2008 Winter Project Week Tractography Meeting Notes

Brief Summary

"We know when the track is bad (or ugly) but we don't know when it is good." CF Westin

1. All sites could analyze data
2. All sites could generate all requested tracks
3. ROIs need work
4. Ugly R fornix

Action Plans

1. Upload the DTI registered (using FSL's FLIRT) Freesurfer generated white matter mask in nrrd format at same resolution as DTI data for each subject (Anastasia, Sylvain, bug-reporter as needed).
2. All sites analyze full n = 9 test/retest data set and sent agreed upon metrics to Sonia (date?)

Discuss now then move agreed upon plan up to Action Plan

1) ROI definitions and impact on different methods

• Force all methods to use the same ROIs or branch and have two or more classes of optimal ROIs?
• For streamline methods larger ROIs than the tract itself good
• For shortest path methods, connections stop at ROI
• Role of DTI Atlas
• Missing arcuate fasiculus

Summary:

• We should all start with the same ROIs, and then allow them to be adapted to specific method, as long as that process is well-documented. Report back the modified/new ROI.
• Need to add end-points

• Expert drawing?

2) Debugging

Create/use existing synthetic tensor data set (Gordon's helix, crossing fiber situation). Good for debugging algorithm method comparison. Nathan Hageman w/input from other algorithm teams will generate helical synthetic data set. Parameters here: Who will upload to SRB? What outcome metrics from it will each team generate?

3) Review and define metrics to compare results, which methods to compare?

Current metrics: tensor invariants (FA, trace, mode, norm), volume, length.

What is a good fiber tract?

Easier question: How similar are two methods?

Parameter settings within a method may change the results comparable to the difference between methods.

Should we compute more complex statistics of fiber track, e.g. FA measure in spatial distribution?

• 12 different tracts
• Tracts voxelized -> volumetric result from tracts
• Missing parts of a tract, how should be penalized?
• Statistics along the tracts, should we sum along the tract to get one measure?
• Distributions across the tract
• Atlas (Casey)

• Bundle thickness: mix of radius of tract and uncertainty
• How to define volume from connectivity map (shortest path, or stochastic)
• Quantification of registration error and its propagation forward and impact on final results

1. Agree on units for volume measures. Suggest cubic centimeters (ml) vs mm3

How to approach Test- Retest data

Start discussion

Milestones and dates for next steps in project

Registration &/or resampling to support visit 1- visit 2 analysis and group analysis

Should we make plans for a NAMIC sponsored MICCAI event: Tractography Grand Challenge e.g. the one done last year for segmentation? See http://mbi.dkfz-heidelberg.de/grand-challenge2007/)

4:00- 5:00 PM NAMIC DTI pipeline discussion

• Registration (we used FLIRT for DTI to structural MR)
• EPI distortion correction (we used FUGUE), ITK?

Notes

From Casey's presentation: Could do tractography for all tracks. Made suggestion to . His Atlas presentation brought up idea of using one set of ROI specifications for an entire study in atlas space and map them back to the individual.

From John's presentation: His method required that he had to shrink the provided ROIs for uncinate fasciulus and corpus callosum to make his method work. Then could generate fiber volumes for each tract. His tracts all STOP at ROIs (end points). If given ROIs farther apart could generate longer tracts. Needed a white matter mask- created his own with Tri by thresholding FA (>0.15). Discussion regarding whether the ROIs based on Susumu Mori's criteria would actually be optimal or even work for every method.

From Tri's presentation: Able to generate all tracks with given ROIs. Needed white matter mask (see above). 30 minutes/tract on 63GB computer. Seed with source, filter with sink. Had to cut uncinate fasciulus. For later statistical analysis, considered only non-zero connectivity value voxels. Reminds us that bundle thickness is related to both uncertainty and physical radius Internal capsule, why so much blue?

From Vince Magnotta's presentation: GTRACT fine for forceps major and minor, internal capusule and Fast Marching algorithm better for fornix, hippocampus-cingulum and cingulum. For the Fast Marching algorithm, modified the algorithm to use cost image generations based on tensor and anisotropy, gradient descent. His algorithm would perform "better" for core of cingulum bundle, needs a smaller, more focal ROI. For uncinate, couldn't get it to bend around, would need intermediate ROIs.

Sonia's preliminary cross-algorithm study:

1. rank order of qualitative accuracy (strongly underscores need to clarify the clinical goal of the tractography results)

From Carlo:

1. Commended us for making effort against this tough problem and for the decision to control pre-processing steps
2. What is the impact of adjusting parameters within an algorithm on the outcome metrics? May need to explore this in addition to the cross algorithm assessments.
3. What is the effect of the tissue properties that surround the track on results within and across algorithm.
4. Differences due to the topology, how to visualize? Binary masks in/out of track volume.

Back to 2008_Winter_Project_Week_Tractography, AHM_2008, Events