Setting up custom annotation projects using MONAI Label and 3D Slicer

[rjkowalski]

I am a newbie to MONAI Label. I am aiming to setup custom annotation/segmentation projects with MR datasets using MONAI Label and 3D Slicer. We require the ability to create custom segments for each project. At the moment, I am deploying a MONAI Label instance under Docker - mounting the volume of our local project datastore to /workspace/dataset. I have the data structured in the format MONAI Label requires: using nii.gz files for volumes and have dataset/labels/original and dataset/labels/final created. Once the Docker container is running on our compute server, I access the MONAI Label Server from the 3D Slicer MONAI Label module. This is working as expected but I can't seem to figure out how to just have a custom list of segment names in the segment editor. I see that you need to specify a model with starting the server so I have chosen one model or all and have set the use_pretrained_model to false: "--conf models all", "--conf use_pretrained_model false".

Is there a way to deploy a MONAI Label server without using any pretrained models and use the 3D Slicer interface to perform manual segmentations? i.e. "monailabel start_server --app=apps/radiology --studies=/workspace/dataset" (no --conf arguments)?

[wyli]

I think you still need to define the network architecture and some hyperparameters which are included in --conf, do you mean training from scratch instead of loading model weights from a bundle?

cc @diazandr3s

[diazandr3s]

Thanks for the ping, @wyli.

@rjkowalski this is a good question. As @wyli mentioned, you have to define the network architecture, labels to segment, etc.
For instance, if using the segmentation model, these are the labels you'll see in Slicer: https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L37-L63

and the network: https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L82

You could start the MONAI Label server with no pre-trained models. But there must be one configuration. i.e. --conf segmentation

I suggest not to use all models as they have different labels.

[rjkowalski]

@diazandr3s @wyli - Thanks for your input. This is very helpful and I will try this strategy to deploy MONAI Label server with a modified segmentation model.

[rjkowalski]

@diazandr3s - one quick follow up. Are the keys in self.labels dictionary required to be lower-case? Just wondering if I will need to modified data that has already been segmented manually. For example, I have 3D Slicer scene files (.mrb) that contain volumetric data at segment names - SPLEEN, NOT-SPLEEN. Just wondering if I can set self.labels = {"SPLEEN": 1, "NOT-SPLEEN": 2} or if they are required to be lower case in configs/segmentation.py.

[diazandr3s]

Hi @rjkowalski,

Good question. They don't need to be lowercase- just try to be consistent.
The most important thing here, though is the label indexes. You have to be consistent with the label indexes. In your example, the spleen label should always be 1 and the not-spleen always be 2.

Also, if you are using the NIfTI format, it doesn't keep the label names. Only the indexes.

Hope this makes sense.

[rjkowalski]

@diazandr3s - yes, I am using compressed NIFTI so I'll be sure to keep the indexes consistent. Thanks very much for your feedback. This is very useful!