abstract

Deep learning-based image segmentation for radiotherapy is intended to speed up the planning process and yield consistent results. However, most of these segmentation methods solely rely on distribution and geometry-associated training objectives without considering tumor control and the sparing of healthy tissues. To incorporate dosimetric effects into segmentation models, we propose a new training loss function that extends current state-of-the-art segmentation model training via a dose-based guidance method. We hypothesized that adding sucha dose-guidance mechanism improves the robustness of the segmentation with respect to the dose (i.e., resolves distant outliers and focuses on locations of high dose/dose gradient). We demonstrate the effectiveness of the proposed method on Gross Tumor Volume segmentation for glioblastoma treatment. The obtained dosimetry-based results show reduced dose errors relative to the ground truth dose map using the proposed dosimetry-segmentation guidance, outperforming state-of-the-art distribution and geometry-based segmentation losses.

Repository Structure

This repository contains the following functionality for reproducing our results:

• Training procedure for the dose predictor training (see main_train_dose.py)
• Training procedure for the segmentation model training (see main_train_segment.py)
• Evaluation procedure for the dose predictor and the segmentation model (see main_test.py)
• Script for dataset construction (see doseseg/construction/dataset_construction.py)

All provided code is located in the ./doseseg/ folder and all persistent and temporary data is found in the ./data/ folder. The datasets (i.e., one for training the dose predictor and one for training the segmentation model) are not included in this repository because no ethical approval for sharing the data was obtained. However, we share with you all non-confidential data and code that is necessary to reproduce our results on your own data.

Training Workflow

Using the proposed method with your own data requires the following steps:

1. Acquire a dataset comprising for each patient:
   • A CT image
   • All necessary MR images (optional)
   • All OAR segmentations
   • A target segmentation
   • A simulated dose map
2. Convert the clinical data to a discretized file format (e.g., NIfTI) and co-register all images. The images and the segmentations should have the same physical properties (i.e., size, spacing, direction, orientation). Converting and harmonizing the data is best achieved using our PyRaDiSe package.
3. Construct a dataset for training the dose predictor and another one for training the segmentation networks (see doseseg/construction/dataset_construction.py)
4. Train the dose predictor (see main_train_dose.py) using the dose predictor dataset
5. Train the segmentation model (see main_train_segment.py) using the segmentation dataset and the previously trained dose predictor
6. Evaluate the trained dose predictor and segmentation model (see main_test.py)
7. (Optional) Use the trained dose predictor and segmentation model for dose prediction and segmentation of new patients
8. Congratulations, you have successfully trained a dose aware segmentation model on your own data :)

Citation

If you use our code for your own work, please cite our paper:

@InProceedings{Ruefenacht2023a,
  author    = {Rüfenacht, Elias and Poel, Robert and Kamath, Amith and Ermis, Ekin and Scheib, Stefan and Fix, Michael K. and Reyes, Mauricio},
  booktitle = {Medical Image Computing and Computer Assisted Intervention – MICCAI 2023},
  title     = {{Dose Guidance for Radiotherapy-oriented Deep Learning Segmentation}},
  year      = {2023},
  publisher = {Springer},
}