NA-MIC Project Weeks

Back to Projects List

Automatic segmentation of priority collections from Imaging Data Commons

Key Investigators

• Andrey Fedorov (BWH, USA)
• Kyle Sunderland (Queen's, Canada)
• Vamsi Thiriveedhi (MGH, USA)
• Paolo Zaffino (Magna Graecia U. of Catanzaro, Italy)
• Lalith Kumar Shiyam Sundar (LMU, Germany)
• Michael Onken (OpenConnections GmbH, Germany)

Project Description

Questions? Want to participate remotely? Join our project Discord channel!

The goal of this project is to further increase availability of anatomic segmentations and accompanying radiomics features for the images available in Imaging Data Commons.

Objective

1. Volumetric segmentations of the anatomy for selected images in IDC.
2. Radiomics features for the segmented labels.
3. Interface to explore the resulting data.

Approach and Plan

1. Develop Terra workflow wrapping selected MOOSE, VIBESegmenator and TotalSegmentator models
2. Extract radiomics features using Radiomics.jl
3. Saving resulting segmenations and radiomics features into DICOM using dcmqi
4. Test on a representative sample from IDC (prioritize processing of Cancer Moonshot Biobank and CPTAC images)
5. Apply to a larger cohort
6. Develop review interface, examine results, investigate possible issues
7. Apply workflow to the APOLLO5 dataset (embargoed)

Progress and Next Steps

This week we made progress on debugging/fixing/imporoving segmentation workflow and the associated tools.

1. Refined post-processing component of the workflow - improved/fixed generation of DICOM SEG.
   • https://github.com/Sunderlandkyl/CloudSegmentator/pull/1
2. Fixed handling of deflate transfer syntax in dcmqi, which was broken on windows (which was also broken in Slicer) https://github.com/QIICR/dcmqi/pull/549
3. Adding SNOMED mapping upstream to the MOOSE segmentator https://github.com/ENHANCE-PET/MOOSE/pull/241
4. Started working on making the segmentation workflow applicable to segmentation of non-IDC data (via private Google bucket access)
5. Added the option to allow Terra to access private Google buckets containing dicom files
6. Added export of output DICOM SEG files into Google Cloud DICOM stores
7. Added output CSVs containing performance metrics and structure metrics to the output

Illustrations

Background and References

• https://github.com/Sunderlandkyl/CloudSegmentator/tree/moose_test/
• https://portal.imaging.datacommons.cancer.gov/
• https://terra.bio/
• https://github.com/ENHANCE-PET/MOOSE
• https://github.com/robert-graf/VIBESegmentator
• https://github.com/pzaffino/Radiomics.jl
• https://github.com/QIICR/dcmqi
• https://github.com/wasserth/totalsegmentator
• https://doi.org/10.1117/1.JMI.13.6.062204