Heart Datathon 2021

Detection Metrics

Dice Coefficient

• \[Dice = \frac{2 |A∩B|}{(|A|+|B|)} = \frac{2 TP}{(2 TP + FP + FN)}\]

Jaccard Index and IoU

• \[JI = \frac{DICE}{2-DICE} = \frac{|A \cap B|}{|A|+|B|-|A\cap B|}\]

Review

EDA

• Train: 1600 (A2C: 800, A4C: 800)
• Validation: 200 (A2C: 100, A4C: 100)
• Too small data
  • Sensitive to apply augmentation
• Ultrasound Image
  • Masks are not strictly following the boundaries of the object.

Worked

• Dehaze Pre-Processing
• CRF Post-Processing
• ReduceLROnPlateau
• HRnet(not used)
  • Limited with model size rule.
  • HRNet + OCR $\approx$ HRNet Unet
• Ensemble(mean)

Not Worked

• more augmentation
  • Hflip, Cutout, CenterCrop
• AdamP
• CosineAnnealingLR

Losses

• Used
  • Tversky
  • DiceLoss
  • LogCosh
  • BCELoss
  • BCE_DICE_Combo (final)
• Still doubt about selecting which loss to use.

Appendix

Reference

Competition info: http://www.hdaidatathon.com/

Competition github: https://github.com/DatathonInfo/H.D.A.I.2021

similar competition: https://www.kaggle.com/c/understanding_cloud_organization/discussion/114093

pytorch-goodies: https://github.com/kevinzakka/pytorch-goodies

Loss Function Library - Keras & PyTorch: https://www.kaggle.com/bigironsphere/loss-function-library-keras-pytorch

torchmetrics: https://torchmetrics.readthedocs.io/en/latest/references/modules.html?highlight=iou#iou

Medical_Img_Seg_and_Enhancement: https://github.com/GeekyGeek3371/Medical_Img_Seg_and_Enhancement

Kaggle Ultrasound Nerve Segmentation: http://fhtagn.net/prog/2016/08/19/kaggle-uns.html

semantic segmentation Loss Survey: https://arxiv.org/pdf/2006.14822.pdf

Deep Learning for Segmentation using an Open Large-Scale Dataset in 2D Echocardiography: https://arxiv.org/pdf/1908.06948.pdf

SegLoss: https://github.com/JunMa11/SegLoss

CRF: https://github.com/lucasb-eyer/pydensecrf