Skip to content

Correcting Annotations #62

@yellenlab

Description

@yellenlab

Hi @eric-czech ,

I uploaded some files to the google storage bucket showing both good and bad annotations. There are some problems with the classifiers for cells, digits, and (I believe) markers, and I included a readme file about each type of incorrect annotation in the specific folders. In brief, there are some images where the apartment is left shifted, which I believe is a marker mis-identification. These typically occur at the edge of the chip, where the pattern changes. There are other errors in digit classifiers, which might be because these new numbers are slightly smaller than the numbers on previous chips (I'm really running out of space for these numbers with the new chip designs). Finally, the cell identifier mostly works when there is less than 25 cells in the apartments, but starts to misbehave when it goes past 50% confluency. I'm worried that the growth rates are being underpredicted. Interestingly, the annotations do not seem to be affected by changes in lighting, e.g., at the edge of the camera vs. the center. And when the focus is good, there doesn't seem to be a specific reason why it annotates some images correctly but not others. Perhaps the cell diameter requirement is too stringent?

After you have a chance to look through these annotations, perhaps we can form a plan to re-train the models. Thanks for your help!

Ben

@jmotschman

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions