Learning From Synthetic CT Images via Test-Time Training for Liver Tumor Segmentation.

Implementation of Learning From Synthetic CT Images via Test-Time Training for Liver Tumor Segmentation.

Overview of the proposed Synthetic-to-Real Test-Time Training (SR-TTT) method for learning from synthetic images in the liver tumor segmentation task.

Usage

1. Packages

   pytorch==1.9.0

2. Data-preprocessing

   ├── Train
   │   ├── D08
   │   │   ├── Config (configure file)
   │   │   ├── Image (CT slices in png format)
   │   │   ├── Liver (liver mask in png format)
   ├── Test
   │   ├── D08
   │   │   ├── Config (configure file)
   │   │   ├── Image (CT slices in png format)
   │   │   ├── Liver (liver mask in png format)
   │   │   ├── Liver_nii (liver mask in nifti format)
   │   │   ├── Gt_nii (groundtruth for evaluation)

• Convert the nifti images to int32 png format, then subtract 32768 from the pixel intensities to obtain the original Hounsfield unit (HU) values, saved in Image folder, similar to the processing steps in Deeplesion.
• The liver regions can be extracted by a leading liver segmentation model provided by nnU-Net, saved in Liver(png format) and Liver_nii(nifti format).
• Config, configure file of the dataset.
• Dataset of MSD08 can be downloaded from the link. LiTS dataset can be processed using the above steps, we do not provide all the processed images due to its large dataset size.

3. Training

   python segment_train.py

4. Test-time Training and Testing

   python segment_test.py

Citation

If you find this repository useful for your research, please cite the following:

F. Lyu, M. Ye, A. J. Ma, T. C. -F. Yip, G. L. -H. Wong and P. C. Yuen, 
"Learning From Synthetic CT Images via Test-Time Training for Liver Tumor Segmentation," 
in IEEE Transactions on Medical Imaging, doi: 10.1109/TMI.2022.3166230.