Create deepedit tutorial

[diazandr3s]

Signed-off-by: Andres Diaz-Pinto [email protected]

Description

Add tutorial to run deepedit model on MONAI Core

Status

Ready

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[Nic-Ma]

Hi @diazandr3s ,

Thanks very much for the tutorial!
As we are going to release MONAI bundle soon, would you like to also contribute a bundle for this deepedit model?
The get_started tutorial and reference example are at: https://github.com/Project-MONAI/tutorials/tree/master/modules/bundles.
And related documentation: https://docs.monai.io/en/latest/bundle_intro.html.

What do you think?
CC @ericspod @wyli

Thanks in advance.

[ericspod]

Thanks for the tutorial!

It's great to have a tutorial introducing DeepEdit but I think it needs more explanation about what's going on and what the moving parts are. The notebook doesn't make clear what the specific elements of DeepEdit are and how they're used, in the first part there is a sequence of transforms which appears to add a landmark into the data dictionary, but the parts which do this aren't explained and the use for this isn't made clear. In the second part a pretrained model is used to segment but how that relates to the first part isn't clear. I understand what the moving pieces are in DeepEdit from other presentations but this needs to be motivated here.

There's a number of functions in the notebook for plotting, you might want to put these into a PR in core to be kept here if they're generic enough. A general routine for overlaying landmarks would be useful in general.

There's also a lot of code in the custom_*.py files, for a tutorial this makes it hard to follow and there seems to be duplication here with what's in MONAILabel itself, could this be reduced or is the point to avoid importing MONAILabel? The notebook imports a number of classes from here that aren't used, it would be clearer if there were no unnecessary imports.

If would be great if this model were packaged into a bundle with an added inference.json file at least, a training.json file would be great as well that replicated what's in the training script. It would be easier to take the model into other contexts such as deployment.

[diazandr3s]

Many thanks for your comments and suggestions, @Nic-Ma @ericspod. They are helpful! :)

I'll create a couple of PRs to MONAI core to include custom transforms, custom interaction, and common plotting functions ... All this to simplify the tutorial. I'll also make the steps/content clearer

I really like the idea of creating a bundle for DeepEdit. I'll work on this once we have the latest version of the tutorial.

Will keep you posted!

[daguangxu]

Many thanks for your comments and suggestions, @Nic-Ma @ericspod. They are helpful! :)

I'll create a couple of PRs to MONAI core to include custom transforms, custom interaction, and common plotting functions ... All this to simplify the tutorial. I'll also make the steps/content clearer

I really like the idea of creating a bundle for DeepEdit. I'll work on this once we have the latest version of the tutorial.

Will keep you posted!

@diazandr3s, we plan to release MONAI 0.9 in mid-June. Can you try to finish this DeepEdit bundle before end of May? Thanks.

[diazandr3s]

Many thanks for your comments and suggestions, @Nic-Ma @ericspod. They are helpful! :)
I'll create a couple of PRs to MONAI core to include custom transforms, custom interaction, and common plotting functions ... All this to simplify the tutorial. I'll also make the steps/content clearer
I really like the idea of creating a bundle for DeepEdit. I'll work on this once we have the latest version of the tutorial.
Will keep you posted!

@diazandr3s, we plan to release MONAI 0.9 in mid-June. Can you try to finish this DeepEdit bundle before end of May? Thanks.

Yes, I'll be working on this next week.

[diazandr3s]

Quick update on this:

• Add all DeepEdit Transforms and Interaction class in MONAI Core. Here is the PR Add DeepEdit transforms and interaction MONAI#4164
• Update the imports in Notebook and train file
• Update the DeepEdit schema and add more information about the way it works

Once I have the MONAI release candidate or MONAI weekly version, I'll update the files

[Nic-Ma]

Have you tested both the DynUNet and UNETR? May I know the difference in the result?

Thanks.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

Nic-Ma commented on 2022-05-09T09:55:05Z
----------------------------------------------------------------

Please add these optional imports in above cell directly.

Refer to other tutorial:

https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/spleen_segmentation_3d_lightning.ipynb

Thanks.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

Nic-Ma commented on 2022-05-09T09:55:06Z
----------------------------------------------------------------

Line #35.        print("Final PLOT:: {} => image shape: {}, pred shape: {}; min: {}, max: {}, sum: {}".format(

I think here it plot too many charts, could you please help reduce them? Maybe less than 10 charts?

Thanks.

[diazandr3s]

Have you tested both the DynUNet and UNETR? May I know the difference in the result?

Thanks.

Yes! It works better than the DynUNET. But I wanted to keep the DynUNET as default as it takes less time training

[diazandr3s]

View / edit / reply to this conversation on ReviewNB

Nic-Ma commented on 2022-05-09T09:55:06Z ----------------------------------------------------------------

Line #35. print("Final PLOT:: {} => image shape: {}, pred shape: {}; min: {}, max: {}, sum: {}".format(

I think here it plot too many charts, could you please help reduce them? Maybe less than 10 charts?

Thanks.

Agree!

[diazandr3s]

View / edit / reply to this conversation on ReviewNB

Nic-Ma commented on 2022-05-09T09:55:05Z ----------------------------------------------------------------

Please add these optional imports in above cell directly.

Refer to other tutorial:

https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/spleen_segmentation_3d_lightning.ipynb

Thanks.

Agree. Just changed it

[Nic-Ma]

View / edit / reply to this conversation on ReviewNB
Nic-Ma commented on 2022-05-09T09:55:06Z ----------------------------------------------------------------
Line #35. print("Final PLOT:: {} => image shape: {}, pred shape: {}; min: {}, max: {}, sum: {}".format(
I think here it plot too many charts, could you please help reduce them? Maybe less than 10 charts?
Thanks.

Agree!

I think you don't need to add another variable num_slides, just for i in range(10) or for i in range(0, 100, 10)?

Thanks.

[diazandr3s]

View / edit / reply to this conversation on ReviewNB
Nic-Ma commented on 2022-05-09T09:55:06Z ----------------------------------------------------------------
Line #35. print("Final PLOT:: {} => image shape: {}, pred shape: {}; min: {}, max: {}, sum: {}".format(
I think here it plot too many charts, could you please help reduce them? Maybe less than 10 charts?
Thanks.

Agree!

I think you don't need to add another variable num_slides, just for i in range(10) or for i in range(0, 100, 10)?

Thanks.

agree!

[wyli]

Thanks, it looks good to me.

[Nic-Ma]

Looks good to me.

Thanks.