Stain Disentanglement Network for staining transformation in WSI.
Python 3.7
PyTorch 1.12.0
A HE stained dataset MITOS-ATYPIA-14 provides frames scanned by two scanners: Aperio Scanscope XT and Hamamatsu Nanozoomer 2.0-HT. We reorganized the file structure as:
dataset/
├── testing
│ ├── A06_00Aa.tiff
│ └── H06_00Aa.tiff
│ └── ...
└── training
├── A03_00Aa.tiff
└── H03_00Aa.tiff
└── ...
In train.py, the process of training SDN ia shown. The default setting is to train SDN on frames whose names are started with H (Hamamatsu Nanozoomer 2.0-HT). During training, the program saves some validation results in ./output.
In test.py, we show the testing process using a reference image for random cropped patch transformation. With the aligned frames scanned by two scanners in MITOS-ATYPIA-14, SSIM and PSNR can be evaluated during testing.