3DUS-US/MRI Rigid Registration

We proposed a vessel-based 3DUS-CT/MRI rigid registration for liver tumor ablation (see paper here). Our previous work shows the way of collecting 3D US images. To achieve the 3D US-CT/MRI registration task, firstly, we trained a nnUNet model to automatically segment the vessels from 3D US and CT/MRI data. Next, the vessel surface models (point clouds) and centerlines are extracted by using the 3D Slicer software. Lastly, the coarse-to-fine registration algorithm can help us align them together. The worklow is shown below.

Citation

If you use this code for your research, please cite our publications:

@article{xing20233d,
  title={3D US-CT/MRI registration for percutaneous focal liver tumor ablations},
  author={Xing, Shuwei and Romero, Joeana Cambranis and Roy, Priyanka and Cool, Derek W and Tessier, David and Chen, Elvis CS and Peters, Terry M and Fenster, Aaron},
  journal={International Journal of Computer Assisted Radiology and Surgery},
  volume={18},
  number={7},
  pages={1159--1166},
  year={2023},
  publisher={Springer}
}
@article{xing20223d,
  title={3d us-based evaluation and optimization of tumor coverage for us-guided percutaneous liver thermal ablation},
  author={Xing, Shuwei and Romero, Joeana Cambranis and Cool, Derek W and Mujoomdar, Amol and Chen, Elvis CS and Peters, Terry M and Fenster, Aaron},
  journal={IEEE Transactions on Medical Imaging},
  volume={41},
  number={11},
  pages={3344--3356},
  year={2022},
  publisher={IEEE}
}

Introduction

• ./vessel_segmentation_nnUNet. The well-trained nnUNet model for segmenting vessels directly from 3D US images. The checkpoint can be directly used, combined with the configured nnUNet enviroment.
• main.m. This file includes the complete coarse-to-refine registration algorithm to align the 3D US and CT/MRI images. Note that the vessels should be formatted as surface models represented as surface point clouds, and centerlines represented as a bunch of sampled points. CloudCompare can be used to extract the surface point clouds, and 3D Slicer is used for extracting centerlines. Both exported file formats are .txts.
• Registration_evaluation_TRE/centerlineDistance.m. These two files are used for evaluating the registration accuracy. "ReadRegisteredCenterlineParameters.m" shows you how to read the registered parameters.

(If you can not download the trained models due to the huge size, please go to file "/vessel_segmentation_nnUNet/dropbox_link.txt". The shared dropbox link is included, you can download that externally.)

Setup

• Matlab 2023b. We used Matlab 2023b to run the registration algorithm. For Matlab, it is quite flexible, any different versions can also be used to run this code.
• nnUNet environment. Please follow this repo to configure the environment, which is to use our trained vessel segmentation model.
• 3D slicer (VMTK module). This module is to automatically extract the vessel centerlines and export the compatible format for our registration algorithm.
• CloudCompare. This software is to convert the vessel surface models (.stl or .obj) to the point clouds format (.txt).