SIIM-ACR Pneumothorax Segmentation

2nd place solution

Model

segmentation

• unet (se-resnext50, se-resnext101) from [pretrained-models]
• unet (efficientnet-b3, efficientnet-b5) from [EfficientNet-PyTorch]
• deeplabv3 (se-resnext50) from [semantic-segmentation]

classification

• unet (se-resnext50, se-resnext101) from [pretrained-models]
• unet (efficientnet-b3) from [EfficientNet-PyTorch]

Augmentations

Used following transforms from [albumentations]

RESIZE_SIZE = 1024 # or 768
train_transform = albumentations.Compose([
        albumentations.Resize(RESIZE_SIZE, RESIZE_SIZE),
        albumentations.OneOf([
            albumentations.RandomGamma(gamma_limit=(60, 120), p=0.9),
            albumentations.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.9),
            albumentations.CLAHE(clip_limit=4.0, tile_grid_size=(4, 4), p=0.9),
        ]),
        albumentations.OneOf([
            albumentations.Blur(blur_limit=4, p=1),
            albumentations.MotionBlur(blur_limit=4, p=1),
            albumentations.MedianBlur(blur_limit=4, p=1)
        ], p=0.5),
        albumentations.HorizontalFlip(p=0.5),
        albumentations.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.2, rotate_limit=20,
                                        interpolation=cv2.INTER_LINEAR, border_mode=cv2.BORDER_CONSTANT, p=1),
        albumentations.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0)
    ])

Loss Function

classification

• Cls loss: BCE + focal loss
• Seg loss: BCE

segmentation

• Seg loss: dice loss

Training Method

image size

Train image size: 768 or 1024 test image size: 768 or 1024

stochastic weight averaging

[swa]

ensemble

• classification: stacking
• segmentation: average

pseudo labels

We predict masks on chexpert dataset using trained model, and then add these pseudo labels(about 1000) to the network and fine-tune model. There was no significant improvement.

File structure

├── configs
│   ├── seg_path_configs.json
├── data              
│   ├── chexpert_data
│   │   ├── chexpert_img
│   │   ├── chexpert_pesudo_label
│   ├── cls_fold_5_all_images
│   │   ├── 5_fold_file
│   ├── cls_fold_5_all_images_p_label_chexpert
│   │   ├── 5_fold_file
|   ├── competition_data
│   │   ├── test_png
│   │   ├── train_png
│   │   ├── sample_submission.csv
│   │   ├── train-rle.csv
│   └── det_fold_5
│   │   ├── 5_fold_file
├── models_snapshot
├── result
├── semantic_segmentation
│   ├── deeplab_model
├── src_unet_cls
│   ├── classification_model_code
├── src_unet_seg
│   ├── segmentation_model_code
├── README.md
└── requirements.txt

Install

pip install -r requirements.txt

How to run code

segmentation model training:

cd src_unet_seg

python B_train_model.py -backbone unet_se50 -img_size 1024 -tbs 4 -vbs 2 -use_chex 1 -save_path seg_unet_se50_1024
python B_train_model.py -backbone unet_se101 -img_size 1024 -tbs 4 -vbs 2 -use_chex 0 -save_path seg_unet_se101_1024
python A_train_model.py -backbone unet_ef3 -img_size 1024 -tbs 6 -vbs 2 -use_chex 1 -save_path seg_unet_ef3_1024
python A_train_model.py -backbone unet_ef5 -img_size 768 -tbs 4 -vbs 2 -use_chex 0 -save_path seg_unet_ef5_768
python A_train_model.py -backbone deeplab_se50 -img_size 1024 -tbs 4 -vbs 2 -use_chex 1 -save_path seg_deeplab_se50_1024

stochastic weight averaging

python swa_models.py -i seg_unet_se50_1024/ -o ./se50_swa_{}.pth.tar -e0 43 -e1 34 -e2 39 --model_num unet_se50 --batch-size 4
python swa_models.py -i seg_unet_se101_1024/ -o ./se101_swa_{}.pth.tar -e0 43 -e1 34 -e2 39 --model_num unet_se101 --batch-size 4
python swa_models.py -i seg_unet_ef3_1024/ -o ./ef3_swa_{}.pth.tar -e0 43 -e1 34 -e2 39 --model_num unet_ef3 --batch-size 6
python swa_models.py -i seg_unet_ef5_768/ -o ./ef5_swa_{}.pth.tar -e0 43 -e1 34 -e2 39 --model_num unet_ef5 --batch-size 4
python swa_models.py -i seg_deeplab_se50_1024/ -o ./deep_se50_swa_{}.pth.tar -e0 43 -e1 34 -e2 39 --model_num deeplab_se50 --batch-size 4

Inference:

python predict_768.py 
python predict_1024.py

Ensemble:

python ensemble_5_model.py 

classification model training:

cd src_unet_cls
python train_model.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 16 -vbs 8 -save_path diy_model_se_resnext50_32x4d_768_normal
python train_model.py -backbone diy_model_se_resnext50_32x4d -img_size 1024 -tbs 8 -vbs 4 -save_path diy_model_se_resnext50_32x4d_1024_normal
python train_model.py -backbone EfficientNet_3_unet -img_size 1024 -tbs 16 -vbs 8 -save_path EfficientNet_3_unet_1024_normal
python train_model.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 16 -vbs 8 -save_path diy_model_se_resnext50_32x4d_768_add_chexpert

stochastic weight averaging

python swa.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 32 -vbs 8 -cp diy_model_se_resnext50_32x4d_768_normal
python swa.py -backbone diy_model_se_resnext50_32x4d -img_size 1024 -tbs 32 -vbs 8 -cp diy_model_se_resnext50_32x4d_1024_normal
python swa.py -backbone EfficientNet_3_unet -img_size 1024 -tbs 32 -vbs 8 -cp EfficientNet_3_unet_768_normal
python swa.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 32 -vbs 8 -cp diy_model_se_resnext50_32x4d_768_add_chexpert

Inference:

python predict.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 4 -vbs 4 -spth diy_model_se_resnext50_32x4d_768_normal
python predict.py -backbone diy_model_se_resnext50_32x4d -img_size 1024 -tbs 4 -vbs 4 -spth diy_model_se_resnext50_32x4d_1024_normal
python predict.py -backbone EfficientNet_3_unet -img_size 1024 -tbs 4 -vbs 4 -spth EfficientNet_3_unet_1024_normal
python predict.py -backbone diy_model_se_resnext50_32x4d -img_size 768 -tbs 4 -vbs 4 -spth diy_model_se_resnext50_32x4d_768_add_chexpert

Submit:

# stacking
python stacking.py

Leaderboard:

• stage1: 0.8883
• stage2: 0.8665