Back to Projects List
- Andrey Fedorov (BWH)
- Andrew Beers (MGH)
- Joost van Griethuysen (NKI)
- Steve Pieper (Isomics)
- Greg Sharp (MGH)
- Chris Gorgolewski (Stanford)
A common task in medical image analysis is to convert the DICOM image series into a 3d volume that is suitable for the analysis using ITK, deep learning studies, etc. The goal of this project is to summarize, discuss, and evaluate various publicly available tools that exist for this task.
Initially, we want to focus on the very narrow task of converting a scalar volume image series.
Approach and Plan
- Identify the list of requirements: what kind of data should be acceptable
- Compile the list of tools suitable for this task. Include main features of the tool: language, license, support/history of development, pros and cons, …
- Collect feedback from the community - what tool do you use? what issues have you encountered? do you feel like you have a solution that is working reliably?
- Compile a list of representative datasets, discuss approaches that could be used to evaluate a tool.
Progress and Next Steps
- Discussed and started developing the toolset for automated evaluation of conversion tools
- Use Docker auto build to make an image containing all of the converters and preloaded data (DICOM + volume reconstructed baseline)
- Use CircleCI initialized from the docker image
- Use ctest to run each converter, compare with the baseline, compare time
- update web page with the comparison dashboard after each run
- identified new interesting sources of reference data
Questions from @lassoan:
- Very interesting and important project, but for me the project scope and level are not very clear yet. Is it only about finding and sorting files that make up an image series? If we include lower levels, then we have to talk about what DICOM toolkits are used (DCMTK, GDCM, vtk-dicom, pydicom, …). If we include higher levels, then we have to consider how images are reconstructed (geometry: tilted gantry, variable slice spacing, multi-orientation scout scans; image intensity: scaling and various LUTs; measurement units, etc.) and how they are stored in memory (ITK, VTK, or other image) and in file (nrrd, nifti, metaio, …; including metadata). Also, should we exclude 3D+t, color, vector, tensor images?
- It is a good idea to focus the discussion on a specific topic (probably “scalar volume loading” topic is already too wide), but then we should keep in mind that we cannot make general conclusions about what toolkit we should standardize on, as scalar images are not the only data type we are interested in (segmentation object, spatial registration object, structured report, RT objects, are just as important).
mriconvert and mcverter
A mockup of how the public-facing page could look: https://andrewbeers.github.io/dicomCompare/
Background and References
- Clunie D. How Thick am I? The Sad Story of a Lonely Slice. 2013. http://dclunie.blogspot.com/2013/10/how-thick-am-i-sad-story-of-lonely-slice.html. Accessed January 5, 2018.
- Investigate NeuroDebian for simplified install of some converters. [https://hub.docker.com//neurodebian/](https://hub.docker.com//neurodebian/)
- Rosetta Bit library of valid file conversions: https://www.nitrc.org/projects/rosetta/
- Potential data source: NIMH Data Archive is releasing a large set of raw DICOMs as part of the ABCD Study in January 2018.
- Another list of alternatives for volume reconstruction: https://www.nitrc.org/plugins/mwiki/index.php/dcm2nii:MainPage#Alternatives