github30 / pfneumonia

Repo for RSNA pneumonia open-source

PFNeumonia

This repo contains the 6th place solution for RSNA Pneumonia Detection Challenge.

All trainings and predictions were done with eight Tesla P100s.

Preparation

• Locate Kaggle RSNA Pneumonia datasets on ./RSNA. The names of subdirectories must be stage_2_train_images and stage_2_test_images. Labels for training dataset must be written in stage_2_train_labels.csv.
• Download ResNet pretrained models from https://github.com/KaimingHe/deep-residual-networks and locate them on ~/.chainer/dataset/pfnet/chainer/models.
• Install requirements by executing pip install -r requirements.txt

How to train

First, shuffle patient IDs by executing the command below. This process is required if you are going to train with a new dataset other than the official Stage2 train or test images.

PYTHONPATH=. python scripts/shuffle_patients.py

Our final submission is an ensemble of 10 models which derive from a 10-fold cross-validation. To run a 10-fold CV, run the command:

./run_cross_validation.sh

How to predict

Predict twice with even (02468) and odd (13579) five models.

./pred02468.sh even.csv
./pred13579.sh odd.csv

Merge these predictions to get the merged CSV.

PYTHONPATH=. python examples/segmentation/ensemble_submission.py even.csv odd.csv -t0.5 -o merged.csv

Finally, adjust threshold of confidence to get the final submission file.

PYTHONPATH=. python scripts/increase_threshold.py -i merged.csv -o final.csv -t 0.3

Links

• Official: https://www.kaggle.com/c/rsna-pneumonia-detection-challenge
• Official Leaderboard: https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/leaderboard