• The code of SdAOF training & test are released (in ./scripts).
• The pipeline of preprocessing Totalsegmentor-hip to Totalsegmentor_Pelvis_Bone_Recon_Dataset to compatable for efficient training is released (in ./Data_Preprocess).
• The dataset for bone mesh reconstruction Totalsegmentor_Pelvis_Bone_Recon_Dataset is released.

• SdAOF introduces continuous occupancy field representation for efficient high-resolution bone shape reconstruction from biplanar X-ray images.
• SdAOF integrates a spatial-division augmented distillation strategy to capture prevalent occlusion relationship within X-ray images.
• SdAOF achieves SoTA performance in biplanar bone shape reconstruction, reconstructing fine-scale bone surfaces and is scalable to different resolutions without retraining.

• The required CUDA version is 11.3, please ensure using this version in your environment, you may refer to this guideline to keep a separate CUDA version for the new environment.

• Download the code, create a conda environment, and install the required packages in requirements.txt by running the following commands:

  git clone https://github.com/xmed-lab/SdAOF.git
  cd SdAOF
  conda create -n SdAOF python=3.9
  conda activate SdAOF
  pip install -r ./requirements.txt

• TIGRE is required for simulating DRRs and geometry utils, install TIGRE by running

  wget https://github.com/CERN/TIGRE/archive/refs/tags/v2.6.zip
  unzip v2.6.zip
  pip install TIGRE-2.6/Python --no-build-isolation

• Pointnet2 is required for calculating chamfer distance and earth mover's distance metrics, install Pointnet 2 by running

  cd ./scripts/metrics/pointnet2
  python setup.py install

• We provide the processed Totalsegmentor_Pelvis_Bone_Recon_Dataset links: BaiduNetdisk; OneDrive; Huggingface

• Since processing the dataset is time-consuming, we suggest directly downloading and using the processed Totalsegmentor_Pelvis_Bone_Recon_Dataset.

• We also provide step-by-step instructions to reproduce the preprocessing pipeline in ./Data_Preprocess. If you are interested, please refer to it for more details.

• We provide the checkpoint of trained Teacher OF Net and SdAOF Net in the following links, you can download them for direct inference

  Net	Link
  Teacher OF Net	BaiduNetdisk; OneDrive
  SdAOF Net	BaiduNetdisk; OneDrive

• All scripts related to training and evaluation are in ./scripts. Change to this directory first: cd ./scripts.

• The training of SdAOF consists of two stages.

  • Train the Teacher OF Net by running

    CUDA_VISIBLE_DEVICES="YOUR_GPU_ID" python train_teacher_OFNet.py --name Teacher_OF_Net --data_root YOUR_DIRECTORY_TO_DATASET --out_res 128 --save_interval 50 --eval_interval 50 

    The saved checkpoint is in ./scripts/logs/Teacher_OF_Net/best_epoch.pth by default.
  • Train the SdAOF Net by running

    CUDA_VISIBLE_DEVICES="YOUR_GPU_ID" python train_SdAOFNet.py --name SdAOF --data_root YOUR_DIRECTORY_TO_DATASET --teacher_ckpt_path YOUR_DIRECTORY_TO_SAVED_TEACHER_OFNET_CHECKPOINT --out_res 128 --save_interval 50 --eval_interval 50 --distill_individual_feat_alpha 0.2  

    The saved checkpoint is in ./scripts/logs/SdAOF/best_epoch.pth by default.
  • During training, you can config your wandb account, and track training statistics by adding configurations as below

    --wandb --project WANDB_PROJECT_NAME --entity YOUR_WANDB_ENTITY

• Evaluate the trained SdAOF by running

  CUDA_VISIBLE_DEVICES="YOUR_GPU_ID" python eval.py --name SdAOF --data_root YOUR_DIRECTORY_TO_DATASET --ckpt_path YOUR_DIRECTORY_TO_SAVED_SdAOFNET_CHECKPOINT --out_res 336 --eval_mesh_npoint 4096 

  Evaluation results is saved in ./scripts/Eval_results.

If you find our paper helps you, please kindly cite our paper in your publications.

@inproceedings{chen2024spatial,
  title={Spatial-division augmented occupancy field for bone shape reconstruction from biplanar x-rays},
  author={Chen, Jixiang and Lin, Yiqun and Sun, Haoran and Li, Xiaomeng},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={668--678},
  year={2024},
  organization={Springer}
}

We sincerely appreciate XrayTo3DPreprocess for sharing the processing code implementations for Totalsegmentor Dataset.