This is the 5th place code for the RSNA-STR 2020 Pulmonary Embolism Detection challenge.
A solution overview can be found here. You can also check the slides in the docs/ folder.
The solution video is here.
A single fold can be trained in 8 hours, and will achieve on the leaderboard 0.162 Private and 0.167 public. The below steps will train on 3 folds which results in a leaderboard score of 0.155 Private and 0.158 public.

The solution was tested on a NVIDIA DGX A100 with Ubuntu 18.04.4 LTS as operating system.
The full solution runs end to end in approx 24 hours on a single A100 card. This card has 40GB GPU memory, if you have less memory, you can use lower batchsize, but increase accumulation to simulate the same large batchsize used originally. It is important to keep the batchsize large due to the number of targets and imbalanced labels.

You are provided with a docker file. You can run without this by ensuring the necessary packages are available. The docker environment can be set up like below (please ensure docker is installed).

cd docker
docker build -t rsnav04 -f DockerFile.docker .
cd ..
# Change the device to your gpu device number
docker run -itd  --ipc=host --name RSNA_CONTAINERV04 -v $PWD:/mount --gpus '"device=5"' --rm rsnav04:latest 
docker attach RSNA_CONTAINERV04 # To enter the docker
cd ../mount # from inside the docker container - this is the home directory for the package

To leave the docker, if you want to keep it running press ctrl+p and ctrl+q in sequence. To exit and close the docker conatiner, type exit.

At minimum, the training and inference steps should be run from inside the docker container.

Download the competition dataset and place in folder data, so we have a file data/rsna-str-pulmonary-embolism-detection.zip. Do not unzip it.

For preprocessing we load the dicom file and window over each CT scan in the dicom, using the below windows,

• RED channel / LUNG window / level=-600, width=1500
• GREEN channel / PE window / level=100, width=700
• BLUE channel / MEDIASTINAL window / level=40, width=400

Each channel is stored in a differnt channel within a jpeg file to easy loading to models.
If you would like more detail on windowing, check out this description.

Run the following to create jpegs from dicoms.

nohup sh bin/run_01_prepare_data.sh &> logs/preprocess_run.out &

This will run in the background for around 3 hours depending on your environment. There are a total of just under 2 million images to be extracted. You can check how many files were processed while the background process is running :
cat logs/preprocess_run.out | wc -l

This script runs a lot faster due to multithreading, however it does fail sometimes on some images. We run this a few times to pick up failed files. At the end, check the tail of the log to ensure no images are left.

$ cat logs/preprocess_run.out | tail -4
2020-10-31 00:43:13,676 - Preprocess - INFO - Success train/13ef0b464626/8e86fda638ec/346da83668c7.dcm
2020-10-31 00:43:13,680 - Preprocess - INFO - Success train/fbe76b0deffe/6b26e3296bf3/81c2a6a8187a.dcm
2020-10-31 00:43:13,681 - Preprocess - INFO - Success train/f673adb91a15/578d5d64aae8/821cc62d3d2c.dcm
2020-10-31 00:43:43,020 - Preprocess - INFO - There are 0 unprocessed files

You can see the logs for the original jobs in the logs/original/ for reference.

Run the following to train 3 of 5 folds. The folds are provided in the data/ folder.

nohup sh bin/run_02_image_features.sh &> logs/train_images_run01.out &

To view progress run cat logs/train_images_run01.out.
Weights will be saved to folder weights/. To extract features run the below.

nohup sh bin/run_03_extract_image_features.sh &> logs/infer_images_run01.out &

Run the following to train 3 of 5 folds. The folds are provided in the data/ folder.

nohup bin/run_04_sequence_model.sh &> logs/train_sequences_run01.out &

To view progress run cat logs/train_images_run01.out.
Weights will be saved to folder weights/. Model configs are stored on configs/.

The trained weights can be found in the following dataset: rsnastr2020weights.
The submitted kaggle kernel with prediction code can be seen in this script kaggle.com/darraghdog/rsnastr2020-prediction.