Scribbles on abdominal CT data

[marianabb]

Hello,

I've been testing a few of the example apps in hopes of using this great tool in our environment. I've now setup a working version of both the 'radiology/segmentation_spleen' and 'radiology/segmentation' apps (our final application would be multi-label, so the 'segmentation' app is the one I'm more interested in). I am using MONAILabel version 0.4.0rc3, 3D Slicer 5.1.0 r30918.

The Scribbles module works as intended in both these applications when using the spleen task dataset. However, it doesn't generate correct results when using the abdominal CT data from https://www.synapse.org/#!Synapse:syn3193805/wiki/217789. I also have the same issues when using my local abdominal CT data.

In most cases, even though foreground scribbles have been drawn, I can see the following in the logs:

[2022-05-13 16:34:04,972] [229768] [MainThread] [INFO] (root:128) - warning: no foreground scribbles received with label 15, adding foreground scribbles to ROI centre

This is the result after I've created a ROI and added some foreground scribbles:

This is another result if I don't include the ROI, which maybe is more of the expected result, but still outside the frame of the original image:

If I add reader="ITKReader" to the LoadImaged pre-transform in class HistogramBasedGraphCut, I get a segmentation in the same frame as the original image, but still inverted (this was also generated without ROI, but if I add one the result is usually just a couple of voxels):

The Auto-segmentation module using the pretrained network works fine, so this seems to be an issue specific to the scribbles part. Any ideas on what to try next?

Thanks in advance!

[SachidanandAlle]

@masadcv can you help here..

[marianabb]

I've also tried this on the reoriented version of the abdominal CT dataset here.

Using a ROI results in just a few voxels segmented in a closer, but still not correct location:

And without ROI, also slightly better but still inverted:

[diazandr3s]

Thanks for the detailed report, @marianabb

While we wait for @masadcv to reply, have you considered using other manual annotation tools available in Slicer? Like, Grow from seeds OR the 3D brush?

[marianabb]

Thanks for the great suggestions @diazandr3s, we will definitely also use those in our pipeline! We're a bit new to 3D Slicer so that video on how to use the 3D brush has already taught me a lot.

The main reason we'd like to use scribbles is to "smartly" edit already existing segmentations generated from our model, although maybe some of the already existing 3D Slicer segmentation tools like scissors or margin might do the job.

[diazandr3s]

Glad to hear this helped :)

[masadcv]

@marianabb many thanks for reporting this. I am looking into the issue and have successfully recreated the error using the Abdominal dataset from https://www.synapse.org/#!Synapse:syn3193805/wiki/217789

The issue seems to be in header information for the images mentioned here. Using the default nibabel to load these images raises the following error message:
[2022-05-16 21:26:56,001] [225596] [MainThread] [INFO] (nibabel.global:274) - pixdim[0] (qfac) should be 1 (default) or -1; setting qfac to 1

I looked at header for one of the images in your screenshots (img0034.nii.gz) using ITKSnap->Tools->Image Information->Metadata (http://www.itksnap.org/) - and I can see pixdim[0] = 0 which does not make sense to me (maybe I am missing something here). Here is this information from ITKSnap:

As a comparison, here is the same information from a volume in MSD Spleen dataset (which has more plausible pxidim[0]):

I will discuss this issue with @SachidanandAlle and @diazandr3s tomorrow and try to work out a solution....

[masadcv]

@diazandr3s maybe we can help @marianabb with reorienting images, so this can be resolved? Is there a code to correct header for these problematic images?

[marianabb]

I had not tried using the ITKReader again, you're right, now this is pretty close to working properly!

I can reorient all my images to RAS for now and hopefully this will be fully resolved later :)

Thanks for all the help!

[diazandr3s]

I think it'll be great to have these images in the NRRD format instead of Nifti. There is now an NRRD reader in MONAI Project-MONAI/MONAI#4259 and a writer in MONAI Label #779

@marianabb would you like to give it a try if Nifti files don't work?

[marianabb]

I believe the root problem (and please correct me if I'm wrong) is not the image formats, but the image orientation. Without the proposed changes the tool only works properly in images with RAS orientation (Nifti or NRRD). With the proposed changes the tool works properly in both formats if no ROI is used to limit the segmentation, the ROI part is still only working properly with images in RAS orientation.

[diazandr3s]

I might be wrong, but I think the issue is the image header.

I meant we first load the images, correct the headers and then save the images in NRRD or Nifti format