Identify melanoma in lesion images

Team Members

Our Approach

We used EfficientNet [B0-B6], Resnest,Resnext, with Sizes 192x192 256x256 384x384 512x512 768x768 384x512[HxW]

Summary

What Worked for Us

  • Heavy TTA (X20)
  • Cutmix
  • Coarse dropout
  • SWA(Stochastic Weight Averaging)
  • Loss-Label Smoothing, BCE
  • Optimizers - AdamW, Adam
  • 2018, 2020 and malignant datasets
  • 5 checkpoints' prediction averaging(stabalised our model's predictions)
  • some models were trained with different height width ratios

What didn't Work for Us

  • Loss functions-Focal loss, dice loss
  • Optimizer- Ranger
  • Hair removal/addition
  • Pseudo labelling
  • 2019 dataset
  • Preprocessing techniques from Aptos Competition
  • Progressive learning

Ensembling techniques

  • Weighted average
  • Power Average
  • Minmax ensemble(didn't help)

Our final Submission

15+ pytorch model (with context) and 15+ tf models (without context) - 0.9627 (public LB) 0.9470 (private LB) 0.9618 (CV) It could have achieved us 6th rank on the private leader board of the competition.