Presentation on theme: "Detecting Subtle Changes in Structure"— Presentation transcript:
1Detecting Subtle Changes in Structure Chris RordenVoxel Based MorphometrySegmentation – identifying gray and white matterModulation- adjusting for normalization’s spatial distortions.Diffusion Tensor ImagingMeasuring white matter integrityTractography and analysis.Many images are from Christian Gaser. You can see his presentations and get his VBM scripts from these sites:fmri.uib.no/workshops/2006/mai/fmri/index.shtmldbm.neuro.uni-jena.de/home/
2Voxel Based Morphometry Most lectures in course focus on functional MRI.However, anatomical scans can also help us infer brain function.Do people with chronic epilepsy show brain atrophy?Which brain regions atrophy with age?Do people with good spatial memory (taxi drivers) have different anatomy than other people?Voxel based morphometry is a tool to relate gray and white matter concentration with medical history and behavior
3Morphometry examines the shape, volume and integrity of structures. Classically, morphometry was conducted by manually segmenting a few regions of interest.Voxel based morphometry conducts an independent statistical comparison for each voxel in the brain.Images from Christian Gaser
4Voxel Based Morphometry VBM has some advantages over manual tracing:Automated: fast and not subject to individual bias.Able to examine regions that are not anatomically well defined.Able to see the whole brainNormalization compensates for overall differences in brain volume, which can add variance to manual tracing of un-normalized images.
5VBM disadvantages VBM has clear disadvantages Crucially depends on accurate normalization.Low power: gray matter random fields are very heterogenous (individual patterns of sulcal folding registration is always poor.Crucially depends on a priori probability maps.Assumes normal gray-white contrast. Focal Cortical DysplasiaLooks for differences in volume, can be disrupted if shape of brain is different: problem for developmental disorders
6SegmentationStart with high quality MRI scanClassify tissue types (gray matter in this example)
7Partitioning Tissue Types VBM segments image into three tissue types: gray matter, white matter and CSF.Typically done on T1 scans (best spatial resolution, good gray-white contrast).Only three tissue types: will not cope with large lesions.Probability map: each voxel has a % chance of being one of the 3 tissue types.T1graywhiteCSFImages from Christian Gaser
8Segmentation I: Image Intensity estimate for GMp=0.95frequencyp=0.05Image intensityWMback-groundGMCSFImages from Christian Gaser
9Segmentation II: Voxel location Maximization of a posteriori probability: Bayesian approach (expectation maximization)Analogy:We know that last year there were 248 of 365 days with rain in Norway (p=0.68)the conditional (or posterior) probability for rain in Bergen will be p>0.5T1WMGMCSFProbability maps (n=152)Images and text from Christian Gaser
10Segmentation overview Intensity based estimate for GMSource ImageFinal resultp=0.95p=0.05p=0.95p=0.90p=0.05p=0.95a priori GM map
12Homogeneity correction crucial Field inhomogeneity will disrupt intensity based segmentation.Bias correction required.no correctionT1WMGMEstimate
13Normalization is crucial Poor normalization has two problemsImage will not be registered with a priori map = poor segmentation.Images from different people will not be registered: we will compare different brain areas.Custom template and prior is usefulAccounts for characteristics of your scanner.Accounts for characteristics of your population (e.g. age).Must be independent of your analysis:Either formed from combination of both groups (control+experimental) or from independent control group.
14Two step segmentation segmentation I segmentation II Step I: Creation of customized templatesegmentation Isegmentation IInorma-lizationaveragingStep II:Optimized segmentationcustomized templatenorma-lizationMNI template
16Overview of ‘Optimized VBM’ segmented Inormalizedsegmented IImaskedsmoothedcustomized templatemask
17SPM5 segmentation Unified segmentation Iterated steps of segmentation estimation, bias correction and warpingImpactWarping of prior images during segmentation makes segmentation more independent from size, position, and shape of prior imagesmuch slower than SPM240 iterationssegmentation40 iterationsbias correction20 iterationswarpingsignificant change of estimateno significant change of estimate
18SPM8 new segmentation SPM8’s default segmentation similar to SPM5 There is a hidden new segmentationFrom Graphics window, Choose Tasks/SPM/Tools/NewSegmentAllow multiple channels, for example combine T1 and T2 forbetter tissue classificationIncludes more tissue types (gray, white, CSF, bone, other soft tissue, air.Not as finicky with regards to starting estimate
19Building a better template: data 25 participants scannedHigh resolution: 0.85mm3 vs 1mm3Large FOV: 320x320 matrix vs 256x256 matrixBoth T1 and T2-weighted images for each person (with same orientation and coverage)Uses revolutionary 3D SPACE sequence that provides high resolutionT2 has opposite contrasts for CSF (bright) and bone (dark)T2
