Uncertainty-Aware Label Rectification for Domain Adaptive Mitochondria Segmentation
Siqi Wu, Chang Chen, Zhiwei Xiong, Xuejin Chen, Xiaoyan Sun. Uncertainty-Aware Label Rectification for Domain Adaptive Mitochondria Segmentation. In MICCAI 2021.
Requirements
Anaconda>=5.2.0 (Python 3.6)
PyTorch>=1.1.0
One or more GPUs with sufficient memory
Memory>=128 GB for data caching
Datasets
To access MitoEM, download files from https://mitoem.grand-challenge.org and convert to arrays.
cd mitoem
python png2npy.py
python tif2npy.py
Otherwise, download .npy files from https://pan.baidu.com/s/1wt1giVGjreYXuArxfOuo1Q (key: us4d).
unzip mitoem-train.zip -d mitoem
unzip mitoem-valid.zip -d mitoem
To access a part of raw images from FAFB and the corresponding labels.
cd fafb-valid/im
cd fafb-valid/seg
Test the pre-trained models
For the reproduction of results listed in Tables 1 and 2.
cd <Name-of-Folder>
python inference.py
Note that, it may take minutes to calculate all of the four metrics.
Train the model
Train an uncertainty-aware model with data from source domain (Rat).
cd u2d-bc-rat-uc-train
python main.py
Generate, rectify, and cache pseudo labels for training.
cd u2d-bc-r2h-train
python inference_4train.py
python generate_mask.py
Train a model with generated labels on target domain (Human).
python main.py