SIIM-FISABIO-RSNA COVID-19 Detection

Identify and localize COVID-19 abnormalities on chest radiographs

Below you can find a outline of how to reproduce our solution.

1. Requirements:

1.1 Hardware:

• GPU : 4x Tesla V100
• GPU Memory : 4x32 GiB
• CUDA Version : 11.0
• Driver Version : 450.119.04
• CPU RAM : 16 GiB
• DISK : 2 TB

1.2 Libraries:

• python-gdcm==3.0.9.1
• pydicom==2.1.2
• joblib==1.0.1
• tensorflow==2.4.1
• torch==1.7.0
• torchvision==0.8.1
• numpy==1.19.5
• pandas==1.2.4
• matplotlib==3.4.2
• opencv-python==4.5.2.54
• opencv-python-headless==4.5.2.54
• Pillow==8.2.0
• PyYAML>=5.3.1
• scipy==1.6.3
• tqdm==4.61.1
• tensorboard==2.4.1
• seaborn==0.11.1
• ensemble_boxes==1.0.6
• albumentations==1.0.1
• thop==0.0.31.post2005241907
• Cython==0.29.23
• pycocotools==2.0
• addict==2.4.0
• timm==0.4.12
• efficientnet==1.1.1

2. External Packages

External Packages with version number are listed on requirements.txt

! pip install -qr requirements.txt

3. Data Preparation

3.1 Description

• Download competition data and extract it to ./data/siim-covid19-detection
• Download chexpert dataset and extract to ./data/chexpert
• Download RSNA competion data and extract it to ./data/rsna-pneumonia-detection-challenge
• Download Ricord dataset and extract it to ./data/ricord

After this ./data directory should look something like this.

.
├── data
│   ├── chexpert
│   │   ├── train
│   │   ├── train.csv
│   │   ├── valid
│   │   └── valid.csv
│   ├── ricord
│   │   ├── MIDRC-RICORD
│   │   └── MIDRC-RICORD-meta.csv
│   ├── rsna-pneumonia-detection-challenge
│   │   ├── GCP Credits Request Link - RSNA.txt
│   │   ├── stage_2_detailed_class_info.csv
│   │   ├── stage_2_sample_submission.csv
│   │   ├── stage_2_test_images
│   │   ├── stage_2_train_images
│   │   └── stage_2_train_labels.csv
│   └── siim-covid19-detection
│       ├── sample_submission.csv
│       ├── test
│       ├── train
│       ├── train_image_level.csv
│       └── train_study_level.csv

In case you are wondering to have a look at complete directory structure, you can see it in data_structure.txt

After this run prepare_data.py. It does the following

• Read training data from RAW_DATA_DIR (specified in SETTINGS.json)
• Run any preprocessing steps
• Save the cleaned data to CLEAN_DATA_DIR (specified in SETTINGS.json)

3.2 Script

prepare_data.py

• --img-size image size in which we want our cleaned to to be
• --debug if given 1, it will only process 100 images

! python prepare_data.py 

4. Training

4.1 Description

Simply run the train.py script. It does the following

• Read training data from TRAIN_DATA_CLEAN_PATH (specified in SETTINGS.json)
• Pretrains classification and detection backbones in chexpert data.
• Finetunes them on competition data and external data.
• Save model to MODEL_DIR (specified in SETTINGS.json)

4.2 Script

train. py

• --settings-path path to SETTINGS.json. Default value uses the correct path.
• --clsbs-path path to json file containing necessary batch sizes for different classification models. Default value uses the correct path.
• --detbs-path path to json file containing necessary batch sizes for different detection models. Default value uses the correct path.
• --debug will process only 100 images

! python train.py

5. Prediction

5.1 Description

Before proceeding download this already trained checkpoints and unzip them into the path specified in CHECKPOINT_DIR in SETTINGS.json.

./checkpoints then should look like

.
├── checkpoints
│   ├── 2cls
│   ├── 4cls
│   ├── det

For predicting on test data run predict.py. It does the following

• Read test data from TEST_DATA_CLEAN_PATH (specified in SETTINGS.json)
• Loads models from MODEL_DIR(specified in SETTINGS.json) when everything is trained from scratch or CHECKPOINT_DIR(specified in SETTINGS.json) when predicting from our previously trained checkpoints.
• Use our models to make predictions on new samples
• Save our predictions to SUBMISSION_DIR (specified in SETTINGS.json)

5.2 Script

predict. py

• --mode if used "full", then it will use the weights saved in MODEL_DIR (which was saved after training from scratch) and when used "fast" it will use the weights saved in CHECKPOINT_DIR (already trained checkpoints)
• --debug if given 1, it will infer on only first 100 images

!python predict.py --mode "fast"

or

! python predict.py --mode "full"

Acknowledgement ❤️

• Weights & Biases for tracking training.
• efficientnet for efficientnet model.

Prepared By Team RTX 4090