breast-cancer-identification

Identifying high-risk breast cancer using digital pathology images

Required folders and files

checkpoints/
    # model checkpoints
csv_dir/
    # mapping relationship
datasets/
    # python pickle (tensor) datasets
models/
    # swin transformer model class
split_biopsies.ipynb
prepare_datasets.ipynb
train_pipeline_resnet50.ipynb
train_pipeline_swin.py
test_ensemble.ipynb
submit.ipynb
license.txt

Usage

Split train/test/holdout set

Run split_biopsies.ipynb

It will export several CSV files in csv_dir folder that record the mapping relationship of downsampled slices information and labels.

Export Python Pickle format train/test/holdout set

Run prepare_datasets.ipynb

The script exports a dictionary for each train/test/holdout set in datasets/ folder. Like

pd.to_pickle({'x': holdout_x_list, 'y': holdout_y_list, 'id': holdout_biopsy_id_list}, f'./datasets/holdout.pkl')

• x: slice tensor. Croped and Normalized to 224x224x3 resolution.
• y: label. {0, 1, 2, 3, 4}
• id: BiopsyID, the slice image belongs to.

Train Deep learning models

• Run train_pipeline_resnet50.ipynb for RestNet-50 model
• Run train_pipeline_swin.py for Swin-Large model

Above two scripts will save model parameters in checkpoints/ folder.

Ensemble two models' outputs for holdout set

Run test_ensemble.ipynb

The script load ResNet and Swin models' parameters and outputs the prediction logits for holdout set.

Then we calibrate and ensemble the two outputs (take average scores) and export the final prediction results for expected holdout set.

Submit results

• Specify prediction result CSV, Run submit.ipynb

(Our team's final best result is in submit_1220.csv)