20Building a better template: normalization Subj1SPM8 affine coregistration to MNI space.ANTS used to generate mean image (25 Core i7 * 70 hours)Mean image normalized to MNI space using SPM8 ‘new seg’.Subj2Mean
21Building a better template: segmentation Crude TPM created:Default newseg TPM for brain caseNeck region generated from mean imageRefined TPM created:Each individual’s T1+T2 scans segmented using newsegTPM created from mean segmentationTPM manually edited to remove errorsTPM smoothed by 2mm FWHMRepeat previous steps four times.TPM : CSF
22Building a better template: results Our new template accurately segments tissuesIncludes neck regionsWorks well, even if only provided with a T1 scanWe no longer need a CT scan for bone, or even a T2 scan…Single subject segmentation
23Building a better template: future MUSC ‘brain attack’ CT protocol:identify people with suspected acute stroke who did not have a stroke.People are identified by neurologist who ensures no subsequent brain injury or other structural abnormalities
24Voxel Based Morphometry We can statistically analyze gray matter atrophyEpilepsy
25This will reduce ability to detect differences Segmentation ProblemIf someone has atrophy, normalization will stretch gray matter to make brain match healthy template.This will reduce ability to detect differencesNormalization will squish this regionNormalization will stretch this region
26Image ModulationAnalogy: as we blow up a balloon, the surface becomes thinner. Likewise, as we expand a brain area it’s volume is reduced.Without modulationSourceTemplateModulated
27Image ModulationOptimized Segmentation can adjust for distortions caused during normalization.Areas that had to be stretched are assumed to have less volume than areas that were compressed.Corrects for changes in volume induced by nonlinear normalizationMultiplies voxel intensities by a modulation matrix derived from the normalization stepAllows us to make inferences about volume, instead of concentration.
28VBM and developmental syndromes Williams SyndromeDevelopmental syndrome: Chromosome 7Manual Morphology shows8-18% decrease in posterior GM/WMMost consistent finding is reduced intra-parietal sulcus depth and superior parietal lobe volume (see figure)Relatively preserved frontal GM/WMCreates unique shapeUnique spatial distribution of gross volume loss influences VBM results depending on whether modulation is usedControl WSEckert et al. 2006b,c
29Shape differences influence modulated data. Modulation and shapeShape differences influence modulated data.Deformation Based Morphometry can identify shape/gross volumetric differences.Eckert et al., 2006a
30Modulation is optional and controversial Modulation will smooth images, specificity will decreaseAlternatively, you can covary overall brain volume by including GM or GM+WM as nuisance regressor.Example showing danger of modulation. This image comes from an elderly participant, with relatively large ventricles. Normalization adjusts ventricle size, but the deformations are spatially smooth, so tissue near the ventricles (e.g. caudate) are also being spatially compressed.[Deformations exaggerated for exposition]
31DBM (from Henson)Deformation-based Morphometry examines absolute displacements.E.G. Mean differences (mapping from an average female to male brain).
32New methods can complement VBM. Cortical ThicknessNew methods can complement VBM.Freesurfer’s cortical thickness is powerful tool.Requires very good T1 scans.Modulated VBMFreesurferAge-related declines ingray matter volume and cortical thickness
33VBM comments Longitudinal VBM: Cross sectional studies Sensitive way to detect atrophy through time. Using the same individual reduces variability.Cross sectional studiesCan compare two distinct populationsCan also examine atrophy through time, though will require more people than longitudinal VBM.VBM findings are first step in understanding structural changes.Bookstein, 2001Davatzikos, 2004
34Diffusion Weighted Imaging T1/T2 scans do not show acute injury. Diffusion weighted scans do.DW scans identify areas of permanent injuryMeasures random motion of water molecules.In ventricles, CSF is unconstrained, so high velocity diffusionIn brain tissue, CSF more constrained, so less diffusion.T2DW
35Diffusion Tensor Imaging (DTI) DTI is an extension of DWI that allows us to measure direction of motion.DTI allows us to measure both the velocity and preferred direction of diffusionIn gray matter, diffusion is isotropic (similar in all directions)In white matter, diffusion is anisotropic (prefers motion along fibers).
36DTIThe amount of diffusion occurring in one pixel of a MR image is termed the Apparent Diffusion Coefficient (ADC) or Mean Diffusivity (MD).The non-uniformity of diffusion with direction is usually described by the term Fractional Anisotropy (FA).MD differsFA differs
37What is a tensor? A tensor is composed of three vectors. Think of a vector like an arrow in 3D space – it points in a direction and has a length.The first vector is the longest – it points along the principle axis.The second and third vectors are orthogonal to the first.Sphere: V1=V2=V3Football:V1>V2V1>V3V3 = V2???:V1>V2>V3
38Diffusion Tensor Imaging To create a tensor, we need to collect multiple directions.Typically directions.More directions offer a better estimate of optimal tensor.