CNILP

A description projection for our Cross-Neighbor-Information based Pseudo-Label for Better Semi-Scribble-Supervised Medical Image Segmentation

Data Preprocess

To prepare the dataset, you can follow the work of WLSMIS

You can also run python code/scribbles_generator.py for scribble label generating and

python code/dataloaders/word_data_processing.py for data preprocessing.

Our work is based on both the 2D medical volumes.

The scribble annotations we used for WORD dataset are available at ./WORD_scribble_labels

Model Training

Run

cd code
python train_CNIPL.py --gpu 0  --check 500 --labeled_ratio 8 --early_stop 20000 --fold fold1 --num_classes 4  --root_path ../data/ACDC --exp ACDC_CNIPL  --max_iterations 60000 --batch_size 16 
python train_CNIPL.py --gpu 0  --check 500 --labeled_ratio 8 --early_stop 20000 --fold fold2 --num_classes 4  --root_path ../data/ACDC --exp ACDC_CNIPL  --max_iterations 60000 --batch_size 16 
python train_CNIPL.py --gpu 0  --check 500 --labeled_ratio 8 --early_stop 20000 --fold fold3 --num_classes 4  --root_path ../data/ACDC --exp ACDC_CNIPL  --max_iterations 60000 --batch_size 16 
python train_CNIPL.py --gpu 0  --check 500 --labeled_ratio 8 --early_stop 20000 --fold fold4 --num_classes 4  --root_path ../data/ACDC --exp ACDC_CNIPL  --max_iterations 60000 --batch_size 16 
python train_CNIPL.py --gpu 0  --check 500 --labeled_ratio 8 --early_stop 20000 --fold fold5 --num_classes 4  --root_path ../data/ACDC --exp ACDC_CNIPL  --max_iterations 60000 --batch_size 16
python val_ours.py

for model training and evaluating. Have fun.

License

This repository is released under MIT License.