CLMS: Bridging Domain Gaps in Medical Imaging with Continual Learning for Source-Free Adaptation

This is the official PyTorch implementation of CLMS (Continual Learning Multi-Scale) .

CLMS: Bridging Domain Gaps in Medical Imaging with Continual Learning for Source-Free Adaptation.

Get Started

Environment

• Python=3.7.7
• torch=1.7.0
• torchvision=0.8.1
• SimpleITK=2.2.0
• Pillow=8.2.0
• opencv-python=4.5.5.64
• scikit-image=0.17.2
• scikit-learn=0.23.2
• scipy=1.2.0

Datasets Preparation

• It is necessary to set the paths for the training data required for training the model in the options/base_options.py file. Set the path for the target domain training dataset in the targetLibPath field of options/base_options, and specify the path for the augmented domain dataset in the augLibPath field.

• Format of Target Domain Training Set Files:

  • Files in targetLibPath are of type .pt.
  • They can be accessed using torch.load().
  • The file format is a dictionary with a key 'slide', and the corresponding value is a list containing paths to target domain images. Example of Target Domain Training Set File (targetDomain_train.pt):
  • {'slide': ['/target/image1.jpg', '/target/image2.jpg', ...] }

• Format of Augmented Domain Training Set Files:

  • Files in augLibPath are also of type .pt.
  • Accessible using torch.load().
  • The file format is a dictionary with a key 'slide', and the corresponding value is a list containing paths to target domain images. Example of Augmented Domain Training Set File (augmentation_train.pt):
  • {'slide': ['/aug/image1.jpg', '/aug/image2.jpg', ...] }

Load Source Model Weights file

• After completing data preparation, it is necessary to load the source domain model weight. In the options/base_options.py file, the checkpoints_dir field represents the path to the source domain model weights.

Train

• Before training, you need to configure the weight saving path in the checkpoints_dir field in options/base_options.py. Set the desired batch size in the batch_size field, and specify the GPUs for training in the gpu_ids field.

• In options/train_options.py, adjust the n_epochs field to set the number of training epochs, and configure the learning rate with the lr parameter.

• For other training-related settings, please carefully review the comments in options/base_options.py and train_options.py.

• After completing the aforementioned setup, execute the command python train.py to commence the training process.

• Please note that this framework supports multi-GPU parallel training.

Test

• After completing the training, you can set the paths for the target domain test set and source domain test set, as well as the loaded model weights in the test/test.py file. Execute the command python test.py to start the testing process.

Reference