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
- 16 cores, 64 GB RAM
- 4 24 GB NVIDIA Quadro RTX 6000 GPU
- Python 3.7.7
- PyTorch 1.6
Setup Python environment
conda create -n rsna-pe python=3.7 pip
pip install -r requirements.txt
kaggle competitions download -c rsna-str-pulmonary-embolism-detection
[Optional] Download trained checkpoints
kaggle datasets download -d https://www.kaggle.com/vaillant/rsna-str-pe-checkpoints
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
Extract PE features
Train heart slice classifier
Obtain OOF predictions (PE/heart slice)
Train time-dependent CNNs
Train 3D RV/LV CNNs
Extract heart features
Obtain OOF predictions (PE/RV/LV exam)
Train linear model (refine RV/LV predictions based on PE exam labels)
Please see public notebook at https://www.kaggle.com/vaillant/rsna-str-pe-submission.