Incorporate insights from Cellpose/Carsen

[rhoadesScholar]

On distance masking:

• Make distance mask for whole crop (save alongside)
• Precompute targets?
• Without distance outside object, no distance mask is needed

Sparse data:

• Consider using targets:
  • boundary pixels (BCE)
  • flows (MSE)
  • probability (BCE)
  • center prediction

Loss weighting:

• Sample crops based on organelles
  • Balance sampling instead of balancing loss weights
  • Changing loss weights changes gradient magnitudes constantly which is difficult for Adam optimizer

BCE for binary class predictions --> MSE doesn’t penalize false positives

• Weight BCE compared to MSE

(Center of object (for flow calculation) is calculated as object-pixel rounded center of mass)
Try 4 convolutions per block in UNet (2x2 —> see CellPose 3)

[davidackerman]

include Cellpose Unet for #9