Sample informative tiles from WSIs taken from TCGA with UASF. The model takes as inputs weak-labeled tiles, where tiles are given their WSI label. The model adopted Monte-Carlo dropout that are used for variational inference to generate not only a prediction probability but also an uncertainty measure.
Create a virtual environment and install the required packages
conda env create --file env.yml
source activate uacnnconda create -n uacnn python=3.7 openslide pytorch torchvision torchaudio cudatoolkit=11.4 -c pytorch -c conda-forge
conda activate uacnn#HistomicsTK
pip install histomicstk --find-links https://girder.github.io/large_image_wheels
#Additional packages
pip install tqdm rasterio large-image-source-openslide tensorboard pytorch-lightningNote: code was run with python 3.7
To insure reproducibility of the results, data splits and tiles coordinates are provided in MICCAI_data.gz To uncompress it, run
tar -xzvf MICCAI_data.tar.gzWe originally downloaded the whole-slide images from the TCGA data portal https://portal.gdc.cancer.gov/ via the gdc-client tool. Go to dedicated folder to store files (all FFPE slides from TCGA is approx. 1.0 TB). Download images using the corresponding manifest file:
mkdir data
cd data
mkdir slides
cd slides
gdc-client download -m gdc_manifest.TCGA_SARC_dx_slides_2022-05-28The code in preprocessing folder is designed to extract non-overlapping tiles directly from whole-slide images. We extracted 256x256 pixel tiles at x10 magnification level. To extract tiles run:
python extract_tiles.py --tile [path/to/slides/folder] --wsi [file_name.svs] -wd [tile_width] -ht [tile_height] -m[magnification] -o [path/to/output/]
#Example
python preprocessing/extract_tiles.py --tile data/slides/ --wsi TCGA-X9-A973-01Z-00-DX5.D6A52779-A0A4-4119-9AB4-1A7A6BD98337.svs -m 10 -wd 256 -ht 256 -o data/tiles/To run an experiment, write first an app file or use the example available in folder app/LMS/. You can modify tiling and training parameters in app/LMS/config file. Launch training experiment with a train-validation split:
cd app/LMS/
python lms_tcga_sarc.py train --img_dir data/tiles/ \
--train_data_desc MICCAI_data/LMS_tiles_clinical_data_fold1_train.csv \
--val_data_desc MICCAI_data/LMS_tiles_clinical_data_fold1_val.csv \
--weights=histossl \
--task=FNCLCC_gradeIn our experiments, we compared the performance of UA-CNN preteined on Imagnet and UA-CNN preteined on Histossl. Histossl is a pretrained model for self supervised histopathology. Reference: Self supervised contrastive learning for digital histopathology. You can download it from here:
https://github.com/ozanciga/self-supervised-histopathology
python lms_tcga_sarc.py eval_uncertainty --img_dir data/tiles/ \
--train_data_desc MICCAI_data/LMS_tiles_clinical_data_fold1_train.csv \
--val_data_desc MICCAI_data/LMS_tiles_clinical_data_fold1_val.csv \
--weights=LMS_SCE_uacnn.pth \
--task=FNCLCC_grade python lms_tcga_sarc.py wsi_uncertainty_map --img_dir data/slides/TCGA-X9-A973-01Z-00-DX5.D6A52779-A0A4-4119-9AB4-1A7A6BD98337.svs \
--weights=LMS_SCE_uacnn.pth --tile_size=256 --mag=10 --task=FNCLCC_grade --nclasses=3 --class2index="0,1,2"