NucleiSegmentationBoundaryModel

[utterances-bot]

NucleiSegmentationBoundaryModel

Bioimage.io -- an AI model repository for deep learning.

https://bioimage.io/?tags=affable-shark&type=model&id=10.5281%2Fzenodo.5764892

[k-dominik]

Was the data trained on the whole training data? So including color images? What transformation has been applied to those. Should this maybe be part of the preprocessing spec? I think it would be great to be able to take any images out of the training/test set and plug them in. But here the step of "flattening" the data to a single channel is not included in the model. It seems that the training already gets grayscale images...

[constantinpape]

The model is trained on this data: https://bioimage.io/#/?tags=affable-shark&type=model&id=ilastik%2Fstardist_dsb_training_data.
It only takes into account the greyscale images from dsb, and the subset was created by the stardist paper.
You can find the training data here: https://github.com/stardist/stardist/releases/download/0.1.0/dsb2018.zip

Note that this information is contained in the metadata in the source field (see https://github.com/ilastik/bioimage-io-resources/blob/main/collection.yaml#L117). (But this field is for some reason not displayed in the card for the dataset.)

[k-dominik]

thx @constantinpape for the update. For some reason, the data link in the card points here: https://www.nature.com/articles/s41592-019-0612-7 (not the stardist paper).

Clicking the data icon, however, does link to the stardist one...

[FynnBe]

(But this field is for some reason not displayed in the card for the dataset.)

this issue: bioimage-io/bioimage.io#312 is still open...

[carlosuc3m]

It would be great if this model had the names of the inputs to the actual DL models specified. There is an issue discussing this: bioimage-io/core-bioimage-io-python#334.

In this particualr model onnx requires the input to be called input.0

[carlosuc3m]

Also the input size for the onnx model seems to be constrained to [1, 1, 256, 256] only