ANTs-based processing of AD PET data

Overview of ADNI and Amyloid PET

The Alzheimer’s disease neuroimaging initiative (ADNI)¹ collects and shares clinical and putative biomarker² data for AD in relation to normal aging. Ideally, these biomarkers would be suggestive of not only the presence of AD pathology, but also the (Braak) stage of that pathology:

Braak stages for amyloid and tau

Only two biomarkers in ADNI are specific to amyloid pathology: florabetapir PET and cerebrospinal fluid measurements³ both of which relate to molecular scale processes. AV45 binds primarily to the \(\beta\) sheet structure of fibrillar amyloid with high affinity and also exhibits favorable pharmacokinetics⁴. AV45 is specific in that it does not bind to \(\tau\) or \(\alpha\)-synuclein (present in ALS or Parkinson’s disease). Note that these amyloid biomarkers do not always agree⁵ but this is likely due to the increased sensitivity of CSF at earlier disease stages. Note: “patients in ADNI1 initially received PIB scans instead of Florbetapir scans. However, the protocol was amended before the conclusion of the study to include the adoption of Florbetapir scans over PIB scans due to processing time limits.”

AV45 PET imaging is a validated method for estimating the likelihood of pathology that can only be verified post mortem. Both visual read and SUVR based on a composite region of interest (Avid methodology) show high sensitivity and specificity in a moderately-sized dataset for detecting neuropathology.

⁶

Interpreting ADNI biomarkers for AD

ADNI is far from definitive. Perhaps the most important concern with ADNI is that we lack ground truth for not only pathology but also in vivo disease stage (a fundamental problem in the field). Staging is not just a question of presence but also of rate i.e. time to conversion. Amyloid pathology may be present without incipient clinical symptoms (up to 25 years in familial AD)⁷. The ADNI cohort is “super-normal”, i.e. not representative of the general population. Furthermore, ADNI data quality ⊕compromised by sequence availability may not always reach “gold-standard” for that methodology, e.g. DTI or ASL acquisitions and anatomical labeling protocols. ⊕Freesurfer is a marginal performer wrt anatomical labeling ADNI also lacks an explicit measurement of the distribution of \(\tau\) pathology⁸ (another general problem in the field). Lastly, the ADNI SUVR recommendations are not validated against pathology but instead performance is inferred across different models. That being said, ADNI is the most mature and comprehensive resource in the USA for investigating in vivo predictors of AD pathology.

Only neuropathological evaluation can confirm AD. CSF and PET provide surrogate measures. The “gold standard” interpretation of the amyloid PET is currently the visual (consensus) read in which the viewer(s) compares uptake in several different regions. The SUVR measurement is a “silver standard” in comparison to visual read but both diagnostics must reference post-mortem in validation studies. ⊕confirmation of an AD biomarker requires post-mortem data examined with immunohistochemistry or related methods

The first Avid AV45 study showed 100% specificity and 93% sensitivity for detecting AD pathology (28 of 29 cases correct, 14 of 15 with AD pathology and \(+\) PET scans). In contrast, clinical diagnosis misclassified 3 subjects. Interestingly, SUVR thresholded at 1.1 appears to yield perfect classification (cf. Table 1, Clark et al.). ⊕35 patients, 6 for training, 74 controls (all negative), 96 percent accuracy

SUVR seeks to compare diseased with normal brain regions by taking a ratio of uptake values. It does not mimic visual read.

\[ \text{SUVR}^{JagustADNI} = \frac{\text{count}_\text{composite region}}{\text{count}_\text{cerebellum}} \]

This simple equation may be more complex than it appears. Assuming a PET acquisition exists (which is a spatiotemporal image), we perform the following steps to enable SUVR calculation:

• motion correction;

• spatiotemporal smoothing;

• registration (or mapping) to brain structure within subject and to a group template brain ⊕canonical templates provide a labeled brain-centric coordinate system;

• segmentation of tissues and parcellation of neuroanatomy;

• a hypothesis about what parts of the brain should be considered “normal” (denominator) versus those that should be considered “diseased” (numerator).

The Jagust lab processing for ADNI recommends a cutoff of 1.11 based on the following steps:

• spatiotemporal smoothing with an 8mm (?) FWHM kernel;

• registration with SPM5 to brain structure within subject and to a group template brain with Freesurfer;

• segmentation of tissues and parcellation of neuroanatomy with Freesurfer;

• recommend cerebellum as control region and a composite region of interest comprising frontal, lateral parietal, lateral temporal and anterior and posterior cingulate cortices in subject space (not template space);

AD Regions

Compare to template space:

if ( file.exists( path.expand( "~/transforms/ch2.nii.gz" ) ) )
  mni = antsImageRead( path.expand( "~/transforms/ch2.nii.gz" ) ) else mni = antsImageRead( getANTsRData( "ch2" ) )
ch2 = antsImageRead( "./adniTemplate/T_template0.nii.gz" )
ch2seg = antsImageRead( "./adniTemplate/antsMalfLabeling_6Labels.nii.gz" )
ch2a = antsImageRead( "./adniTemplate/antsMalfLabeling.nii.gz"  )
invisible( plot( ch2 , domainImageMap=mni ) )

invisible( plot( ch2, ch2seg, domainImageMap=mni  ) )

invisible( plot( ch2, ch2a , domainImageMap=mni ) )

The recommended SUVR values are provided in a variable called SUMMARYSUVR_WHOLECEREBNORM in ADNI databases.

