Detailed and exhausitive annotations on whole-slide images (WSI) are extremely labor-intensive and time-consuming. In this repository, we provide implementation of two methods -- (1) Deep k-NN (DkNN); (2) Label Cleaning Multiple Instance Learning (LC-MIL) -- for refining these coarse annotations, and producing a more accurate version of annotations. The figure below shows an example of the coarse annotations and the refined annotations produced by one of our method (LC-MIL).
Noticeably, although both methods are machine learning based, the refinement can be conducted on each single slide, and NO external data is needed.
- CAMELYON16
- PAIP2019
- PAIP2020
- One WSI form CAMELYON16 for test :
test_016.tif
python train_model.py
positional arguments:
slide_root The location of the whole-slide image.
slide_ID Name of the whole-slide image.
slide_format Dataformat the whole-slide image. Permitted format can
be `.svs`, `,ndpi`, and `.tif`.
ca_path The path to the coarse annotations. File format should
be `.sav`
model_save_root Where model will be saved
optional arguments:
-h, --help show this help message and exit
--remove_blank REMOVE_BLANK
How to remove blank regions (i.e. identify tissue
regions) of WSI. We provide three functions: 0:
convert to HSV and then OTSU threshold on H and S
channel; 1: apply [235, 210, 235] on RGB channel,
respecitively; 2: convert to gray image, then OTSU
threshold. Default is 0. For new dataset, the user is
encouraged to write customed function
--focal_loss FOCAL_LOSS
Whether or not to use focal loss (True: using focal
loss; Flase: using cross entropy), default is false
--patch_shape PATCH_SHAPE
Patch shape(size), default is 256
--unit UNIT Samllest unit when cropping patches, default is 256
--gpu GPU gpu
--lr LR Initial Learning rate, default is 0.00005
--step_size STEP_SIZE
Step size when decay learning rate, default is 1
--reg REG Reg,default is 10e-5
python apply_model.py
positional arguments:
slide_root The location of the whole-slide image.
slide_ID Name of the whole-slide image.
slide_format Dataformat the whole-slide image. Permitted format can
be `.svs`, `,ndpi`, and `.tif`.
ca_path The path to the coarse annotations. File format should
be `.sav`
model_dir Where to load the model (to conduct feature
extraction)
feature_save_root Where the mapped features will be saved
knn_save_root Where the KNN results (distance and index) will be
saved
heatmap_save_root Where the predicted heatmap will be saved
optional arguments:
-h, --help show this help message and exit
--remove_blank REMOVE_BLANK
How to remove blank regions (i.e. identify tissue
regions) of WSI. We provide three functions: 0:
convert to HSV and then OTSU threshold on H and S
channel; 1: apply [235, 210, 235] on RGB channel,
respecitively; 2: convert to gray image, then OTSU
threshold. Default is 0. For new dataset, the user is
encouraged to write customed function
--focal_loss FOCAL_LOSS
Whether or not to use focal loss (True: using focal
loss; Flase: using cross entropy), default is False
--patch_shape PATCH_SHAPE
Patch shape(size), default is 256
--unit UNIT Samllest unit when cropping patches, default is 256
--gpu GPU gpu
cd DkNN
python train_model.py ../Data test_016 .tif ../coarse_annotations.sav .
python apply_model.py ../Data test_016 .tif ../coarse_annotations.sav model_test_016.pth . . .
We can not actually upload our test WSI, test_016.tif to this repository due to the space limit of Github, but you can find it in the google drive
python train_model.py
positional arguments:
slide_root The location of the whole-slide image.
slide_ID Name of the whole-slide image.
slide_format Dataformat the whole-slide image. Permitted format can
be `.svs`, `,ndpi`, and `.tif`.
ca_path The path to the coarse annotations. File format should
be `.sav`
model_save_root Where model will be saved
optional arguments:
-h, --help show this help message and exit
--remove_blank REMOVE_BLANK
How to remove blank regions (i.e. identify tissue
regions) of WSI. We provide three functions: 0:
convert to HSV and then OTSU threshold on H and S
channel; 1: apply [235, 210, 235] on RGB channel,
respecitively; 2: convert to gray image, then OTSU
threshold. Default is 0. For new dataset, the user is
encouraged to write customed function
--length_bag_mean LENGTH_BAG_MEAN
Average length of bag (Binomial distribution),default
= 10
--num_bags NUM_BAGS Number of bags to train,default = 1000
--focal_loss FOCAL_LOSS
Whether or not to use focal loss (True: using focal
loss; Flase: using cross entropy), default is FL
--patch_shape PATCH_SHAPE
Patch shape(size), default is 256
--unit UNIT Samllest unit when cropping patches, default is 256
--gpu GPU gpu
--lr LR Initial Learning rate, default is 0.00005
--step_size STEP_SIZE
Step size when decay learning rate, default is 1
--reg REG Reg,default is 10e-5
python apply_model.py
positional arguments:
slide_root The location of the whole-slide image.
slide_ID Name of the whole-slide image.
slide_format Dataformat the whole-slide image. Permitted format can
be `.svs`, `,ndpi`, and `.tif`.
model_dir The path to the MIL model
heatmap_save_root Where the predicted heatmap will be saved
optional arguments:
-h, --help show this help message and exit
--remove_blank REMOVE_BLANK
How to remove blank regions (i.e. identify tissue
regions) of WSI. We provide three functions: 0:
convert to HSV and then OTSU threshold on H and S
channel; 1: apply [235, 210, 235] on RGB channel,
respecitively; 2: convert to gray image, then OTSU
threshold. Default is 0. For new dataset, the user is
encouraged to write customed function
--length_bag_mean LENGTH_BAG_MEAN
Average length of bag (Binomial distribution),default
= 10
--num_bags NUM_BAGS Number of bags to train,default = 1000
--focal_loss FOCAL_LOSS
Whether or not to use focal loss (True: using focal
loss; Flase: using cross entropy), default is FL
--patch_shape PATCH_SHAPE
Patch shape(size), default is 256
--unit UNIT Samllest unit when cropping patches, default is 256
--gpu GPU gpu
cd LC_MIL
python train_model.py ../Data test_016 .tif ../coarse_annotations.sav .
python apply_model.py ../Data test_016 .tif model_test_016.pth . . .
We can not actually upload our test WSI, test_016.tif to this repository due to the space limit of Github, but you can find it in the google drive
Post-processing procedure for both methods (DkNN and LC-MIL), and the illustration can be found in Post-process.ipynb.
@misc{wang2021label,
title={Label Cleaning Multiple Instance Learning: Refining Coarse Annotations on Single Whole-Slide Images},
author={Zhenzhen Wang and Aleksander S. Popel and Jeremias Sulam},
year={2021},
eprint={2109.10778},
archivePrefix={arXiv},
primaryClass={cs.CV}
}