PAvel meeting: Nov 29, 2019

What we have achieved 

• Sparse identification of correspondences helps in recovering the transformation
• Once we have the correct transformation, we can associate the nuclei using Hungarian Matching

• Next step is segmentation of the membrane channel in the live, time lapse image to associate gene expression with the correct cell
• By pass, specification of few correspondences

Next few challenges

What I am working on the side

1. GenSeg: An alternative to StarDist that works better for non-convex shapes

2. DL-based multiview fusion: on KellerLab dataset, Johannes OpenSpim dataset

3. SimCare: N2V on SIM data

4. Marko Brankatschk data, Jana data ...

Back to Platynereis ...

1. Graph Matching instead of Point Cloud Matching

So far we have addressed this as a problem of point cloud matching. But does graph matching with binary costs provide a better framework? (for example, 'Active Graph Matching' from  Kainmueller et al, 2014 suggests that including binary costs helps in better matching accuracy)

2.Learning Graph Matching

Caetano et al suggest that learning allows taking into conditions under which the two sets of data have been procured. Should work for partial and complete annotation sets.

Back to Platynereis ...

Dense Matching

For learning, we should have multiple specimens of a stereotypic individual which have been segmented and annotated. Such data exists for C. elegans. (Mean and variance for all nuclei is available - Proposal for Experiment One!)

Can we also get access to the raw images? Florian suggested that PAvel should write to the authors.

Sparse Matching

For Platynereis, we should look at similar time points for the three individuals and perform a similar analysis.

Platynereis: Ch2: Membrane Segmentation

Performing Voronoi Tesselation did not work so well

Performing Nearest Neighbour allocation to cell nuclei

works better

Using StarDist to segment cell membranes is difficult because of the excessive noise in the membrane channel

PT mentioned that focus on generating the lineage. Use tesselation. maybe skip the macromeres. Show him the results during the hackathon. PT specified - close the loop.

FJ mentioned that genseg should be presented on Friday. Also there is paper reading on Friday. Also spoke to TP to join.