Cellpose 3

Overview

Basic idea:

• Use fixed Cellpose model
• train another model to denoise Poisson noise
• The denoised image shoudld have good segmentation performance on cellpose

Losses:

• Reconstruction loss - denoised vs clean
• Perceptual loss - feature similarity in cellpose encodings
• Segmentation loss - pred vs target flows of the fixed cellpose model

Performance:

• Doesn't really compare with other approaches, but shows qualitative and quantitative results
• Clean Tissuenet AP@0.5 = 0.85