Medical-Detection3d-Toolkit

PyTorch implementation of the volumetric landmark detection engine proposed in the paper SkullEngine: A Multi-stage CNN Framework for Collaborative CBCT Image Segmentation and Landmark Detection, MICCAI workshop 2021.

Installation

The code is tested with python=3.8.8, torch=2.0.0, and torchvision=0.15.0 on an A6000 GPU.

git clone https://github.com/qinliuliuqin/Medical-Detection3d-Toolkit
cd Medical-Detection3d-Toolkit

Create a new conda environment and install required packages accordingly.

conda create -n det3d python=3.8.8
conda activate det3d
pip3 install -r requirements.txt

Data

First, the users need to prepare medical images and their corresponding landmark annotations. The assets folder contains an example image (case_001.nii.gz) and landmark annotation file (case_001.csv). Then, generate landmark masks (e.g, case_001_landmark_mask.nii.gz) given the pairs as demonstrated in this notebook. The meaning of labels in the landmark mask:

Label	Meaning
positive integer	these are positive samples of landmarks (e.g., voxels with value 1 represent the first landmark.)
0	these are negative samples (i.e, background voxels)
-1	these are boundary voxels (i.e., between the positive and negative ones) that are not involved in training.

Finally, prepare dataset splitting files for training (train.csv) and testing (test.csv).

Training

Run the following code for training with a single GPU. The user may need to modify training settings in ./config/lmk_train_config.py. By default, the model will be saved in ./saves/weights.

cd detection3d
python lmk_det_train.py --input ./config/lmk_train_config.py --gpus 0

Evaluation

Run the following code to evaluate a trained model on a single GPU.

python lmk_det_infer.py -i ../assets/case_001.nii.gz -m ./saves/weights -o ./saves/results

Citation

@article{liu2021skullengine,
  title={SkullEngine: A Multi-stage CNN Framework for Collaborative CBCT Image Segmentation and Landmark Detection},
  author={Liu, Qin and Deng, Han and Lian, Chunfeng and Chen, Xiaoyang and Xiao, Deqiang and Ma, Lei and Chen, Xu and Kuang, Tianshu and Gateno, Jaime and Yap, Pew-Thian and others},
  journal={arXiv preprint arXiv:2110.03828},
  year={2021}
}