We used iterative-stratification with 5 folds (https://github.com/trent-b/iterative-stratification) stratified by study level classes and number of boxes.

We used efficientnet-b7, v2s, v2m, and v2l with aux branches after 3 different blocks. The models were trained on 3 folds and on different image resolutions (512, 640, 768) to produce 14 classifiers.

We used simple averaging for ensembling the models. LB mAP for study level was ~62.3-62.4

We used mmdetection library to train detectoRS50, universeNet50, and universeNet101. detectoRS50 and universeNet50 were trained on one fold, and universeNet101 was trained on each fold + pseudo labels for public data using the universeNet50 model.

WBF did not work for us, so we decided to use NMW for ensembling from https://github.com/ZFTurbo/Weighted-Boxes-Fusion.

TTA: HorizontalFlip for detectoRS and multi-scale TTA for all universeNet models on [(640, 640), (800, 800)].

Binary classifiers were trained in the same manner as study level models, 3 fold ensemble was used.

Our augmentations include HorizontalFlip, RandomCrop (for study level), ShiftScaleRotate, CLAHE, RandomGamma, Cutout from albumentations library (https://albumentations.ai/ ).

1. Change the paths to your dcm pickle files in the config files for universeNet and detectoRS.
2. The detector can be trained inside the folder UniverseNet as follows:

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM}

1. To train classifier run train_normal.py. To change the data root, the dcm pickle files, and the fold number use --data_root, --dcm_folds_train, --dcm_folds_val, and --fold. The auxilary branches can be changed as well via --aux. For example:

CUDA_VISIBLE_DEVICES=0,2,3,4 python train_normal.py --scheduler plateau --fold 0 --aux 5678 --v2_size m --batch_size 32 --data_root SIIM-FISABIO-RSNA-COVID-19-Detection --dcm_folds_train /dcm_folds/data_train_dcm_fold0.pickle --dcm_folds_val /dcm_folds/data_val_dcm_fold0.pickle