Neural Rendering for Stereo 3D Reconstruction of Deformable Tissues in Robotic Surgery

Implementation for MICCAI 2022 paper Neural Rendering for Stereo 3D Reconstruction of Deformable Tissues in Robotic Surgery by Yuehao Wang, Yonghao Long, Siu Hin Fan, and Qi Dou.

A NeRF-based framework for Stereo Endoscopic Surgery Scene Reconstruction (EndoNeRF).

[Paper] [Website] [Sample Dataset]

Demo

Setup

We recommend using Miniconda to set up an environment:

cd EndoNeRF
conda create -n endonerf python=3.6
conda activate endonerf
pip install -r requirements.txt
cd torchsearchsorted
pip install .
cd ..

We managed to test our code on Ubuntu 18.04 with Python 3.6 and CUDA 10.2.

Dataset

To test our method on your own data, prepare a data directory organized in the following structure:

+ data1
    |
    |+ depth/           # depth maps
    |+ masks/           # binary tool masks
    |+ images/          # rgb images
    |+ pose_bounds.npy  # camera poses & intrinsics in LLFF format

In our experiments, stereo depth maps are obtained by STTR-Light and tool masks are extracted manually. Alternatively, you can use segmentation networks, e.g., MF-TAPNet, to extract tool masks. The pose_bounds.npy file saves camera poses and intrinsics in LLFF format. In our single-viewpoint setting, we set all camera poses to identity matrices to avoid interference of ill-calibrated poses.

Training

Type the command below to train the model:

export CUDA_VISIBLE_DEVICES=0   # Specify GPU id
python run_endonerf.py --config configs/{your_config_file}.txt

We put an example of the config file in configs/example.txt. The log files and output will be saved to logs/{expname}, where expname is specified in the config file.

Reconstruction

After training, type the command below to reconstruct point clouds from the optimized model:

python endo_pc_reconstruction.py --config_file configs/{your_config_file}.txt --n_frames {num_of_frames} --depth_smoother --depth_smoother_d 28

The reconstructed point clouds will be saved to logs/{expname}/reconstructed_pcds_{epoch}. For more options of this reconstruction script, type python endo_pc_reconstruction.py -h.

We also build a visualizer to play point cloud animations. To display reconstructed point clouds, type the command as follows.

python vis_pc.py --pc_dir logs/{expname}/reconstructed_pcds_{epoch}

Type python vis_pc.py -h for more options of the visualizer.

Evaluation

First, type the command below to render left views from the optimized model:

python run_endonerf.py --config configs/{your_config_file}.txt --render_only

The rendered images will be saved to logs/{expname}/renderonly_path_fixidentity_{epoch}/estim/. Then, you can type the command below to acquire quantitative results:

python eval_rgb.py --gt_dir /path/to/data/images --mask_dir /path/to/data/gt_masks --img_dir logs/{expname}/renderonly_path_fixidentity_{epoch}/estim/

Note that we only evaluate photometric errors due to the difficulties in collecting geometric ground truth.

Bibtex

If you find this work helpful, you can cite our paper as follows:

@inproceedings{wang2022neural,
    title={Neural Rendering for Stereo 3D Reconstruction of Deformable Tissues in Robotic Surgery},
    author={Wang, Yuehao and Long, Yonghao and Fan, Siu Hin and Dou, Qi},
    booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
    pages={431--441},
    year={2022},
    organization={Springer}
  }

Acknowledgement

• Our code is based on D-NeRF and NeRF-pytorch.
• Evaluation code is borrowed from this repo by @peihaowang.