MONAI Label - Pathology

[SachidanandAlle]

Semantic segmentation for WSI Images

1. LOAD Image in Viewer (DigitalSlideArchive/QuPath) and select ROI
   

2. Annotate ROI (e.g. Tumor, Stroma etc..)
   

3. Submit ROI to MONAI Label (WSI/ROI + Annotations) for Training

4. Train/Finetune a model

5. Run Inference on WSI and produce/show label masks

6. Repeat steps 1 to 5 on same WSI image for different ROI or try other WSI image

[andreped]

@SachidanandAlle In my PhD I am working with digital pathology, and I just recently published a paper on this exact topic (see here). However, our pipeline consisted of using a mixture of open solutions instead of integrating everything into a single platform. Having everything into a single platform would be more ideal, but it is very challenging and tedious to do so. Therefore we took advantage of the solutions that "did it best", and made it possible to communicate and transfer files between solutions easily.

Nonetheless, packing this into M-L could be very relevant for our research group (and the digital pathology field as well), as we could very easily integrate our architecture and pytorch code into the M-L framework.

Therefore, I am very positive to the idea! Just remember that AFAIK Slicer has very little (or no) support for rendering WSIs. For reading these images we often use OpenSlide. Also, if you wish to annotate, you would need very good tools to do so, especially if you wish to have similar annotation user experience as with QuPath. Could be quite a lot of work... If you wish to render images in Python, you could use pyFAST, which we are developing at SINTEF. Could be an idea to start from there and build upon pyFAST instead of starting from scratch.

There are also other people working on having a one-platform-does-all solution for active learning-based digital pathology (e.g., QuickAnnotator). Would recommend checking how they did it :]

[SachidanandAlle]

For early access, please find the sample pathology app for MONAI Label under:
https://github.com/Project-MONAI/MONAILabel/tree/main/sample-apps/pathology

NOTE:: Things are in progress/under-development..

[SachidanandAlle]

Need some developers from QuPath to add a plugin over here:
https://github.com/Project-MONAI/MONAILabel/tree/main/plugins

The prototype is here.. (which is kind of simple reference implementation to access MONAI Label APIs for Infer)
https://github.com/SachidanandAlle/qupath/tree/main/qupath-extension-monailabel
(Few things are still kind of hack.. like enable ROI for MONAI Label using Ctrl+Click)

[SachidanandAlle]

Still getting openslide error?

[SenhorUnk]

I figured out, that openslide only supports a specific image format. I was using some sample .png, and .jpg, converted them to .tiff and still got the same error. I found a .svs file online and tried it too. The error changed to:

The error appears to be related to openslide. Do you by any chance know, how to not use openslide on the MONAI server side, since my input images come as normal image. Thanks already for your help!

[SachidanandAlle]

You can use normal png images.. use the latest alpha version.. and use monailabel from main branch (developer mode.. uninstalled pip package).. and u don't need to include your images in the studies folder

[SenhorUnk]

Alpha Version of QuPath: 0.3.2 ?

I will try as suggested - thank you!

[SenhorUnk]

It worked :) thank you!

[andreped]

Might be that @petebankhead is interested in contributing.

Would be a great extension to QuPath and pathologists would surely embrace what this add-on has to offer. I also think it is a smart option to integrate digital pathology-related stuff into QuPath rather than Slicer, as pathologists are more familiar with QuPath. That everything is built upon MONAI also secures that new technology, architectures, and concepts/designs can easily be added and improved over time. Great idea, @SachidanandAlle !

[petebankhead]

Looks interesting!

I'm the only QuPath developer these days (but recruiting soon...) and completely swamped with tasks, so I'm afraid I can't do anything much directly right now.

However, I'm putting together plans for a webinar and hackathon starting (probably) 25 April*. It could be good to discuss further during that week, and perhaps work on any necessary extension then. By that time, I hope to have a new release, better developer documentation, and various enhancements available that could help.

*-No official announcement yet, but I plan to post about it on the website, twitter and image.sc within the next couple of weeks.

[SachidanandAlle]

@andreped have you tried the latest 0.4 release candidate and pathology example.. either on DSA or QuPath..?

Until we wait for QuPath developers to contribute and make it better, I could just give simple plugin to do basic things like annotate, submit and retrain the model.. of-course interaction can be made better.. specially in case of click interactions (and reduce mouse clicks to run the model)

[andreped]

@SachidanandAlle I have been quite occupied developing my own digital pathology software(s) recently :P, which just now had its first major release, FastPathology.

However, FastPathology is purely a deployment solution, and does not, nor intend to, support training and development of new models, as MONAI Label is meant for.

I could test MONAI Label with QuPath, if you'd like. I also have a pathologist who will definitely be interested in testing the plugin and providing feedback. @SahPet: Do you have time and are interesting in testing the plugin? We could do an interation together, if you'd like, @SahPet?