In the myeloma project, we pre-segment some structures for our medical partners that they have to inspect and correct with MITK. The newest version (v2021.02) does load and display the segmentation but does not show it in the segmentation plugin which is needed to correct it.
I tried with some older versions of MITK and it worked there.
Example data is here: https://hub.dkfz.de/s/aF5iKe6YFXjSQ9C
| rMITK MITK | |||
| Restricted Differential Revision | rMITK165ad5f01412 T28356 Make_use_of_coordinate_precision_check | ||
| Restricted Differential Revision | rMITK453b6b57f01c Fixe review remarks (typos) | ||
| Restricted Differential Revision | rMITKc2c4e9649bfc BREAKING!!! | ||
I had a first quick check and for example T1.nrrd and S.nrrd (and S.nrrd converted to LabelSetImage) have the same geometries. My first assumption would have been that they are different. @floca Any idea?
It is, once again the precision of the geometry check. :sigh:
@kislinsk could we have a littel chat somewhen?
@kleina : Which tool /script generated the masks? nnUnet?
The orientation matrix is differs a bit. Just a bit e-5. But enough to currently fail our checks. One of the questions is why it is different in the first place if it generated from the input image. Could be a series of rounding errors or different valur representations...
@kleina Even though the problem is now fix by being less pedantic, I realy would identify the reason for the fidelty loss in the mask geometry. So the information which tool produced the mask is still welcome. Especially if it one of our tools/nets so that we can fix it.
@floca André is on vacation. The original mask in nifty was produced by nnUNet, but we are using some of our Python scripts to split segmented tissues to different labels and to convert it into .nrrd format. Maybe there could be a rounding error during loading and saving the image header, we have to look at it.
@j038k Hi I you look for the problem, keep in mind that the rounding error can occure when it serialized into the file (if it is in text mode). E.g. some code that saves nrrd only does it with standard precission which is in some cases 8 characters (including the dot, sign etc. this can then be not enough and you loose decimal places).
Sorry, I was on vacation.
@kleina : Which tool /script generated the masks? nnUnet?
I use nnUnet to create the initial masks and then a self-written python script to split the multi-label masks into separate files.
@kleina Do they use multi label segmentation view or the old one?
To be honest I'd have to ask the medical partners. If I remember correctly, they use the old one.
@kleina Even though the problem is now fix by being less pedantic, I realy would identify the reason for the fidelty loss in the mask geometry. So the information which tool produced the mask is still welcome. Especially if it one of our tools/nets so that we can fix it.
Thank you for the suggestions. I will take a look.