Label Documentation

[dpardi02]

Is there a place I could look to find documentation regarding how to format my premade labels into MONAI Label? I have had struggles thus far where the labels I submit are not recognized and are not contributing to the training and learning. They are all Nifti files but there might be other parameters I have configured differently for my files. Thanks for the help.

[diazandr3s]

Hi @dpardi02,

Thanks for asking this.

In this video, I commented on how the volumes and labels should be organized to use MONAI Label: https://youtu.be/-HAryYAO5J4?si=OaZMhMQW9lh-Njvb&t=778

Here is more information: https://github.com/Project-MONAI/MONAILabel#local-datastore-for-the-radiology-app-on-single-modality-images

Are you working with non-overlapping segments? How many annotated volumes you've premade?

Hope this helps,

[dpardi02]

Hi @diazandr3s,

Thanks for that additional information. I am pleased to say we have figured out our label discrepancy and were able to correctly configure our labels to run the training. Thanks so much for all your help through this process.

Best,
Daniel

[diazandr3s]

Thanks for the update, @dpardi02.
If you can, please share more details so others can benefit from this.

[dpardi02]

Sure thing @diazandr3s

One issue was the orientation of some of our labels where we were exporting them from ended up switching planes in some instances so they had to be permuted back to match the raw data.

However, the main issue was the content of our labels. Upon further inspection, we found that some errors that were being given was pertaining to the use of multichannel labels. After converting our labels to strict binary we found success.

Again thanks for your help. I have just one final question before I close this ticket if you don't mind. What customizable options are there to the training and does it all take place in the configs document? I know that the number of max epochs can be changed from here but was wondering if there were any other options. Additionally, say I open my dataset, run training, and then close the slicer and MONAI Label session, is there a way to see the saved training somewhere and apply it to images elsewhere? In other words, can the product of a training session be used at a different time or does it have to be applied to images then and there?

[diazandr3s]

Many thanks for the update @dpardi02
Glad you found the source of the problem.

About your questions:

What customizable options are there to the training and does it all take place in the configs document?

Yes, most of them are in the config file. However, you can go deeper and modify in the training file the learning rate, optimizer, loss function, pre-transforms (data augmentation), etc: https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/trainers/segmentation.py#L61

is there a way to see the saved training somewhere and apply it to images elsewhere? In other words, can the product of a training session be used at a different time or does it have to be applied to images then and there?

When you trigger the training process, you'll find the best model in .pt format, several checkpoints and Tensorboard file that you can use to monitor the training - all this inside the folder radiology/model.

You can use the .pt file and run it as any PyTorch model. No need to specifically use MONAI Label.

Another way is by running MONAI Label without starting the server. For this, you can run inference using the main function of the Radiology app: https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/main.py#L267

I hope this helps,

[dpardi02]

Perfect, thank you @diazandr3s for all the help with my questions. This has greatly progressed the work on my project. Much appreciated.