JoaoSantinha / sasan

SASAN

Self-Attentive Spatial Adaptive Normalization for Cross-Modality Domain Adaptation

Requirements

• python >=3.6
• pytorch >=1.6
• tensorflow == 1.15
• medpy
• kornia

Training

• To launch the training please run train.py. The hyperparameters can be updated in def main function as a dictionary.

• For faster convergence, please pretrain the attention module for the domain whose segmenation labels are available, by running python train_segmentation.py attention_mr

• For training the upper bount U-Net on MRI modality, use the following command - python train_segmentation.py mr

• To evaluate the trained model, please run python run_evaluation.py sasan ct for evaluating the performance of MRI to CT domain adaptation. For the other direction CT to MRI, run python run_evaluation.py sasan mr.

Pre-trained models, datasets, code:

• Link to our pre-trained models on Whole Heart Multimodal dataset and code.
• Link to Whole Heart Multimodal dataset pre-processed training tf_record files can be found here. The test mr data is available here and test ct data is available here

Data preprocessing

• To convert the tf_records training data to .npy format please use the script convert_tfrecords.py <modality>, where <modality> is either mr or ct.

Licence

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.