## RoCoSDF: Row-Column Scanned Neural Signed Distance Fields for Freehand 3D ultrasound Imaging Shape Reconstruction

--------------------------------------

The official implementation code for MICCAI 2024 paper:

[RoCoSDF: Row-Column Scanned Neural Signed Distance Fields for Freehand 3D Ultrasound Imaging Shape Reconstruction](https://chenhbo.github.io/RoCoSDF/)

by [Hongbo Chen](https://chenhbo.github.io/), Yuchong Gao, Shuhang Zhang, Jiangjie Wu, [Yuexin Ma](https://yuexinma.me/) and [Rui Zheng](https://sist.shanghaitech.edu.cn/zhengrui_en/main.htm).

RoCoSDF is a framework built on neural implicit signed distance functions for shape reconstruction of multi-view freehand 3D ultrasound imaging.

<div align="center">

<img src="img/Fig_RoCoScan.png" style="zoom:15%" alt="Data Aquisition Protocol"/>

</div>

## Demo

* Thoracic Vertebra T4 from ultrasound transducer 1 (UT1)

<div align="center">

<img src="img/Fig_Result_T4.png" style="zoom:14.7%" alt="Framework"/>

</div>

<br />

* The example mesh results of RoCoSDF are in `outs/T4_RoCo/outputs/*.ply`.

--------------------------------------

## Usage

Our code is implemented in NVIDIA 3090, Ubuntu 18/20, Python 3.8, PyTorch 1.12.1 and CUDA 11.6.

### Install Dependencies 

For 20/30x GPU:

```

conda create -n rocosdf python=3.8

conda activate rocosdf

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge

pip install tqdm pyhocon==0.3.57 trimesh PyMCubes scipy matplotlib

pip install visdom open3d scikit-image plyfile

```

For 40x GPU, cuda 11.8:

```

conda create -n rocosdf python=3.10

conda activate rocosdf

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install tqdm pyhocon==0.3.57 trimesh PyMCubes scipy matplotlib

pip install visdom open3d scikit-image plyfile

```

--------------------------------------

### Data Preparation

- Convert the row-scan and column-scan segmented volumetric mask to point clouds file *.ply.

- Both the row-scan and column-scan point clouds should be in a same tracking 

space or manually aligned in a unified space.

- Put the row-scan and column-scan point clouds data in ./data.

```

RoCoSDF/

│

├── data/

│   ├── T4_Co.ply            % your own data

|   ├── T4_Ro.ply            % your own data

│   ├── T4_Co_ds.pt          % generated during data preprocessing, downsampled point clouds for training

|   ├── T4_Ro_ds.pt          % generated during data preprocessing, downsampled point clouds for training

│   ├── T4_Co_sampler.pt     % generated during training

|   ├── T4_Ro_sampler.pt     % generated during training

|

|

├── outs/

│   ├── T4_Co/

│       └── outputs/

|           └── *.ply

│   ├── T4_Ro/

│       └── outputs/

|            └── *.ply

│   ├── T4_RoCo/

│       └── outputs/

|            └── *.ply

|

```

--------------------------------------

### Run RoCoSDF

In Linux, directly train the model through `sh train.sh` OR using command as below.

```

python runRoCoSDF.py --gpu 0  --conf confs/conf.conf --dataname T4_Co --dataname2 T4_Ro  --dir T4_Co --dir2 T4_Ro --dir3 T4_RoCo --mode train

 ```

### Run SDF Refinement Only

In Linux, directly train the model through `sh train_refine_only.sh` OR using command as below.

```

python runRoCoSDF.py --gpu 0  --conf confs/conf.conf --dataname T4_Co --dataname2 T4_Ro --dir T4_Co --dir2 T4_Ro --dir3 T4_RoCo --mode train_refine

 ```

### Contact

For any queries, please contact [chenhb[at]shanghaitech.edu.cn](mailto:chenhb@shanghaitech.edu.cn).

### Citation

If you use RoCoSDF in your research, please cite the paper:

```

 @InProceedings{chenRoCoSDF,

   author="Chen, Hongbo

   and Gao, Yuchong

   and Zhang, Shuhang

   and Wu, Jiangjie

   and Ma, Yuexin

   and Zheng, Rui",

   title="RoCoSDF: Row-Column Scanned Neural Signed Distance Fields for Freehand 3D Ultrasound Imaging Shape Reconstruction",

   booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024",

   year="2024",

   publisher="Springer Nature Switzerland",

   address="Cham",

   pages="721--731",

   isbn="978-3-031-72083-3"

   }

```

--------------------------------------

### References

The reference codes are from the following links.

We appreciate all the contributors.

* FUNSR(UNSR in the proceeding): https://github.com/chenhbo/FUNSR

* DeepSDF: https://github.com/facebookresearch/DeepSDF

* NeuralPull: https://github.com/mabaorui/NeuralPull-Pytorch

* GenSDF: https://github.com/princeton-computational-imaging/gensdf

* CSGSDF: https://github.com/zoemarschner/csg_on_nsdf