UDA_Med_Landmark

Code of MICCAI 2023 paper: "Unsupervised Domain Adaptation for Anatomical Landmark Detection".

Installation

1. Clone this repository.

git clone https://github.com/jhb86253817/UDA_Med_Landmark.git

2. Create a new conda environment.

conda create -n uda_med_landmark python=3.9
conda activate uda_med_landmark

3. Install the dependencies in requirements.txt.

pip install -r requirements.txt

Datasets Preparation

• Head: Source domain of cephalometric landmark detection (Download Link). Put the downloaded RawImage and AnnotationsByMD under data/Head/.
• HeadNew: Target domain of cephalometric landmark detection (Download Link). If the link is not available, please read the Section 4 "Usage Notes" of paper (https://arxiv.org/pdf/2302.07797.pdf) for data downloading. Put the downloaded Cephalograms and Cephalometric_Landmarks of the training set under data/HeadNew/.
• JSRT: Source domain of lung landmark detection (Download Link). Put the downloaded All247images under data/JSRT/. Collect the landmark annotations from HybridGNet and put them under data/JSRT/annos/. Run the preprocess.py under data/JSRT to generate Images.
• MSP: Target domain of lung landmark detection, which consists of three datasets: Montgomery (Download Link), Shenzhen (Download Link), and Padchest (Download Link). 1) For Montgomery, put the downloaded CXR_png under data/Montgomery/. Collect the landmark annotations from HybridGNet and put them under data/Montgomery/annos_RL/ and data/Montgomery/annos_LL/. Run the preprocess.py under data/Montgomery to generate Images; 2) For Shenzhen, put the downloaded CXR_png under data/Shenzhen/. Collect the landmark annotations from HybridGNet and put them under data/Shenzhen/annos_RL/ and data/Shenzhen/annos_LL/. Run the preprocess.py under data/Shenzhen to generate Images; 3) For Padchest, download the landmark annotations from here, unzip it and put it under data/Padchest/annos. Then select those images with landmark annotations and put them under data/Padchest/Images/.

You will have the following structure:

UDA_Med_Landmark
-- data
   |-- Head
       |-- RawImage
       |-- AnnotationsByMD
   |-- HeadNew
       |-- Cephalograms
       |-- Cephalometric_Landmarks
       |-- img2size.json
       |-- img2dist.json
       |-- train_list.txt
       |-- test_list.txt
   |-- JSRT
       |-- All247images
       |-- preprocess.py
       |-- Images
       |-- annos
   |-- Montgomery
       |-- CXR_png
       |-- preprocess.py
       |-- Images
       |-- annos_RL
       |-- annos_LL
       |-- train_list.txt
       |-- test_list.txt
   |-- Shenzhen
       |-- CXR_png
       |-- preprocess.py
       |-- Images
       |-- annos_RL
       |-- annos_LL
       |-- train_list.txt
       |-- test_list.txt
   |-- Padchest
       |-- Images
       |-- annos
       |-- train_list.txt
       |-- test_list.txt

Training

Take cephalometric landmark detection as example.

1. Go to folder lib, run preprocess.py Head and preporcess.py HeadNew to preprocess the two datasets, respectively.
2. Back to folder UDA_Med_Landmark, configure the command in run_train.sh as needed, then run bash run_train.sh to start training.

Testing

Take cephalometric landmark detection as example.

1. Preprocess Head and HeadNew the same way as in training.
2. Back to folder UDA_Med_Landmark, configure the command in run_test.sh as needed, then run bash run_test.sh to start testing.

Trained model weights:

• cephalometric landmark detection model
• lung landmark detection model