Skip to content

Latest commit

 

History

History
79 lines (62 loc) · 3.52 KB

README.md

File metadata and controls

79 lines (62 loc) · 3.52 KB
Raw

LiftReg: Limited Angle 2D/3D Deformable Registration

arXiv

This is the official repository for

LiftReg: Limited Angle 2D/3D Deformable Registration.
Lin Tian, Yueh Z. Lee, Ra'ul San Jos'e Est'epar, Marc Niethammer
MICCAI 2022.

In this work we propose LiftReg, a 2D/3D deformable registration approach. LiftReg is a deep registration framework which is trained using sets of digitally reconstructed radiographs (DRR) and computed tomography (CT) image pairs. By using simulated training data, LiftReg can use a high-quality CT-CT image similarity measure, which helps the network to learn a high-quality deformation space. To further improve registration quality and to address the inherent depth ambiguities of very limited angle acquisitions, we propose to use features extracted from the backprojected 2D images and a statistical deformation model. We test our approach on the DirLab lung registration dataset and show that it outperforms an existing learning-based pairwise registration approach. Model Structure

This repository is under active maintenance. Here is the todo list:

  • Improve the process to replicate the result
  • Upload subspace basis to cloud drive
  • Upload model weights to cloud drive

Setup environment

cd LiftReg
conda create -n liftreg python=3.7
pip install -r requirements.txt
pip install -e .

Generating DRRs

The process to generate the testing data is to run the following two commands:

python tools/preprocessing.py -o [PATH_TO_OUTPUT_FOLDER] -d [DATASET_NAME_OF_YOUR_CHOICE] --data_type val
python tools/preprocessingDRR.py --data_path [PATH_TO_OUTPUT_FOLDER]/[DATASET_NAME_OF_YOUR_CHOICE] --drr_folder_name [DRR_FOLDER_OF_YOUR_CHOICE] -g 0

Pre-built subspace and model weights

You can download the pre-built subspace and the weights for the registration network at Google Drive

Evaluate with the trained model

If the model is trained following the commands in the previous section, skip step 1. The paths are set correctly by main.py script and the setting.json file is saved at the OUTPUT_FOLDER/disp_subspace/DATE/ folder.

Step 1. Set the paths in cur_task_setting.json

To run the code with pretrained model, set the following paths in cur_task_setting.json:

  1. ["dataset"]["data_path"]
  2. ["dataset"]["drr_folder_name"]
  3. ["train"]["model"]["pca_path"]
  4. ["train"]["output_path"]
  5. ["train"]["test_from"] - Path to the network weights.

Step 2. Run evaluation script

Run the following command:

cd LiftReg
python eval.py -s ./exps/disp_subspace/2022_02_17_20_12_57/cur_task_setting.json -g GPU_ID

Train with built subspace

Assume the two files (pca_vectors.npy and pca_mean.npy) are saved in PCA_PATH and the path to the dataset folder is DATA_PATH.

  1. Set the "pca_path" in cur_task_setting.json to PCA_PATH
  2. Run the following commands
cd LiftReg
python main.py -o OUTPUT_FOLDER -d DATA_PATH -e disp_subspace -s ./ -g GPU_ID

Citation

@inproceedings{tian2022liftreg,
  title={LiftReg: Limited Angle 2D/3D Deformable Registration},
  author={Tian, Lin and Lee, Yueh Z and San Jos{\'e} Est{\'e}par, Ra{\'u}l and Niethammer, Marc},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={207--216},
  year={2022},
  organization={Springer}
}