Rotations in images and noise in segmentations?

[ThomasBudd]

Hi all,

first of all, a huge thanks for this great dataset and sharing it with the community. It is of great value for us and I think I will use it a lot in the future:)

At the moment we want to segment the different lung lobe and I've had some problems with the Totalsegmentator Dataset. The two problems I've been seeing is that there was a rotation in some images, but not in the labels. The image and labels match, but due to a rotation on x or y axis, some parts of the image contain only background (due to the extrapolation), but the segmentation is still present. See for example here (s0001):

I could remove these segmentations in the void partially by checking for rays in x or y direction in which the image only contains background, but this didn't work for images with the rotation in both directions such as these ones (s0044, s0067):

Also some segmentations look noisy or incomplete to me. I wonder if everyone have had the same issue before. See for example s0036, s0130, s0149

The last thing I've noticed is that some lung lobes look weird to me, for example in s0015. Do you know what is happening here?

I've been loading the images and segementations just with nibabel in python. I've noticed that the segmentations were stored as floats instead of ints and some rounding errors appeared due to the compression. For the segmentations I used seg = np.round(nib.load(PATH_TO_SEGMENTATION).get_fdata()).astype(int).

Is it possible that something in the dicom to nifti export went wrong? Sorry for the long post :)

Many thanks again!

[wasserth]

The segmentations are not always pixel perfect. At the end of a image the segmentation can go into the black a bit. And sometimes segmentations can be a bit noisy. There are also a few subjects with abormal lung anatomy or a major pathology in the lung. In those cases it is not always possible even for a doctor to properly find the lung lobes. Then the segmentation can be a bit fuzzy.

You should also look at the images in a Medical Image Viewer (e.g. 3D Slicer or MITK). There you can see if the images are properly aligned or not.