Cellpose 2

Overview

https://www.nature.com/articles/s41592-022-01663-4.pdf

Basic idea:

• Finetune Cellpose with just a few samples
• And get pretty good results

Performance:

• Clean Tissuenet AP@0.5 = 0.85 performance seen in Cellpose 3 seems to be this fine-tuned model

Interesting facts about Tissuenet

• Another class of interactive approaches known as ‘human-in-the-loop’ start with a small amount of user-segmented data to train an initial, imperfect model. The imperfect model is applied to other images, and the results are corrected by the user. This is the strategy used to annotate the TissueNet dataset, which in total took two human years of crowdsourced work for 14 image categories
• The outlines in the Cellpose dataset were drawn to include the entire cytoplasm of each cell, often biased toward the exterior of the cell (Fig. a(iii)). Some TissueNet categories also included the entire cytoplasm (Fig. 1b(i)), but others excluded portions of the cytoplasm (Fig. 1b(ii),(iii)) or even focused exclusively on the nucleus (Fig. 1b(iv)).
• (...) nuclei were not segmented in the Cellpose dataset if they were missing a cytoplasm or membrane label (Fig. 1a(i)), but they were always labeled in the TissueNet dataset
• The TissueNet dataset contained 13 image categories with at least ten training images each