Kaggle-SIIM

PyTorch implementation for Kaggle SIIM-ACR Pneumothorax Segmentation Challenge

42th / 1475 (top 3%)

Leaderboard scores (the mean of the Dice coefficients for each image in the test set):

• Stage-1
  • Without using leak: 0.8696
  • With using leak: 0.8742
• Stage-2
  • Without retraining: 0.8502
  • With retraining: 0.8512

Method

• Small models
  • UNet [1] (ResNet18, 48 MB) for segmentation (10-fold)
  • BiSeNet [2] (ResNet18, 53 MB) for segmentation (10-fold) and classification (5-fold)
  • With auxiliary heads and dual attention [3] in both models
• Losses
  • Weighted binary cross-entropy
  • Symmetric Lovász-hinge [4] with margin
  • Additional binary cross-entropy for classification
• Optimizer: RAdam [5]
• Scheduler: CosineAnnealingLR (1 cycle)
• Augmentation: horizontal flip, random crop and resize
• Resolution: 768x768
• Segmentation part
  • Batch size: 6 * 4 (with gradient accumulation)
  • Using only pneumothorax data
• Classification part
  • Batch size: 6 * 8 (with gradient accumulation)
  • Using pneumothorax:non-pneumothorax = 50%:50% in each batch
  • Using dilated masks (128x128 kernel) to roughly locate pneumothorax region
• Without using external data and TTA

Usage

Modify the appropriate dataset path in config.json

Training for each model

• bash run_train_seg-bisenet.sh
• bash run_train_seg-unet.sh
• bash run_train_cls-bisenet.sh

Inference for each model

• bash run_test_seg-bisenet.sh
• bash run_test_seg-unet.sh
• bash run_test_cls-bisenet.sh

Ensemble all the classification and segmentation models

• Submission1: non_empty_ratio = 0.17
  1. bash run_merge_cls1.sh
  2. bash run_merge_seg1.sh
• Submission2: non_empty_ratio = 0.18
  1. bash run_merge_cls2.sh
  2. bash run_merge_seg2.sh

References

[1] U-Net: Convolutional Networks for Biomedical Image Segmentation (U-Net)

[2] BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation

[3] Dual Attention Network for Scene Segmentation

[4] The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

[5] On the Variance of the Adaptive Learning Rate and Beyond