The CIN algorithm is a weakly supervised learning method based on Densenet [1] architecture which is pre-trained using ImageNet images [2].

The algorithm classifies genome CIN+/CIN- based on H&E histology whole slide images (WSI)

To run the CIN framework please follow these steps:

Install the following softwares used in this project:

TensorFlow2: Follow the instruction from here: https://www.tensorflow.org/install/

OpenSlide： Python version. See more details here: https://openslide.org/

HistomicsTK: Used for color normalization. Follow the instrction from here to install: https://digitalslidearchive.github.io/HistomicsTK/installation.html

See codes for genome CIN score calculation at '/Codes/Genome CIN/CIN_SCORE_CALCULATION.R'

Follow the steps below to tile WSI from svs file and crop WSI into multiple patches of each patient. See codes under the folder '/Codes/Preprocessing'

Use the function of tiling_wsi from TilingWSI.py to get the best WSI from raw svs file. The default setting will get WSI on 2.5x magnification with dimension of 2048px2048p. Through setting tilesize and overlap can change the step size of sliding windows.

You need to define two list objects ahead: filepaths and samplenames. Also set the argument of tile_dir for the location to store WSI.

filepaths : list of svs file paths

samplenames : list of sample names with the same order of filepaths. For example: 'TCGA-A1-A0SE-01Z-00-DX1.04B09232-C6C4-46EF-AA2C-41D078D0A80A'

tile_dir : location to store WSI. For example: '/Image/WSI/'

Use the function of WSIcropping from WSIcropping.py to crop input images into 8x8 nonoverlapping grids and save patches into target location. You can set up a tissue percentage threshold for QC. This function outputs patches and organize them in the structure of one folder per patient. Please refer to '/Image/Patch/'

inputdir : input directory. example: /Image/WSI/

targetdir : target directory. example: /Image/Patch/

pct : tissue percentage threshold. Default is 80%

This step used transfer learning to extract patch features and aggregate into patient level features. It will return a numpy array with dimension of n * 1024. n denotes number of patients.

See codes at '/Codes/Feature Extraction/FeatureExtraction.py'

First, set lab reference for color normalization.

ref_img_path='/Image/WSI/IMAGE_NAME.jpg'

Then, use the function of Densenet121_extractor to extract patient level features from cropped patches generated by last step stored in '/Image/Patch'. You will need to organize a pandas DataFrame object or a .csv file (WSI_df) to store the image information including patch location and CIN score that will be used as labels after. A sample dataframe can be found in '/File/WSI_df.csv'. This function will automatically conduct color normalization.

WSI_df : a pandas dataframe. All patients' patches' in the dataframe will be extracted and aggregated. It should contain a column with name of 'Barcode_Path' to store the paths of patients' folder. All patch images are stored under this folder. example: /Image/Patch/TCGA-3C-AALJ

target : target path of patient level features to be stored. example: '/Bottleneck_Features/features_densenet121.npy'

This is the last step to train your final model. Make sure you have loaded WSI_df (summarise DataFrame) and features (bottleneck features array) before training top MLP.

#load summarise dataframe
WSI_df=pd.read_csv('/File/WSI_df.csv', sep=',',index_col=0).reset_index(drop=True)

#load bottleneck features array
features=np.load('/Bottleneck_Features/features_densenet121.npy')

Run codes at '/Codes/MLP/MLP.py' and use function of MLP_train in MLP.py to train top MLP layers. The function will automatically save the best model along with ID_train, ID_test, y_train, y_test, X_train and X_test for furthur model evaluations.

You need to set up the following arguments:

WSI_df: a dataframe of summarise file, the same one used in previous steps. make sure you have columns indicating label and patient ID

features: patient level feature array.

Label, ID: column names in WSI_df. example: ID='Barcode',Label='Cin_Label'

testratio: percentage of test dataset, default is 0.15. You don't need to set validation dataset because the function will automatically split training and validation dataset as ratio of 0.85/0.15 after setting aside of hold-out test set.

seed: used to split training and test dataset. default is 1001

Model_Name: name to store the model, example: 'Model.h5'

Learning parameters: layer_num,nodes_num_1,nodes_num_2,dropout,lr you can set up to 2 hidden layers. Too many hidden layers is not likely to have good performance by experiments. Use layer_num to set number of hidden layers.Use nodes_num_1 and nodes_num_2 to set nodes number of two hidden layers. Only set nodes_num_2 when layer_num=2. You can set up dropout and learning rate. default setting: dropout=0,lr=0.00001

After tuning the hyperparameters from the last step and getting your final model. You can load your saved model and make predictions and evaluate your model.

from tensorflow.keras.models import load_model
import sklearn.metrics as metrics
#make predictions
model=load_model('Model.h5')
y_pred=model.predict(X_test)
#evaluate model accuracy:
metrics.accuracy_score(y_true=y_test, y_pred=y_pred.round())

[1] Huang, G. et al. (2017). Densely Connected Convolutional Networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). DOI :10.1109/CVPR.2017.243.

[2] Deng, J. et al. (2009). ImageNet: A Large-Scale Hierarchical Image Database. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). https://www.image-net.org/papers/imagenet_cvpr09.bib