6 May 2019:
- Upload AIEchoDx_demo and AIEchoDx_demo_notebook
5 Jan 2020:
- Upload AIEchoDx_demo 1.0.0
- Upload 25 examples of patients' echocardiographic videos in
AIEchoDx/AIEchoDx_demo/video_examples/
AIEchoDx, which stands for “Artificial Intelligence Echocardiogram Diagnosis Network”, is a two-stage neural network to diagnose patients with atrial septal defect (ASD), dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), prior myocardial infarction (prior MI), and normal controls. It reads 45 frames of echocardiographic videos to make the diagnosis.
[1] Will uploade in the future
- 1. Installation and Requirements
- 2. Training AIEchoDx network
- 3. Prediction
- 4. Compare with human physicians
- 5. CAM analysis
- 6. DCM patients clustering
- 7. Examples of patients' echocardiographic videos
System Requirements: Windows (>= 7), Mac OS X (>= 10.8) or Linux
All other software dependencies are listed below:
python version = 3.6.5
tensorflow version = 1.10.0
keras version = 2.2.2
cv2 = 3.1.0
PIL = 5.2.0
pydicom = 1.2.2
seaborn = 0.9.0
Download software from github and cd to AIEchoDx folder
git clone https://github.com/superresolution/AIEchoDx.git
cd ~/AIEchoDx
Make and activate a new Python 3.6 environment
conda env create -f requirement.yml
./setup.sh
After installation, activate aiechodx enviroment:
conda activate aiechodx
This enviroment could be deactivate by:
conda deactivate
The echocardiographic images need pre-preocessing before training in our AIEchoDx network. Here we provided seveal useful python functions to read, write and optimize echocardiographic images in util.py.
In detals, the python functions load_dcm_video and load_avi_video are used to read and transfer .dcm and .avi format videos into numpy matrix. The python function load_dcm_information is used to read patients informations. The function limited_equalize is used to optimize the contracts of medical images. The functions remove_info and remove_info2 are used to remove the information of patients.
Othervise, echocardiographic images need to be cut and save into smaller .png format before training.
AIEchoDx is a two-stage network. We train the network separatedly. But in the future, we can combine the two stage networks and train them togather.
Make a data file to store the single images splited from echocardiographic videos; make a model_weight file to store the models and loss value.
cd AIEchoDx_demo
python train_inception_v3.py -d <file>
python prepare_data_for_diagnostic_network.py -m <file1> -s <dir1>
*file1: model weight
*dir1: save data to dir
python train_diagnosis_network.py -t <dir1> -v <dir2> -f <frames>
*dir1: training .txt files
*dir2: validation .txt files
*frames: number of frames; 45 were set as default
python predict.py -v <filename1> -i <filename2> -d <filename3> -x 45 -n 0
*filename1: where the video is saved
*filename2: retrained inception v3 model
*filename3: retraiend diagnostic model
*-x: the last number of the 45 frames in the video
*-x: the first number of the 45 frames video
Will upload in the future
Please see codes 4_CAM_Figure_5.ipynb in AIEchoDx_demo_notebook
The PHATE software could be installed as by pip:
pip install --user phate
Please also see codes 5_DCM_patients_analysis_Figure_6.ipynb in AIEchoDx_demo_notebook
Split videos into single images and save the images in /AIEchoDx-demo/data
python video2images.py
The paper has been accepted in Scientific Reports. Please cite as below:
Liu, B., Chang, H., Yang, D. et al. A deep learning framework assisted echocardiography with diagnosis, lesion localization, phenogrouping heterogeneous disease, and anomaly detection. Sci Rep 13, 3 (2023). https://doi.org/10.1038/s41598-022-27211-w