How do we interpret ADNI SUVR? We might take these values as a probability estimate for having Alzheimer’s disease pathology. For reference:

ADNI histogram

ADNI plot

shows the histogram and a plot of SUVR in ADNIGO/2 controls. Note the presence of “borderline” cases. How should we think about these subjects on the scale of AD progression?⁹

In a recent paper, Kandel, et al¹⁰ used SUVR threshold as a binary indicator variable for the presence of pathology and showed that white matter hyperintensities jointly predict this diagnostic variable along with APOE status (multivariate \(p < 0.01\) for both predictors). However, it is possible that these values may be related to cerebral amyloid angiopathy or other past or present comorbidities. That is, amyloid depositions appearing in AV45 may be complementary to or even independent of Alzheimer’s disease.¹¹ Thus, for several reasons, care must be taken when using AV45 SUVR as a surrogate ground truth.

Calculating SUVR from ADNI data

We assume that you ran antsCorticalThickness.sh on your subject data while employing the adniTemplate as your reference space.

library( "ANTsR" )
data( DesikanKillianyTourville )
labelData = DesikanKillianyTourville
bd = path.expand( "./elderly/" )
myid = "0001"
mydt = "19451231"
outname <- paste( bd ,myid,'_',mydt,sep='')
template = antsImageRead( "./adniTemplate/T_template0.nii.gz" )
malf = antsImageRead( "./adniTemplate/antsMalfLabeling.nii.gz" )
segimg = Sys.glob(  paste( outname, "*BrainSegmentation.nii.gz", sep=''))[ 1 ]
segimg = antsImageRead( segimg )
t1img = Sys.glob(  paste( outname, "*BrainSegmentation0N4.nii.gz", sep=''))[ 1 ]
t1img = antsImageRead( t1img ) #  * thresholdImage( segimg, 1, 6 )
temaff = Sys.glob(  paste( outname, "*SubjectToTemplate0GenericAffine.mat" ,sep=''))[ 1 ]
if ( length( temaff ) > 0 ) {
  temwrp = Sys.glob( paste( outname, "*SubjectToTemplate1Warp.nii.gz",sep=''))[ 1 ]
  itemwrp = Sys.glob( paste( outname, "*TemplateToSubject0Warp.nii.gz",sep=''))[ 1 ]
}
txc2mni = temaff # c( temwrp, temaff ) # warps too large for github
mni2txc = temaff # c( temaff, itemwrp )
malf2sub = antsApplyTransforms( t1img, malf, transformlist = mni2txc, 
                                interpolator="NearestNeighbor", 
                                whichtoinvert=c(TRUE) )
malfCTXsub = malf2sub * thresholdImage( segimg, 2, 2 )

petfn = Sys.glob(  paste( outname, "*av45pet.nii.gz", sep='') )
fdgimg = antsImageRead( petfn )

pofn = paste( outname, "_AV45SUVR.nii.gz", sep='' )
print( paste( "begin", pofn ) )

## [1] "begin ./elderly/0001_19451231_AV45SUVR.nii.gz"

fdgProc = petSUVR( fdgimg, t1img, segimg, smoothingParameter = 2.5,  labelValue = 6, noMotionCorr=T )
# fdg2template = antsApplyTransforms( template, fdgProc$petSUVR, transformlist = txc2mni )
# plot( fdg2template, malfCTXsub, axis=1 )
# antsImageWrite( fdg2template, pofn )

Display the T1 image.

plot( t1img  )

## NULL

Display the amyloid data overlaid on the T1 image.

plot( t1img, fdgProc$petSUVR, window.overlay=c( 0.9, max( fdgProc$petSUVR ) ) )

## NULL

Display the cortical segmentation and labeling on the amyloid data.

plot( fdgProc$petSUVR, malfCTXsub )

## NULL

Now, we compute amyloid load.

# compute amyloid load
data( DesikanKillianyTourville )
labelData = DesikanKillianyTourville
adregions = c(
  grep( "inferior temporal", labelData$label_name ),
  grep( "osterior cing", labelData$label_name ),
  grep( "anterior cing", labelData$label_name ),
  grep( "angular gyrus", labelData$label_name ),
  grep( "middle frontal gyrus", labelData$label_name ),
  grep( "superior frontal gyrus", labelData$label_name ) )
locmask = thresholdImage( malfCTXsub, 1, Inf)
labmat = labels2matrix( malfCTXsub, locmask, labelData$label_num[adregions], missingVal=0 )
labmat = labmat / rowSums( labmat )
proj = t( labmat %*% fdgProc$petSUVR[ locmask == 1 ] )
colnames( proj ) = paste( 'ADREGIONS', labelData$label_name[ adregions ], sep="_" )
finalval = mean( proj )

The composite SUVR value is 1.35186 which indicates that the subject should be considered \(A\beta^+\).

Conclusions

• SUVR is a principled method for quantifying likelihood of amyloid pathology.

• AV45 is specific but may not exclusively detect AD-related amyloid pathology (concerns of sensitivity in broader populations? Do we truly understand normal variability in/causes for amyloid deposition in the brain?).¹²

• AV45 is more specific than measures such as hippocampus volume and less subjective and more repeatable than commonly used cognitive measurements such as CDR.

• Amyloid pathology does not imply near-term cognitive decline.

• Some caveats to using ADNI:
  • the super-normality of the cohort;
  • no clear way to define time to conversion (but see ADNI2);
  • lack of post-mortem validation for ADNI diagnostics;
  • many complex processing choices are opaque to the user of the summary measurements.

Despite these issues, ADNI is a great dataset. Enjoy!