Auto Segmentation of 2D Images -- Fail

[mgorin1985]

I am using MonaiLabel and Slicer 3D

I have a project with 2D images, all converted to grey scale, and in nii.gz format

I used MonaiLavel + Slider 3D to train how to segment a structure contained within the 2D images. I get validation dice scores north of 90%.

I then load an image that the model didnt train with (prepared the same way, grey scale, 2D, same size, .nii.gz etc) and then run the auto segmentation feature and every pixel in the image is included as part of the segmentation mask.

What am I doing wrong?

I have a feeling the model the auto segmentation tool is running is not my newly trained model. Could this be the case? If so, how do I tell Slicer to now use my model.

Thank you all!
Mike

[diazandr3s]

Hi @mgorin1985,

The reference segmentation model is implemented for 3D volumes.

If you'd like to use it for 2D annotation, you should change:

• the label names (https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L33-L47),
• network configuration (https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L69-L77) to work on 2D
• ROI (https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L66) i.e. (96, 96),
• spacing (https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L64) i.e. (1.0, 1.0) and
• pre transforms for train (https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/trainers/segmentation.py#L94-L97).

Hope this helps.

[mgorin1985]

Thank you so much for your reply. Can you be a bit more specific about what I would change in each of those files? For label names I get that. I have only 1 label name now and that is prostate, but the rest is a bit foreign to me as a newbie.

Mike

[diazandr3s]

The links I sent point to some of the parameters that need to be changed for the model to work on 2D images. Most of the default values are for 3D images. i.e. ROI. A region of interest (ROI) for 2D images should be a 2D array not 3D as it is by default. The same for spacing and the network configuration.
Are you using X-rays images? If yes, this discussion may also help: #829 (comment)

[kaufmannb]

Thanks for your reply. Unfortunately I have the same issue. Although I made all changes as described above, It's not possible to login to the server due to errors in the code.

Would it also be possible to use Pathology or Endoscope? I want to use the automated segmentation with 2D ultrasound scans.

Many thanks.

[diazandr3s]

It depends on which viewer you're more comfortable with. The pathology app is more for histopathological images using QuPath and the endoscopy app is more for endoscopy videos using CVAT viewer.
I think the most suitable viewer for this task is Slicer. But again, the segmentation model in the radiology app needs some changes to be made. This discussion may be useful: #829 (comment)