MONAI AutoSeg3D for CBCT (CMF)

I tried MONAI AutoSeg3D auto-segmentation on CMF scan (CBCT).
I see that the segmentation results appear like as if it was done on a downsampled image.

Please see the snapshot here:

I have used Whole Body segmentation TS2. I think TS2 implies Total Segmentator V2. I Observed a similar result when I used the Total Segmentator extension.

I also see the following log in the AutoSeg3d:


Using orientation_ras

Using crop_foreground

Using resample with resample_resolution [1.5, 1.5, 1.5]

Running Inference ...

So it appears it resamples the image to 1.5x1.5x1.5

Is it possible to run the AutoSeg3D to get a better(finer) segmentation mask result? The extension currently does not provide such a control.

Thanks

Hello,

To my knowledge Monain Auto3DSeg has not been trained on those kind of data for now. You could give a try to DentalSegmentator : https://discourse.slicer.org/t/new-extension-dentalsegmentator-for-automatic-dental-ct-cbct-segmentation

Best,
Gauthier

@Gauthier thanks for your reply.
Yes, I have used DentalSegmentator and AMASSS. They perform better. I wanted to see the performance of this tool as well. Basically, then use the best one.

On Dental Segmentator: Only the extension code is open source, however I could not find the implementation on auto segmentation model training. Is that also available, say if someone wants to extend work on Dental Segmentator?

Regards,
Sukhraj

Hello,

Thanks for your reply.

The pretrained nnU-Net model for DentalSegmentator is shared here : DentalSegmentator nnU-Net pretrained model for CBCT image segmentation

Best regards,
Gauthier

1 Like

So essentially, learn how to use and train further a pre-trained model in nn-Unet, is all that is needed to extend Dental Segmentator.

Yes that would be the idea! nnU-Net offers a basic implementation of finetuning: nnUNet/documentation/pretraining_and_finetuning.md at master · MIC-DKFZ/nnUNet · GitHub

We haven’t used it for now, so I could not tell you more about the performance bumps that coud be obtained. It will probably depend a lot on what is your objective!

1 Like

Have you tried the whole head model in Monai AutoSeg3D? You will get the skull, the mandible, the rachis, eyeball, brain, and masticatory muscles.

1 Like

@philippepellerin I can’t find the Head Model int the model list, please see the snapshot:
image

Am I missing something? I am using Slicer 5.7.x

@philippepellerin Thanks for noticing in detail.

I observed a head model “Whole Head Segmentation” is available in 5.6.2 version of slicer.