SFusion: Self-attention based N-to-One Fusion Block

Our implementation is on an NVIDIA RTX 3090 (24G) with PyTorch 1.8.1.

Datasets

We use the BraTS2020 dataset, an open-source dataset.
Please download and unzip the 'MICCAI_BraTS2020_TrainingData' into ./dataset.
Then, please cd ./process and run the following commands to prepare the data:

python split.py

Training Examples

python train.py --phase train --model_name TF_RMBTS

Saved models can be found at ./checkpoint. model_name includes : 'TF_U_Hemis3D', 'U_Hemis3D', 'RMBTS', 'TF_RMBTS', 'LMCR', 'TF_LMCR' .

Note that 'TF_RMBTS' refers to 'SF_FDGF'.

Test Examples (Please train the model before test.)

python train.py --phase test --model_name TF_RMBTS

Brain tumor segmentation results for test data can be found at ./checkpoint.

Evaluation

python evaluation.py --model_name TF_RMBTS