In this work, we have formulated an ACMT-Net incorporating a novel CPSA module. ACMT-Net is designed to accurately predict the change of one point set prompted by the movement of another point set. We further proposed a novel k-NN-based contrastive learning approach for pre-training the attentive correspondence between bony and facial point sets, enhancing its capability to model spatial correspondence. The proposed ACMT-Net attains the same level of accuracy as the state-of-the-art FEM simulation method, while considerably reducing the computational time required during the surgical planning processes. For more details on the network, please refer to our MICCAI 2022 paper, MedIA 2024 paper or the pre-print version available on arXiv.
Our work has been awarded MICCAI 2022 Young Scientist Publication Award and MICCAI 2022 Student Travel Award.
Prerequisites
- Linux (tested under Ubuntu 16.04 )
- Python (tested under 2.7)
- TensorFlow (tested under 1.4.0-GPU )
- numpy, h5py
The code is built on the top of PointNET++. Before run the code, please compile the customized TensorFlow operators of PointNet++ under the folder "/Prediction_net/tf_ops".
Train and test
To trian a model:
python -u run.py --mode=train --gpu=0
To test the trained model:
python -u run.py --mode=test --gpu=0
Note that the train and test hdf5 files have been set in the program. If you have errors when running this code please check ALL the path first.
**Conference version
@inproceedings{fang2022deep,
title={Deep Learning-Based Facial Appearance Simulation Driven by Surgically Planned Craniomaxillofacial Bony Movement},
author={Fang, Xi and Kim, Daeseung and Xu, Xuanang and Kuang, Tianshu and Deng, Hannah H and Barber, Joshua C and Lampen, Nathan and Gateno, Jaime and Liebschner, Michael AK and Xia, James J and others},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={565--574},
year={2022},
organization={Springer}
}
**Journal version
@article{fang2024correspondence,
title={Correspondence attention for facial appearance simulation},
author={Fang, Xi and Kim, Daeseung and Xu, Xuanang and Kuang, Tianshu and Lampen, Nathan and Lee, Jungwook and Deng, Hannah H and Liebschner, Michael AK and Xia, James J and Gateno, Jaime and others},
journal={Medical Image Analysis},
pages={103094},
year={2024},
publisher={Elsevier}
}