1 of 12

Image analysis competitions

25.11.2020

2 of 12

Agenda

3 of 12

Cassava Leaf Disease Classification

  • Evaluate crop healthiness
  • Simple accuracy
  • EDA

4 of 12

HuBMAP: Hacking the Kidney

  • Segment glomeruli in kidney tissue images
  • Dice coefficient
  • Huge images (HUGE)
  • EDA

5 of 12

Augmentations

  • A way to make data more diverse, thus making network more robust
  • Active area of research for image classification, not so for segmentation
  • Most people use rotations, crops and flips, they do not help much
  • Very little results for medical images, but with clear improvement (e.g. for MRI glioma segmentation)

6 of 12

Augmentations: overview

  • Input transformation
    • Rotate, translate, flip, scale, affine, etc.
  • Input masking
    • Coarse dropout
    • Hide-and-Seek
    • CutOut
    • GridMask
  • Input mixing
    • MixUp
    • CutMix
    • GridMix
  • Combined
    • AugMix

7 of 12

Input transformation

Easy, fast, available out of the box, widely used

8 of 12

Input masking

Ours = GridMask (figure from their paper), currently the best

9 of 12

Input mixing

  • Generally better than masking (Cutout and CutMix have the same sampling)
  • Labels are weighted according to the image proportions (beta distribution)

10 of 12

Input mixing: GridMix

  • My idea to blend CutMix+GridMask (because mix > mask)
  • Helps for cell image classification, needs to be tested (+parameters)

11 of 12

Combined: AugMix

Combine different augmentations in one image

12 of 12

Augmentations for segmentation

  • Masking: model well borders (might not be too useful given other effects)
  • Mixing:
    • CutMix/GridMix: emulate mixed cultures, could help with 7CL
    • MixUp: emulate two-layered culture, will increase robustness
  • Transformation: more data patterns

Best methods are very fresh, not tested for medical images