This is an official implementation of "Temporal-consistent segmentation of echocardiography with co-learning from appearance and shape" (MICCAI 2020 early accept, Oral)
Cardiac Acquisitions for Multi-structure Ultrasound Segmentation, CAMUS
- Visualization of temporal consistency
- Cardiac (endocardium, epicardium, and left atrium) segmentation
| Methods | Endo | Epi | LA | ||||||
|---|---|---|---|---|---|---|---|---|---|
| ED phase | Dice | HD | MAD | Dice | HD | MAD | Dice | HD | MAD |
| U-Net | 0.936 | 5.3 | 1.7 | 0.956 | 5.2 | 1.7 | 0.889 | 5.7 | 2.2 |
| ACNNs | 0.936 | 5.6 | 1.7 | 0.953 | 5.9 | 1.9 | 0.881 | 6.0 | 2.3 |
| CLAS | 0.947 | 4.6 | 1.4 | 0.961 | 4.8 | 1.5 | 0.902 | 5.2 | 1.9 |
| ES phase | Dice | HD | MAD | Dice | HD | MAD | Dice | HD | MAD |
| U-Net | 0.912 | 5.5 | 1.7 | 0.946 | 5.7 | 1.9 | 0.918 | 5.3 | 2.0 |
| ACNNs | 0.913 | 5.6 | 1.7 | 0.945 | 5.9 | 2.0 | 0.911 | 5.8 | 2.2 |
| CLAS | 0.929 | 4.6 | 1.4 | 0.955 | 4.9 | 1.6 | 0.927 | 4.8 | 1.8 |
Note: ED & ES: end-diastole and end-systole phases; HD: Hausdorff distance; MAD: Mean absolute distance
- Volumes (EDV & ESV) and ejection fraction (EF) estimation
| Methods | EDV | ESV | EF | ||||||
|---|---|---|---|---|---|---|---|---|---|
| corr | bias(ml) | std | corr | bias(ml) | std | corr | bias(%) | std | |
| U-Net | 0.926 | 7.2 | 15.6 | 0.960 | 4.4 | 10.2 | 0.845 | 0.1 | 7.3 |
| ACNNs | 0.928 | 2.8 | 15.5 | 0.954 | 2.0 | 10.1 | 0.807 | 0.3 | 8.3 |
| CLAS | 0.958 | -0.7 | 15.1 | 0.979 | -0.0 | 8.4 | 0.926 | -0.1 | 6.7 |
Note: corr: Pearson Correlation Coefficient
- Requirement
pip install torch matplotlib scikit-image opencv-python SimpleITK scipy imageio pillow
- Download dataset with following storage format:
./data/training/patient-id/*
./data/testing/patient-id/*
- Run
Read and preprocess training data
python ./utils/read_data.py
After preprocess, A2C/A4C.npy with size (450, 10, 256, 256) and A2C_gt/A4C_gt.npy with size (450, 2, 256, 256) were generated for apical two-chamber and four-chamber views in the folder "./data/". 450 means the number of patients for training, 10 means the number of sampling frames for each sequences, and only end-diastole and end-systole frames have ground truths.
Training
python train.py
Testing
python test.py
Please cite our paper if you find anything helpful:
@InProceedings{CLAS,
author={Wei, Hongrong and Cao, Heng and Cao, Yiqin and Zhou, Yongjin and Xue, Wufeng and Ni, Dong and Li, Shuo},
title={Temporal-Consistent Segmentation of Echocardiography with Co-learning from Appearance and Shape},
booktitle={Medical Image Computing and Computer Assisted Intervention -- MICCAI 2020},
year={2020},
publisher={Springer International Publishing},
pages={623--632},
isbn={978-3-030-59713-9}
}