MURA bone X-rays

Networks (Results)

To beat : kappa 0.843

• pretrained resnet34 + adamW + weight decay + mixup + oevrsampling
  • kappa : 56.8% (not much training)
• pretrained resnet34 + adamW + weight decay + progressive resizing (64-> 128->256->128)
  • kappa : 64.7%
• xresnet34 (not pretrained)+ adamW + weight decay + progressive resizing (64-> 128->256->128)
  • kappa : 66.9 %
  • wayy faster without pre trained
• pretrained densenet + adamW + weight decay + progressive resizing (128->256->128)
  • kappa : %

Conclusion

• mixup and over sampling isnt always useful so dont use it all the time
• xresnet >>> resnet