search

qupath / qupath-extension-instanseg

The official QuPath extension for InstanSeg
https://github.com/instanseg/instanseg
Apache License 2.0
10 stars 5 forks source link

Using `Pixel Width` and `Pixel Height` is not enough #99

Closed lacan closed 1 month ago

lacan commented 1 month ago

I have a dataset where the cells are rather large, taken with a spinning disk microscope

Pixel size is set to 0.5980 um as per the metadata of the microscope

Runing InstanSeg yields image

Changing the pixel size to 0.2 um yields more expected results (Still wrong, but that's the data)

image

Perhaps an assumption based solely on the pixel size is not sufficient...

petebankhead commented 1 month ago

I think it's more an issue of training data, which I discussed a bit here.

I don't think there's anything in the training set that remotely resembles the spinning disk image.

We want to release multiple models and also simplify training custom models, but this will take a bit of time.

It's possible to adjust the downsample in a script if you want to explore different effects without needing to change the image metadata.

Because I don't think this is an extension bug, I think it's more a topic for the main instanseg repo or (preferably) image.sc.

It's really important though. I would like to have a bigger discussion around assembling a good, fully-open training set. I plan to post something about that to image.sc, probably later in the week.