aijinrjinr / MLB-Seg

Consistency-guided Meta-Learning for Bootstrapping Semi-Supervised Medical Image Segmentation (MLB-Seg)

Pytorch implementation for MLB-Seg

Data

All data will be stored in a folder data

 MLB-Seg (our repo)
    ├── data├── LA├──train
            |     ├──meta_train
            |     ├──original_data
            |     └──split_info.mat
            |
            └── Prostate├──train
                        ├──meta_train
                        ├──original_data
                        └──split_info.mat

• Please download the original LA/PROMISE12 dataset and put it in the corresponding original_data folder which would be used during validation.
• Store each 2D slice from the training set in the corresponding train folder and each 2D slice from the meta set in the corresponding meta_train folder.
• For data in train, make sure it has the format shown below

XX.npy
  ├──'img'
  ├──'label'
  └──'noisy_label'

• 'img' represents the original slice, 'label' is the ground-truth segmentation and 'noisy_label' is the imperfect label (could be noisy annotations or generated labels for unlabelde data).
• For data in meta_train, make sure it has the format shown below

XX.npy
  ├──'img'
  └──'label'

• split_info.mat store the information (name) for each partient in the training/meta/validation set which has the format shown below

split_info.mat
  ├──'train'
  ├──'meta'
  └──'test'

Train

python train.py --dataset Prostate --train_root ./data/Prostate/train/  --meta_root ./data/Prostate/meta_train/ --vali_root ./data/Prostate/original_data/ --checkpoint ./checkpoint/pretrained_model.pth --datasplitpath ./data/Prostate/split_info.mat

• Store the pretrained model (training on the meta set) in checkpoint folder.