Convenience wrapper for eigenanatomy decomposition.

Decomposes a matrix into sparse eigenevectors to maximize explained variance. Note: we do not scale the matrices internally. We leave scaling choices to the user.

sparseDecom(
  inmatrix = NA,
  inmask = NULL,
  sparseness = 0.1,
  nvecs = 10,
  its = 5,
  cthresh = 50,
  statdir = NA,
  z = 0,
  smooth = 0,
  initializationList = list(),
  mycoption = 0,
  robust = 0,
  ell1 = 1,
  getSmall = 0,
  verbose = FALSE,
  powerit = 0,
  priorWeight = 0,
  maxBased = FALSE
)

Arguments

inmatrix	n by p input images , subjects or time points by row , spatial variable lies along columns
inmask	optional antsImage mask
sparseness	lower values equal more sparse
nvecs	number of vectors
its	number of iterations
cthresh	cluster threshold
statdir	place on disk to save results
z	u penalty, experimental
smooth	smoothness eg 0.5
initializationList	see initializeEigenanatomy
mycoption	0, 1 or 2 all produce different output 0 is combination of 1 (spatial orthogonality) and 2 (subject space orthogonality)
robust	rank transform input data - good for data checking
ell1	the ell1 grad descent param
getSmall	try to get smallest evecs (bool)
verbose	activates verbose output
powerit	alternative power iteration implementation, faster
priorWeight	scalar weight typically in range zero to two
maxBased	boolean that chooses max-based thresholding

Value

outputs a decomposition of a population or time series matrix

Author

Avants BB

Examples

mat<-replicate(100, rnorm(20))
mydecom<-sparseDecom( mat )
mat<-scale(mat)
mydecom2<-sparseDecom( mat )
# params that lead to algorithm similar to NMF
mydecom3<-sparseDecom( mat, z=1, sparseness=1 )

if (FALSE) {
# for prediction
if ( usePkg("randomForest") & usePkg("spls")  & usePkg('BGLR') ) {
data(lymphoma) # from spls
training<-sample( rep(c(TRUE,FALSE),31)  )
sp<-0.02 ; myz<-0
ldd<-sparseDecom( lymphoma$x[training,], nvecs=5 , sparseness=( sp ),
  mycoption=1, z=myz ) # NMF style
traindf<-data.frame( lclass=as.factor(lymphoma$y[ training  ]),
  eig = lymphoma$x[training,]  %*% as.matrix(ldd$eigenanatomyimages ))
testdf<-data.frame(  lclass=as.factor(lymphoma$y[ !training ]),
 eig = lymphoma$x[!training,] %*% as.matrix(ldd$eigenanatomyimages ))
myrf<-randomForest( lclass ~ . ,   data=traindf )
predlymp<-predict(myrf, newdata=testdf)
print(paste('N-errors:',sum(abs( testdf$lclass != predlymp ) ),
  ' non-zero ',sum(abs( ldd$eigenanatomyimages ) > 0 ) ) )
# compare to http://arxiv.org/pdf/0707.0701v2.pdf
# now SNPs
data(mice)
snps<-quantifySNPs( mice.X, shiftit = TRUE )
numericalpheno<-as.matrix( mice.pheno[,c(4,5,13,15) ] )
nfolds<-6
train<-sample( rep( c(1:nfolds), 1800/nfolds ) )
train<-( train < 4 )
lrmat<-lowrankRowMatrix( as.matrix( snps[train,] ) ,  50 )
lrmat=scale(lrmat)
snpd<-sparseDecom( lrmat-min(lrmat), nvecs=20 , sparseness=( 0.001), z=-1 )
projmat<-as.matrix( snpd$eig )
snpse<-as.matrix( snps[train, ]  ) %*% projmat
traindf<-data.frame( bmi=numericalpheno[train,3] , snpse=snpse)
snpse<-as.matrix( snps[!train, ]  ) %*% projmat
testdf <-data.frame( bmi=numericalpheno[!train,3] , snpse=snpse )
myrf<-randomForest( bmi ~ . , data=traindf )
preddf<-predict(myrf, newdata=testdf )
cor.test(preddf, testdf$bmi )
plot(preddf, testdf$bmi )
} # check for packages
# prior-based example
set.seed(123)
ref<-antsImageRead( getANTsRData("r16"))
ref<-iMath(ref,"Normalize")
mi<-antsImageRead( getANTsRData("r27"))
mi2<-antsImageRead( getANTsRData("r30"))
mi3<-antsImageRead( getANTsRData("r62"))
mi4<-antsImageRead( getANTsRData("r64"))
mi5<-antsImageRead( getANTsRData("r85"))
refmask<-getMask(ref)
refmask<-iMath(refmask,"ME",2) # just to speed things up
ilist<-list(mi,mi2,mi3,mi4,mi5)
for ( i in 1:length(ilist) )
{
ilist[[i]]<-iMath(ilist[[i]],"Normalize")
mytx<-antsRegistration(fixed=ref , moving=ilist[[i]] ,
  typeofTransform = c("Affine") )
mywarpedimage<-antsApplyTransforms(fixed=ref,moving=ilist[[i]],
  transformlist=mytx$fwdtransforms)
ilist[[i]]=mywarpedimage
}
mat=imageListToMatrix( ilist , refmask )
kmseg=kmeansSegmentation( ref, 3, refmask )
initlist=list()
for ( k in 1:3 )
 initlist[[k]]=
   thresholdImage(kmseg$probabilityimages[[k]],0.1,Inf) *
   kmseg$probabilityimages[[k]]
eanat<-sparseDecom( mat,
  inmask=refmask, ell1=0.1,
  sparseness=0.0, smooth=0.5, verbose=1,
  initializationList=initlist, cthresh=25,
  nvecs=3, priorWeight=0.5 )
ee=matrixToImages( eanat$eigenanatomyimages, refmask )
eseg=eigSeg(  refmask, ee )
priormat=imageListToMatrix( initlist, refmask )
cor( t(eanat$eigenanatomyimages), t(priormat) )
plot( ref, eseg )
}