overview of pre-processing pipelines for vbm their impact on spatial normalisation and statistical...
Post on 20-Dec-2015
214 views
TRANSCRIPT
![Page 1: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/1.jpg)
Overview of pre-processing pipelines for VBM
Their impact on spatial normalisation and statistical
analysis
Julio Acosta-Cabronero
Department of Clinical NeurosciencesUniversity of Cambridge
![Page 2: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/2.jpg)
PART I
Skull Stripping
![Page 3: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/3.jpg)
Skull StrippingSingle Strategy
• Brain Extraction Tool v.1/2 (BET/BET2)Deformable model which evolves to fit the brain’s surface
• Brain Surface Extractor (BSE)Edge-based method that employs anisotropic diffusion filtering
![Page 4: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/4.jpg)
• Hybrid Watershed Algorithm (HWA)Combines watershed algorithms and deformable surface models
Skull StrippingHybrid Algorithms
![Page 5: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/5.jpg)
HWA Using Atlas Information
![Page 6: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/6.jpg)
Optimised BSE
![Page 7: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/7.jpg)
Optimised BET
![Page 8: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/8.jpg)
Optimised BET2
![Page 9: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/9.jpg)
Default vs. OptimisedSimilarity (J)
Optimal/DefaultSpecificity Automation
Proc Time
BSE
Default is unreliable, but a systematic method was found to obtain optimised skull-stripped volumes.
0.94/0.50
Very specific. It delineates accurately the brain boundary, but it may remove too much brain tissue. It usually excludes sinuses.
Poor. Running only under Windows (Brainsuite2). Although optimised volumes can be obtained in a simple manner without need of visual inspection.
5 sec
BET
Default is reliable. Optimal f and g (maximising J) are within the ranges [0.4, 0.5] and [-0.1, 0], respectively. It was found empirically that reducing f to 0.4, undesired FN were avoided.
0.94/0.94
Good definition of brain boundary. More conservative than BSE. It usually includes bits of sinuses.
Good. Fixing f to 0.4 (g=0) is a simple method to automate it reliably. Too conservative at times, but it does not usually remove brain tissue.
16 sec
BET2 Same as BET 0.95/0.95 Same as BET Same as BET 27 sec
HWA
Optimisation by using atlas information slightly improved the output volumes.
0.84/0.83Very conservative. It includes big chunks of CSF and sinuses
Very good. No input parameters.
9 min
![Page 10: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/10.jpg)
1.- BET & BET2 are very similar, but BET2 seems to be slightly more accurate
2.- HWA is the most conservative method, but it ensures a very low FN rate and it does not require user intervention
3.- BET & BET2 could be automated in a conservative manner – f=0.4, g=0; although it does not ensures FN rate ~ 0
4.- BSE can be very useful if very specific skull-stripped volumes are required or if high-quality scans are used, otherwise it may undesirably remove essential brain tissue
Skull StrippingSummary
![Page 11: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/11.jpg)
PART II
Bias Correction
![Page 12: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/12.jpg)
Intensity Non-Uniformity (r.f. Bias) Correction
Locally-Adaptive Methods
• Non-parametric Non-uniform Intensity Normalisation (N3)Iterative modelling of low-frequency spatial variations in the data to maximise high-frequency information in the intensity histogram of the corrected volume.
• Bias Field Corrector (BFC)It also utilises an approach based on normalisation of regional tissue intensity histograms to global values.
![Page 13: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/13.jpg)
Optimised BET2
Optimised BET2 + N3
![Page 14: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/14.jpg)
eRMS BSE BET BET2 HWA
N3 4.29 4.32 4.26 4.34
BFC 4.60 5.67 5.51 7.34
Bias CorrectionPhantom Work
1.- N3 outperformed BFC
2.- N3 performs similarly for all skull-stripping methods
![Page 15: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/15.jpg)
PART III
Unified Segmentation
![Page 16: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/16.jpg)
Unified SegmentationPhantom Work – Standard Space
Method J FN (%) FP (%) N
Full Volume 0.84 9.0 8.3 -0.7
HWA + N3 0.80 8.3 15.2 6.9
BSE + N3 0.90 6.4 4.4 -2.0
BET2, f=0.4 + N3 0.86 5.5 9.5 3.9
BET2, f=0.5 + N3 0.89 4.8 6.9 2.1
![Page 17: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/17.jpg)
PART IV
Statistical Analysis
![Page 18: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/18.jpg)
Statistical AnalysisArtificial Lesion
![Page 19: Overview of pre-processing pipelines for VBM Their impact on spatial normalisation and statistical analysis Julio Acosta-Cabronero Department of Clinical](https://reader030.vdocuments.us/reader030/viewer/2022032800/56649d4c5503460f94a29f46/html5/thumbnails/19.jpg)
Statistical AnalysisArtificial Lesion