3rd place solution for the 2018 RSNA Pneumonia Detection Challenge

by Phillip Cheng, MD MS

Hello! Below is an outline of how I prepared the data, trained my models, and predicted bounding boxes for the RSNA Pneumonia Detection Challenge. Further details are in the model documentation submitted to Kaggle. For questions please contact me at phillip.cheng@med.usc.edu

Directory structure

• data: contains raw and processed data from Kaggle
• keras-retinanet: modified version of keras-retinanet
• models: selected resnet-50 and resnet-101 model snapshots used for the prediction submissions
  • Note: the .h5 model files are too large to be stored in the Github repository, so you will need to manually download the .h5 files from the release and place them in this folder.
• snapshots: folder for saved snapshots after each epoch of training
• submissions: folder to hold output from predict.py

My hardware

• Dell Precision T5500 with 3.07 GHz Intel Xeon X5675 CPU
• 12 GB NVIDIA Titan Xp GPU
• 512 GB HDD
• 16 GB DDR3 RAM

Software

(python packages are detailed separately in requirements.txt):

• Ubuntu 18.04 LTS
• Python 3.6.6
• CUDA 9.0
• cuDNN 7.0.5
• NVIDIA driver v.390

Data and keras-retinanet setup

(assumes the Kaggle API is installed and configured)

Run the following shell commands from the top level directory

cd data
kaggle competitions download -c rsna-pneumonia-detection-challenge
unzip stage_2_test_images.zip -d stage_2_test_images
unzip stage_2_train_images.zip -d stage_2_train_images
cd ../keras-retinanet
python setup.py build_ext --inplace
cd ..
python prepare_data.py

Note that settings.json is configured for creating training and validation sets from the Stage 1 training labels, which are what I used to train the models for the competition. I did not perform any training on the Stage 2 training set, which included Stage 1 test images.

Model training

There are two RetinaNets used in my solution, which are trained by the following scripts.

./train50.sh
./train101.sh

Snapshots after each epoch of training are saved in snapshots. For the competition, I manually selected the best snapshots from several runs based on the precision score and maximum Youden index reported during training.

Bounding box prediction

python predict.py

Output .csv is saved to submissions folder