2021-SIIM-COVID19-Detection

6th place solution for SIIM-FISABIO-RSNA COVID-19 Detection Challenge.

Installation

• Run this docker container: nvcr.io/nvidia/pytorch:20.10-py3 (Please see the release note for packages detail if you like to install them by yourself.)
• pip install -r requirements.txt

Data preparation

SIIM covid 19 dataset

• Download the main competition dataset then extract to ./dataset/siim-covid19-detection

cd dataset/siim-covid19-detection
python dcm2png.py
python make_image_symlink.py

External datasets

• Download kaggle_chest_xray_covid19_pneumonia then extract to ./dataset/chest-xray-covid19-pneumonia
• Download kaggle_covidx_cxr2 then extract to ./dataset/covidx-cxr2
• Download kaggle_chest_xray_pneumonia then extract to ./dataset/chest-xray-pneumonia
• Download kaggle_curated_chest_xray_image_dataset_for_covid19 then extract to ./dataset/curated-chest-xray-image-dataset-for-covid19
• Download kaggle_covid19_xray_two_proposed_databases then extract to ./dataset/covid19-xray-two-proposed-databases
• Download kaggle_ricord_covid19_xray_positive_tests then extract to ./dataset/ricord-covid19-xray-positive-tests
• Download CXR14 then extract to ./dataset/NIHCC_CXR14/NIHCC_CXR14
• Download CheXpert then extract to ./dataset/CheXpert-v1.0/
• Download pneumonia_RSNA then extract to ./dataset/rsna-pneumonia-detection-challenge/

cd dataset/chest-xray-covid19-pneumonia && python make_image_symlink.py && cd -
cd dataset/covidx-cxr2 && python unify_file_extensions.py && python make_image_symlink.py && cd -
cd dataset/chest-xray-pneumonia && python make_image_symlink.py && cd -
cd dataset/curated-chest-xray-image-dataset-for-covid19 && python make_image_symlink.py && cd -
cd datasetcovid19-xray-two-proposed-databases && python unify_file_extensions.py && python make_image_symlink.py && cd -
cd dataset/curated-chest-xray-image-dataset-for-covid19 && python make_image_symlink.py && cd -
cd dataset/CheXpert-v1.0 && python make_image_symlink.py && cd -
cd dataset/rsna-pneumonia-detection-challenge && python dcm2png.py && cd - # I did not notice that this preprocessing may not be correct. You can use more correct processing (But may not reproduce my results.)

dataset structure should be ./dataset_structure.txt

Processed label description (The processing of labels is trivial and is done by the author)

For clarity, the SIIM Covid class index is assigned as the following.

• Class 0: Negative for Pneumonia
• Class 1: Typical Appearance
• Class 2: Indeterminate Appearance
• Class 3: Atypical Appearance

For the class indexing of the external dataset, please refer to the comment in metadata.yaml.
If the patient ID is present, we split the dataset by it.

Classification label column

• ACCNO: file name without file extension

Detection label column

• ACCNO: file name without file extension
• labels_for_model: label string where each paranthesis is made up of (x1, y1, x2, y2, 1). The coordinates are corresponding to the image resised to 1024.

Solution summary

• Multi-task classification:
  • Shared-backbone multi-head classifier
• Sharpness-aware minimization (SAM):
  • Foret, Pierre, et al. "Sharpness-aware minimization for efficiently improving generalization." arXiv preprint arXiv:2010.01412 (2020).
• Inverse focal loss:
  • replace (1-p_t) ** gamma with p_t ** gamma in the original focal loss to suppress outlier samples.
• Stochastic weight averaging (SWA):
  • Izmailov, Pavel, et al. "Averaging weights leads to wider optima and better generalization." arXiv preprint arXiv:1803.05407 (2018).
• Attentional-guided context FPN (ACFPN):
  • Cao, Junxu, et al. "Attention-guided context feature pyramid network for object detection." arXiv preprint arXiv:2005.11475 (2020).
• Attentional feature fusion (AFF):
  • Dai, Yimian, et al. "Attentional feature fusion." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2021.
• Fixed feature attention (FFA):
  • We use feature pyramid from the classification model for attention.

We ensemble 5-fold CV study-level and image-level models respectively for our final submission.

Training

Let $SIIM_root denote the root folder of this package.

Ploting learning curve

When training each model, you can use plot_metrics.ipynb to plot the learning curve by simply modifying the model_folder with the work_dir of each training run.

Multi-task classification pre-training

cd $SIIM_root/src/classification
python train_multitask_classification.py
python transfer_multitask_pretrained_backbone.py

The average AUROC of all validation task is 0.9054

SIIM Covid19 study-level training

Note: If you want to train the models without using the pre-trained weights, you can set args_dict['pretrained'] = 'imagenet' in the config file.

cd $SIIM_root/src/classification
python train_classification.py configs/config_cv0.py
python train_classification.py configs/config_cv1.py
python train_classification.py configs/config_cv2.py
python train_classification.py configs/config_cv3.py
python train_classification.py configs/config_cv4.py

	Fold 0	Fold 1	Fold 2	Fold 3	Fold 4
validation mAP of 4 classes	0.5735	0.5905	0.5841	0.5825	0.5942

The final trained study-level model will be saved at $SIIM_root/src/classification/work_dir/covid19_kaggle_train_cv_5_*_run1/model_best.pth where * should be replaced with 0 to 4.

RSNA pneumonia detection pre-training

cd $SIIM_root/src/detection
python train_detection.py configs/config_RSNA_pna_pretraining.py

The validation [email protected] is 0.4352

SIIM Covid19 image-level training

Note: If you want to train the models without using the pre-trained weights, you can set args_dict['fine_tune'] = False in the config file.

cd $SIIM_root/src/detection
python train_detection_with_cls_feats.py configs/config_cv0.py
python train_detection_with_cls_feats.py configs/config_cv1.py
python train_detection_with_cls_feats.py configs/config_cv2.py
python train_detection_with_cls_feats.py configs/config_cv3.py
python train_detection_with_cls_feats.py configs/config_cv4.py

validation [email protected]

	Fold 0	Fold 1	Fold 2	Fold 3	Fold 4
w/o fixed feature attention	0.5504	0.5412	0.5845	0.5446	0.5952
with fixed feature attention	0.5547	0.5456	0.5841	0.5472	0.6047

The final trained image-level model will be saved at $SIIM_root/src/detection/work_dir/covid19_kaggle_train_cv_5_*_run1/model_best.pth where * should be replaced with 0 to 4.

Inference

Use our Kaggle kernel for model inference. Simply replace args_dict['cls']['ckpt_paths'] and args_dict['det']['ckpt_paths'] with your trained model.

The LB scores of ensemble models are shown below. For each TTA, only horizontal flip is performed.

Study-level ensemble + Image-level ensemble	Public LB	Private LB
w/o TTA w/o fixed feature attention	0.633	0.621
w/o TTA	0.633	0.625
study-level TTA	0.634	0.626
image-level TTA	0.635	0.627
both TTA	0.636	0.628

Wish you have fun and save more lives in the world! 🏥