Example workflow to predict IDH mutation status from histopathological whole slide images (WSI) using deep learning.
The first step is to tile the slides into patches and store them into specified folder. A csv file with two columns ('SVS_Path','PatientID') is required to store svs file paths and IDs. See csv folder for example file for --df_path argument.
User has to assign two arguments: --df_path and --target_path
python3 Tiling.py --df_path '/path/to/csv/tilingslides.csv' --target_path '/root/folder/path/for/tiles'
Some optional arguments:
'--workers',type=int,default=8
'--tilesize',type=int,default=256
'--stride',type=int,default=256 (when set to a number smaller than tilesize, tiles will have overlaps)
'--tissuepct_value',type=int,default=0.7 (for quality control)
'--magnification',type=str,default='multi' (by default 2.5x, 5x, 10x, 20x will all be tiled, otherwise assign one target magnification). Note that for training, tiles from a specific magnification will be used. It is recommended to examine the validation performance using different magnifications to choose the optimal one.
python3 Tiling.py --df_path '/path/to/csv/file.csv' --target_path '/root/folder/path/for/tiles' --tissuepct_value 0.5 --magnification '2.5x'
After tiling the slides, we can start training models.
Prepare a csv file for --df_path is required. Example can be found in csv folder with name of training.csv. Basically, three columns are required: 1. label column, use --y_col to assign label name; 2. 'Path': paths for the tiles generated from each slide in the previous step; 3. 'Train_Test': containing values of 'Train'/'Validation'/'Test'. Train and Validation are required.
Some important arguments:
'--result_dir', type=str': folder path for saving the model
'--df_path': path to the training meta data
'--gpu': at least two gpus are required, default is '0,1,2,3'
'--patch_n': patch number to sample for each slide during each iteration
'--balance': weights for balancing loss function for each class; default is 0.5. If weighting loss function for class imbalancement, --balance_training_off is suggested
'--balance_training': used to balance class imbalancement during training. Will automatically sample the same number of samples for each class. Suggest leaving --balance as default
'--CNN' choose from resnet and densenet
'--y_col' label name in the csv file. Values must be numeric: 0 or 1
'--freeze_batchnorm' suggest to set this for more stable results
'--pooling': select from attention, max and mean for aggregating embedding layers of tiles from one slide
'--A' if set pooling as attention, set a number to A for number of nodes in attention network. Default is 16.
python3 Train.py --result_dir ‘/path/for/model’ --df_path ‘/path/to/training.csv’ --workers 16 --CNN densenet --no_age --patch_n 200 --spatial_sample_off --n_epoch 100 --lr 0.00001 --optimizer Adam --use_scheduler --balance 0.5 --balance_training --freeze_batchnorm --pooling mean --notes model0
python3 code/Evaluation.py --df_path 'path/to/dataframe/' --y_col='IDH' --Model_Folder 'path/to/your/trained/model' --key_word 'Test' --no_age --two_forward_off
conda env create N where N is the name of your environment
conda activate N where N is the name of your conda environment
nvidia-smi
conda install pytorch torchvision cudatoolkit=11.0 -c pytorch
4. Download and install cuDNN by following these instructions
cp cuda/include/cudnn*.h /path/to/your/environment/include
cp cuda/lib64/libcudnn* /path/to/your/environment/lib
#! /bin/bash -l
#SBATCH --partition=panda-gpu
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=15
#SBATCH --job-name=training
#SBATCH --time=1-00:00:00
#SBATCH --mem=64G
#SBATCH --gres=gpu:2
source ~/.bashrc
conda activate /path/to/your/environment
python3 code/Train.py --result_dir 'data/model/' --df_path 'MetaData_training.csv' --workers 15 --CNN densenet --no_age --y_col 'Gleason_HighLow' --patch_n 200 --spatial_sample_off --n_epoch 100 --lr 0.00001 --optimizer Adam --use_scheduler --balance 0.5 --balance_training --freeze_batchnorm --pooling mean --notes model0 --gpu 2
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent