A description projection for our JAMIT presentation `Semi-Scribble MRI Image Segmentation via Confidence- and Distance-based Pseudo-label Refinement'
Download the dataset from the ACDC website and prepare the dataset following WSL4MIS
It is important that the whole 150 cases of the ACDC dataset should be preprocessed.
Train the model with 8 labeled volumes:
cd code
python train_dist_unce.py --gpu 0 --labeled_ratio 8 --check 500 --early_stop 10000 --fold fold5 --num_classes 4 --root_path ../data/ACDC --exp ACDC_dist_unce --max_iterations 60000 --batch_size 16 &
python train_dist_unce.py --gpu 1 --labeled_ratio 8 --check 500 --early_stop 10000 --fold fold5 --num_classes 4 --root_path ../data/ACDC --exp ACDC_dist_unce --max_iterations 60000 --batch_size 16 &
python train_dist_unce.py --gpu 2 --labeled_ratio 8 --check 500 --early_stop 10000 --fold fold5 --num_classes 4 --root_path ../data/ACDC --exp ACDC_dist_unce --max_iterations 60000 --batch_size 16 &
python train_dist_unce.py --gpu 3 --labeled_ratio 8 --check 500 --early_stop 10000 --fold fold5 --num_classes 4 --root_path ../data/ACDC --exp ACDC_dist_unce --max_iterations 60000 --batch_size 16 &
python train_dist_unce.py --gpu 4 --labeled_ratio 8 --check 500 --early_stop 10000 --fold fold0 --num_classes 4 --root_path ../data/ACDC --exp ACDC_dist_unce --max_iterations 60000 --batch_size 16
python val_ours.py
The result of your experiment should be shown in output_8.csv
Have fun!
Acknowledgement
A part of our code is from WSL4MIS and DMPLS
I am grateful to Dr. Zhengzhou for the technical support.