This is the DeepLearning Approach of PANDA Challenge @Kaggle

• Model:

  • EfficientNet - B0
  • Pre-trained From ImageNet

• Label & Loss:

  • Classification Bin Label, Cross-Entropy Loss

• Ensemble:

  • 4 - fold Ensemble

• Noisy Label Cleaning:

  • Direct Exclusion if abs(pred - true) >= 1.5

• Final Performance & Ranking:

  • Weighted Quadratic Kappa: 0.93123 Private Score
  • Accuracy: ~ 0.66
  • Ranking: ~ 1%