DRL-STNet: Unsupervised Domain Adaptation for Cross-modality Medical Image Segmentation via Disentangled Representation Learning

Introduction

This algorithm is for Task 3 in FLARE Challenge, which was held at MICCAI 2024. We are ranked ${\textsf{\color{red}5th}}$ regarding the accuracy!

If you find our work is useful in your research, please consider citing:

(1) DRL-STNet: Unsupervised Domain Adaptation for Cross-modality Medical Image Segmentation via Disentangled Representation Learning

@article{lin2024drl,
  title={DRL-STNet: Unsupervised Domain Adaptation for Cross-modality Medical Image Segmentation via Disentangled Representation Learning},
  author={Lin, Hui and Schiffers, Florian and L{\'o}pez-Tapia, Santiago and Tavakoli, Neda and Kim, Daniel and Katsaggelos, Aggelos K},
  journal={arXiv preprint arXiv:2409.18340},
  year={2024}
}

Dataset

To obtain the FLARE Challenge dataset

Shared weights

Folder Structures

• The input for stage 0:

   FLARE/raw_data/  
         ├── source/
         │   ├── FLARE22_Tr_0001_0000.nii.gz
         │   ├── FLARE22_Tr_0002_0000.nii.gz
         │   ├── ...
         ├── source_labels/
         │   ├── FLARE22_Tr_0001.nii.gz
         │   ├── FLARE22_Tr_0002.nii.gz
         │   ├── ...
         ├── target/
         │   ├── amos_XXXX_0000.nii.gz
         │   ├── amos_XXXX_0000.nii.gz
         │   ├── ...
  

• After stage 0 and stage 1

   FLARE/Translation/  
         ├── translation_exp/
         │   ├── translation_checkpoints
         │   ├── translation_train_visual
         ├── translation_data/
         │   ├── train/
         │   │   ├── source_imgs.npy
         │   │   │── target_imgs.npy
         │   │   │── source_selected_files.txt
         │   │   │── target_selected_files.txt
         │   ├── val/
         │   │   ├── source_imgs.npy
         │   │   │── target_imgs.npy
         │   │   │── source_selected_files.txt
         │   │   │── target_selected_files.txt
  

• After stage 2

  FLARE/Translation/  
        ├── prediction/
        │   ├── FLARE22_Tr_0001_0000.nii.gz
        │   ├── FLARE22_Tr_0002_0000.nii.gz
        │   ├── ...
        ├── translation_exp/
        ├── translation_data/
  

Instructions

stage 0 Prepare data for training the translation model

Create your experiment_name folder (e.g., FLARE/), and place the raw_data inside it.

python "/home/hln0895/DRL-STNet/stage0.py" --config "path/to/your/config.yaml"

stage 1: train the translation model

python "/home/hln0895/DRL-STNet/stage1.py" --config "path/to/your/config.yaml"

stage 2: generate fake target scans (MRI)

python "/home/hln0895/DRL-STNet/stage2.py" --config "path/to/your/config.yaml"

stage 3: train a seg model using fake target scans (MRI)

python "/home/hln0895/DRL-STNet/stage3.py" --config "path/to/your/config.yaml"

Contact Us

Feel free to contact me at huilin2023@u.northwestern.edu