Meeting notes

[JLrumberger]

02/07/23 End of February: present mike the results to decide on overall picture and start writing

Alex updated the pixie notebook so that it only needs cell_table as input and it'll be pip installable soon

My tasks:

1. PR Multi dataset training
2. Run experiments
3. Prepare datasets (re-annotate tonic first and send it to Noah and Candace)

11/29/22

data annotation:

• half way through the third sample which is super dense and takes time
• once done with the third sample I'll calculate metrics with noisy labels vs ground truth and predicted vs ground truth

MSK dataset:

• waiting for the segmentation masks from rami if available
• run mesmer and use thresholds from excel col "Analysis Inputs"
• the images I got are just the first few of roughly 1k images

Code changes ahead:

ModelBuilder

• add multi dataset training to ModelBuilder.prep_data
• add multi dataset validation to ModelBuilder.tensorboard_callbacks
• do a more clever validation split (seeded shuffle before split) and write out all validation fovs

SegmentationTFRecords

• subclass to prepare data when we have a ground truth map available instead of an instance segmentation map, a cell_table and a conversion_matrix
• add tissue type to tf records dataset

[JLrumberger]

Note that for some markers (GLUT1, VIM) the stitched output shows stitching artifacts.