stain-normalizer

A color normalization method for whole slide image(WSI) tiles, using cupy and tiffile with GPU to accerate the process, which result in a at least 100X faster speed.

Requirements

cupy
one GPU with according cuda installed

How to use

import tifffile as tif
from stain-normalizer import Normalizer  #if u copy StainNormalizer.py file in working directory
template_tile = tif.TiffSeqence(template_path).asarray() ## u can choose one or several template tiles, but make sure shape of template_tile is 4-D array
target_tile = tif.imread(target_tile_lists) ## return a 4-D array
normalizer = Normalizer()
normalizer.fit(template_tile)     
normalizer.normalize(target_tile)

Comparement

read 2000 tiles(512* 512 * 3) into numpy
- tifffile.imread read: 1m17s
- Image.open read: 9m38s
- cv2.imread : 10m46s
color normalization 30k tiles with chunksize=100,V100(16G) without writing tiles to disk: 21m 46s.

Suggestions

use tifffile to load image
use multi GPU. Cupy uses one GPU by default.
Use LMDB or H5DF to load or save image. Or save as numpy format data ".npy". If use tensorflow, tfrecords is suggested. Single image format is least recommeded.
Use multiprocessing or simply cut images into several chunks and run several jobs.

Other implements

pytorch implement: https://github.com/EIDOSlab/torchstain
numpy: https://github.com/schaugf/HEnorm_python
matlab: https://github.com/mitkovetta/staining-normalization
python modules with different methods:https://github.com/Peter554/StainTools

Referrence

[1] Macenko, Marc, et al. "A method for normalizing histology slides for quantitative analysis." 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE, 2009.