search

microsoft / InnerEye-DeepLearning

Medical Imaging Deep Learning library to train and deploy 3D segmentation models on Azure Machine Learning
https://aka.ms/innereyeoss
MIT License
557 stars 142 forks source link

Suggestion to include SEG and not only RTSTRUCT IODs for the segmentations #382

Open athanell opened 3 years ago

athanell commented 3 years ago

Hi guys, This was a suggestion that I brought up to the InnerEye-CreateDataSet repo, but I was advised to repost it here:

Suggestion: For the case where the image+segmentation sets are in DICOM please consider the option of adding a proper DICOM IOD for the segmentations in your options (more details in http://dicom.nema.org/dicom/2013/output/chtml/part03/sect_A.51.html ) Why: The RTStruct is an old format that has been created for radiation therapy. It creates planar contours (surfaces) instead of volumes. When the geometry of the segmented structure is complex, converting the RTSTRUCT IOD to a binary labelmap (like in a typical nifti format segmentation with 0 values on all voxels that are of no interest and 1 for all the voxels of interest) will lead to mistakes and segmentation quality degradation. That's the reason modern segmentation software that can segment DICOM images will save the segmentations as SEG objects and not RTSTRUCT. Since though a lot of segmentation material exist in the form of RTSTRUCT IODs, the converter of RTSTRUCT to NIFTI of this utility will certainly be beneficial.

Best Regards

AB#3915

ant0nsc commented 3 years ago

Hey @athanell , thanks for this good suggestion. We are presently working on code that does the conversion of segmentations to Dicom, we'll keep that in mind there. I might reach out to you for details later.