RSNA STR Pulmonary Embolism Detection

2nd place solution for the RSNA STR Pulmonary Embolism Detection competition on Kaggle.

Solution overview available at: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/193401

Environment

• 16 cores, 64 GB RAM
• 4 24 GB NVIDIA Quadro RTX 6000 GPU
• Python 3.7.7
• Anaconda
• PyTorch 1.6

Setup Python environment

conda create -n rsna-pe python=3.7 pip
pip install -r requirements.txt

Download data

mkdir data 
cd data
kaggle competitions download -c rsna-str-pulmonary-embolism-detection
unzip train.zip

Extract-transform-load

cd src/etl
python 00_extract_metadata.py
python 01_create_cv_splits.py

[Optional] Download trained checkpoints

mkdir checkpoints
cd checkpoints
kaggle datasets download -d https://www.kaggle.com/vaillant/rsna-str-pe-checkpoints
unzip rsna-str-pe-checkpoints.zip

Note: This uses distributed training across 4 GPUs. You may need to edit the commands in each script to match your environment. You will also likely have different checkpoint names if training models from scratch. Please change those as well for each script performing feature extraction/inference.

Train PE feature extractors

bash 0_run_kfold_dist.sh

Extract PE features

bash 1_extract_kfold_dist.sh

Train transformers

bash 2_run_transformer_kfold_dist.sh

Train heart slice classifier

bash 3_run_heart_slices_dist.sh

Obtain OOF predictions (PE/heart slice)

bash 4_predict_features_kfold_dist.sh
bash 5_predict_heart_slices_dist.sh
cd src/etl
python 08_create_train_df_with_pe_and_heart_probas.py

Train time-dependent CNNs

bash 6_run_mk3d_dist.sh

Train 3D RV/LV CNNs

bash 7_run_heart_dist.sh

Extract heart features

bash 8_extract_rvlv.sh

Obtain OOF predictions (PE/RV/LV exam)

bash 9_predict_features_cls_kfold_dist.sh
cd src/etl
python 09_create_proba_dataset.py

Train linear model (refine RV/LV predictions based on PE exam labels)

bash 10_run_linear.py

Inference

Please see public notebook at https://www.kaggle.com/vaillant/rsna-str-pe-submission.