How to give you more training data?

[LeonardoLupori]

Idea: crowd-sourcing training data from Verso

Hey there!

First off, thank you for DeepSlice, it's a great tool and has saved us a lot of time.

I'm one of the main developers of Verso (https://github.com/LeonardoLupori/verso), where we've integrated DeepSlice. We're gradually wrapping up early development and starting to test the tool with colleagues.

As people use Verso, they'll be producing manually curated ground-truth data across all sorts of staining patterns. Often from real-world slice images that are imperfect (cuts, damage, incomplete sections, and so on). That got me thinking: it could be really nice to add a feature that packages a minimal set of this data in a standardized format, so users can upload or send it to you for training future iterations of the model.

I think DeepSlice has big potential if we can find a way to crowd-source training data, and making it a few-button-clicks process could be a great UX incentive: a way to collect data even from people who otherwise wouldn't have the time, ability, or interest to contribute.

If this idea resonates with you, I'd love to discuss the best way to go about it, both in terms of file formats and file transfer.

All the best,
Leo

[PolarBean]

Hey Leo!

Thanks for this initiative. I agree this data is really valuable. I would definitely be interested in it. I think the best way forward would be to model the feature based on what cellpose does. Which is to have a button that manually submits a users aligned data to a server.

In terms of file formats there is a long standing discussion around a standard for these kinds of data but its a ways off. For the moment as long as we have the anchoring vectors, images, and displacement fields that should be sufficient.

I have a huge amount of serial section images collected via the DeepSlice website but unfortunately this does not include user registrations which is the really valuable part.

How are you thinking about proceeding with this?

All the best,

Harry

[LeonardoLupori]

Hi Harry,

Thanks a lot for the response. I'm glad this looks interesting for you.

About the data format.

At the moment in Verso, we already have:

• anchoring vectors for each slice
• information for non-rigid registration in the form of pairs of control points [src_x, src_y, dst_x, dst_y] (similar to Visualign)
• "low" resolution thumbnail images. These are multichannel ome-tiff files with a max resolution of ~2000 on the long side.

I think here there are two main things to clarify:

1. For the displacement field, do you need a full field of dx,dy displacement for the entire image, or is the collection of control points sufficient?
2. For images would you need 299x299 downsampled versions? What file format? (At the moment, users in Verso might have images with more than 3 channels). Is 8-bit depth ok?

About the file upload

This is a bit out of my comfort zone and would probably be the biggest effort.
One naive way I can think of would be having, on your side, an API running on a server that the users can send data to through HTTPS.
On the receiving side, this obviously requires some sort of validation and security checks. Ideally, we could package everything in verso and send just an archive that is easy to checksum.

What are your thoughts?

[PolarBean]

1. Control points are sufficient if we have the algorithm to generate the deformation from these (and this contains all necessary information to replicate).
2. For DeepSlice at the moment 299x299 is all it takes in but I think there are likely gains from training a new model on higher resolution images. likely 1000x1000 is enough.

Im not so worried about the security side of things to be honest, just having some way of receiving files zipped or otherwise. Would you want me to set up an endpoint at www.DeepSlice.com.au for this or would you want to handle the data yourself. both are fine :)