Labels are being aggregated when retrieving via Slicer Plugin

[LG1234]

Hi everyone,

I have a bit of a problem i cant seem to solve. I am serving pre-annotated images of a spine via monai label. This includes cortical bone, trabecular bone, bone islands and foreign material for every vertebral body, the pelvis and the femur on both sides, so quite some labels.
To keep it consistent, i am using 1+XX for cortical bone, 2+XX for trabecular bone, 3+XX for bone islands and 4+XX for foreign material (XX being the number of the vertebra), so the segmentation IDs are going up to 460 or something.

When opening the files locally, everything is fine with the segmentation masks.
However, when serving over the label server and opening via the MonaiLabel extension (either via localhost pr a tunnel), all trabecular bone data is aggregated unter 455 (i.e. the entire 2+XX segmentation masks are delivered in the single 455 mask), and some other IDs are mismatched (1+XX is served as 3+XX).
I am serving without a json sidecar, and the label maps are injected in the server script.

Does anyone have an idea on how to resolve this?

Thank you!

[production23]

Hi,

This sounds very familiar, we’ve run into the same kind of label-aggregation issue in medical imaging projects where large ID sets need to stay consistent across server + client. It usually comes down to how the label server is mapping values, and we’ve built workflows to make sure nothing gets collapsed or misassigned.

If it helps, we’d be glad to set up a quick pilot with you to streamline this and ensure your label maps stay fully intact across Monai Label and Slicer.

Best,
Aram @ EvoLearns

[lassoan]

Labels that you have in your source data are usually not the same as labels that you use in your training data (for example, because you put together the training data set from 5 different sources, each using different labels). To avoid complications, I would recommend that when you prepare your training data, assign labels to structures without leaving huge gaps.

I would also recommend to not rely on particular label values in your source data, but use standard terminology codes instead. For example, in each segmentation file, specify mapping from label value to terminology codes (as it is done in .seg.nrrd files), and use slicerio package to create training data with any label assignment you prefer.

[production23]

Thanks Andras, this is a common pain point we’ve seen. We help teams solve it by setting up workflows that keep labels consistent and training-ready across multiple sources. Happy to explore a quick pilot if that’s of interest.

[lassoan]

Sure, I would be happy to discuss this further. How do you specify the meaning of each segment and where do you store mapping to label values? Which file formats do you use (DICOM, NRRD, NIFTI, ...)?

[LG1234]

Thank you for your help. This was indeed an overflow issue, probably because I had too many labels overall with too many "empty labels". I packed them all from 0-161, and then it worked.