Difference between revisions of "User:Inorton/Slicer4:DTMRI Thoughts"

From NAMIC Wiki
Jump to: navigation, search
Line 9: Line 9:
 
** DTI
 
** DTI
 
* Preprocess DTI  
 
* Preprocess DTI  
** convert to NRRD
+
** convert to NRRD (module: DicomToNRRDConverter)
** estimate tensors
+
** estimate tensors (module: Diffusion Tensor Estimation)
* Coregister all images using BrainsFit
+
* Coregister all images (BrainsFit)
** DTI Baseline -> T2  
+
** DTI Baseline -> T2 (affine first, then rigid)
** T2 -> T1
+
** T2 -> T1 (rigid)
 
** additional images as available
 
** additional images as available
 
* Segment tumor or pathology region
 
* Segment tumor or pathology region

Revision as of 17:05, 4 February 2011

Clinical perspective

We are using DTI in the context of 2-4 Neurosurgery planning cases per week. We use Slicer FiducialSeeding, tractography selection on a commercial neuronavigation program, and occasionally other Slicer DTI tools as case and surgeon requests dictate. We have used the Slicer3 FiducialSeeding functionality in the OR (with OpenIGTLink/BioImageSuite/BrainLab driving the seeding location).

Workflow

  • Scan and push images to clinical PACS and our clinical fMRI processing workstation (has DICOM SCP and database capability)
  • Pull/copy images from clinical PACS and processing workstation to research laptop/workstation
  • Load all images:
    • T1, T2, (sometimes CT, PET, CBV from perfusion MR, etc.)
    • 2-4 thresholded fMRI activation volumes (coregistered and resliced to structural in SPM)
    • DTI
  • Preprocess DTI
    • convert to NRRD (module: DicomToNRRDConverter)
    • estimate tensors (module: Diffusion Tensor Estimation)
  • Coregister all images (BrainsFit)
    • DTI Baseline -> T2 (affine first, then rigid)
    • T2 -> T1 (rigid)
    • additional images as available
  • Segment tumor or pathology region
    • manually in editor with levelset selection or drawing tools
    • semi-automatically in editor using GrowCutSegment
  • Bring to clinician to use FiducialSeeding to explore area around tumor.
  • (if requested, re-load DICOMs on clinical navigation suite and re-select tracts identified in Slicer)

Background Thoughts

Audiences

  • Clinical/research end users: need simple, efficient, relatively intuitive workflow to generate tractography and perform selection and statistics operations.
  • Clinical developers: integrate DTI functionality for domain-specific purposes (neurosurgery, neurology, etc.)
  • Pipeline developers: the underlying implementations need to be abstracted sufficiently to allow creation of pipeline tools for large-study purposes.
  • DTI Researchers:
    • Can use Slicer+ipython+numpy+... instead of matlab and custom code. Advantages: image conversion (DICOM) and visualization taken care of by Slicer. Challenges: learning curve; the python suite is less integrated than matlab, but it's getting better; pure matlab is relatively more stable (mex stuff can introduce problems).
    • Implementation of new algorithms in Slicer opens up larger potential userbase.

Slicer advantages

See big list of DTI software here: User:inorton/DTI_Software_List.

There are several excellent DTI-centric applications. What advantages does Slicer have for DTI work?

  • More user-friendly data loading: TrackVis requires command line preprocessing; MedInria and TrackVis require manual gradient entry; DTI studio is limited to ROI exploration only (as far as I know)
  • Many segmentation options already available - no external tool (TrackVis, DTI Studio) or separate interface (MedInria) required.
  • Already integrated with intra-operative systems via OpenIGTLink functionality
  • Open-source license (TrackVis closed, MedInria non-commercial, DTI Studio closed)
  • DicomToNRRDConverter test suite: validate DICOM loading from many different scanner types, with special emphasis on DTI private header information.

Slicer improvement areas

(this is referring to Slicer3 interactive DTMRI tools: these areas need improvement in Slicer 4)

  • FiducialSeeding is too slow with high-density seeding, low step-values, or large seeding regions. This is probably intrinsic. Possible improvements include using a multi-threaded seeding filter, or using the GPU (?). But that may be superfluous: with good selection tools, dense pre-seeding in whole-brain or at least extended area of interest may be a better option.
  • Current fiber data model is inefficient for interactive use on large sets (tens of thousands) of fiber tracts.
  • Missing good interactive ROI selection, clustering, and editing capability for pre-computed fibersets.
  • Need subset selection and coloring.
  • Need interactive, user-friendly tract/bundle statistics generation.
  • Labelmap seeding is not multi-threaded so whole-brain tractography takes.. longer than I would like.
  • Disjointed interface: no one-stop integrated GUI for full DICOMs->tracts->measurements workflow.

Existing NA-MIC resources

Existing open-source external resources

See big list of DTI software here: User:inorton/DTI_Software_List