validation of dti analysis
DESCRIPTION
Validation of DTI Analysis. Guido Gerig, Clement Vachet, Isabelle Corouge, Casey Goodlett UNC Chapel Hill. Validation Concepts. Validity: Comparison to truth / gold standard Reliability, Reproducibility: Repeated measures Outcome measures: Sufficient to answer clinical question. - PowerPoint PPT PresentationTRANSCRIPT
NA-MICNational Alliance for Medical Image Computing http://na-mic.org
Validation of DTI Analysis
Guido Gerig, Clement Vachet, Isabelle Corouge, Casey Goodlett
UNC Chapel Hill
National Alliance for Medical Image Computing http://na-mic.org
Validation Concepts
• Validity: Comparison to truth / gold standard
• Reliability, Reproducibility: Repeated measures
• Outcome measures: Sufficient to answer clinical question
National Alliance for Medical Image Computing http://na-mic.org
Validation: DTI Image Analysis
• Validity:– In vivo human imaging: Truth not known– Ex vivo pathological samples: tissue dynamics not
preserved– Animal imaging:
• Hi-res DT-MRI with animal scanner• Possibility to induce pathology: lesions, de-myelination
• Goldish or Silver Standard: Phantom– Nylon fibers embedded in water, interweaved to
simulate crossings: Far from simulation of axons and myelination sheaths
– Computational phantom: Simulated diffusion imaging?– Others?
National Alliance for Medical Image Computing http://na-mic.org
Validation: DTI Image Analysis
• Reliability, reproducibility:– Reproducibility of scanning: Repeated scans single or
multi-site– Reproducibility of image analysis and/or individual steps
• Key Problems:– Metric to determine differences– Imaging and image analysis come with large set of
parameters– Validation of single processing steps versus validation
of whole system– Relation of validation metric to clinical question and
group findings
National Alliance for Medical Image Computing http://na-mic.org
Very preliminary attempts
• Repeatability of quantitative tractography-based FA/MD (ISMRM 2005)
• Tool for geometric fidelity of fiber tract ROI (Casey Goodlett)
National Alliance for Medical Image Computing http://na-mic.org
Validation: Fiber Tracking
• Goal: Evaluate the reproducibility of quantitative tractography (FiberViewer)
• Study: – 6 scans of the same patient with small motion
(MacFall, Duke)– 3 callosal fiber tracts
• Method: – Fiber Extraction
• Selection of a ROI on one scan• Registration for the 5 other scans• Tracking of Fibers
– Analysis across the 6 scans• Diffusion properties: FA, ADC, λ1, λ2, λ3
6 repeated scans DTI average
National Alliance for Medical Image Computing http://na-mic.org
Validation: 6 repeated scans
Scan1
Scan2…
T B01B02
… Scan6
DTI Average
T B01B06
Extraction
Extraction
Extraction
Scan 2…
…Scan 6
DTI Average
Selection of a ROI
Registration of ROI
National Alliance for Medical Image Computing http://na-mic.org
Scans' Comparison: Number of Fibers (ROI:essai3-th69)
0
20
40
60
80
100
120
140
160
180
Dif 1 Dif 2 Dif 3 Dif 4 Dif 5 Dif 6 Average
Nu
mb
er
of
Fib
ers
Number of “Fibers” (streamlines): Really not meaningful
DTI Average
The fallacy of #”fibers”
National Alliance for Medical Image Computing http://na-mic.org
Tract-based diffusion propertiesScans' Comparison: FA (ROI: essai3-th69)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
FA
Mean Mean+std Mean-std
Scans' Comparison: ADC (ROI: essai3-th69)
0
2
4
6
8
10
12
14
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
AD
C
Mean Mean+std Mean-std
Statistics across 6 repeated scans:
Curves of MeanFA and MeanMD, with Standard Deviation
FA
FA
MD
National Alliance for Medical Image Computing http://na-mic.org
Tract-based diffusion propertiesStatistics across 6 repeated scans: eigenvalues
1Scans' Comparison: Lambda2 (ROI: essai3-th69)
0
2
4
6
8
10
12
14
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
La
mb
da
2
Mean Mean+std Mean-std
Scans' Comparison:Lambda3 (ROI: essai3-th69)
0
2
4
6
8
10
12
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
Lam
bda2
Mean Mean+std Mean-std
Scans' Comparison: Lambda1 (ROI: essai3-th69)
0
5
10
15
20
25
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
Lam
bda1
Mean Mean+std Mean-std
2 3
National Alliance for Medical Image Computing http://na-mic.org
Tract-based diffusion properties
FA
Scans' Comparison: FA (ROI: essai3-th69)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
FA
Mean-FA DTIAverage-FA
Scans' Comparison: ADC (ROI: essai3-th69)
0
2
4
6
8
10
12
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
AD
C
MeanADC DTIAverage-ADC
FA
MD
Curves of MeanFA and MeanMD in comparison to the Average DTI
National Alliance for Medical Image Computing http://na-mic.org
Scans' Comparison: Lambda1 (ROI: essai3-th69)
0
2
4
6
8
10
12
14
16
18
20
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
Lam
bd
a1
Mean-Lambda1 DTIAverage-Lambda1
Scans' Comparison: Lambda2 (ROI: essai3-th69)
0
2
4
6
8
10
12
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
Lam
bda2
Mean-Lambda2 DTIAverage-Lambda2
λ1
λ2
λ3
Scans' Comparison:Lambda3 (ROI: essai3-th69)
0
1
2
3
4
5
6
7
8
9
10
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
Points
Lam
bda2
Mean-Lambda3 DTIAverage-Lambda3
Tract-based diffusion properties
Curves of mean eigenvalues in comparison to the eigenvalues of the average DTI
National Alliance for Medical Image Computing http://na-mic.org
Quantitative Results
Table: Results at the center, i.e. at the position of the midsagittal plane, are listed. The second row lists the number of streamlines obtained for each experiment. Although the number of streamlines varies due to the instability of tractography, the resulting diffusion tensor statistics are approximately within 5%std for FA and MD but only 2%std for the first eigenvalue λ1. The FA of the DTI average (last column) is slightly lower than the 6 case mean.
Guido Gerig , Isabelle Corouge, Clement Vachet, Ranga Krishnan and James MacFall, Quantitative Analysis of Diffusion Properties of White Matter Fiber Tracts: A Validation Study, International Society of Magnetic Resonance ISMRM, May 2005 (peer reviewed long abstract)
National Alliance for Medical Image Computing http://na-mic.org
Geometric reproducibility of tracts
• How to compare geometry tractography results
• Proposed method (BIRN): Map tractography result to binary voxel map, binary overlap test (Dice, kappa stats))– very low enthusiasm
• Brainstorming with Casey Goodlett:– Streamline tracking unstable, ill-posed– Key question: Correspondence?
National Alliance for Medical Image Computing http://na-mic.org
Work in progress
• Distance measure between sets of bundles
• Give up on point-to-point correspondence between streamlines
• Solution/Tool: Measure geometric distance of points in bundle A to closest point in bundle B
• Measure is non-symmetric
National Alliance for Medical Image Computing http://na-mic.org
Validation: Tract Comparison
Comparison of scan 1 with original and modified tracking method
National Alliance for Medical Image Computing http://na-mic.org
Validation: Tract Comparison
Comparison of scan 1 with scan 4
National Alliance for Medical Image Computing http://na-mic.org
Conclusions
• NAMIC could develop concept, strategy and tools for validation of DTI
• Validation of reproducibility of diffusion properties (tensor stats)
• Validation of geometry of extracted bundles: Metric for bundle distance
• ….