deepPatchMatch: deepPatchMatch patch match two images with deep features

In stnava/patchMatchR: Local patch matching with flexible transformation models in n-dimensional images

deepPatchMatch patch match two images with deep features

Description

High-level function for deep patch matching that makes many assumptions and therefore minimizes the number of parameters the user needs to choose.

Usage

deepPatchMatch(
  movingImage,
  fixedImage,
  movingImageMask,
  fixedImageMask,
  movingPatchSize = 32,
  fixedPatchSize = 32,
  knn = 1,
  knnSpatial = 0,
  featureSubset,
  block_name = "block2_conv2",
  switchMatchDirection = FALSE,
  kPackage = "RcppHNSW",
  vggmodel,
  subtractor = 127.5,
  patchVarEx = 0.95,
  meanCenter = FALSE,
  initialMovingTransform,
  verbose = FALSE
)

Arguments

movingImage	input image from which we extract patches that are transformed to the space of the fixed image
fixedImage	input image that provides the fixed reference domain.
movingImageMask	defines the object of interest in the movingImage
fixedImageMask	defines the object of interest in the fixedImage
movingPatchSize	integer greater than or equal to 32.
fixedPatchSize	integer greater than or equal to 32.
knn	k-nearest neighbors ( should be >= 1 )
knnSpatial	k-nearest neighbors for spatial localization (optional). this will constrain the search to more proximal locations. will perform better if the images are in the same physical space. currently, the units for the spatial distance is in voxels. may add physical space option later. FIXME - allow a transformation to be passed to this step s.t. moving points can be transformed to fixed space before distance assessment.
featureSubset	a vector that selects a subset of features
block_name	name of vgg feature block, either block2_conv2 or integer. use the former for smaller patch sizes.
switchMatchDirection	boolean
kPackage	name of package to use for knn
vggmodel	prebuilt feature model
subtractor	value to subtract when scaling image intensity; should be chosen to match training paradigm eg 127.5 for vgg and 0.5 for resnet like.
patchVarEx	patch variance explained for ripmmarc
meanCenter	boolean to mean center each patch for ripmmarc
initialMovingTransform	optional initial transformation to moving image
verbose	boolean

Value

correspondence data

Author(s)

Avants BB

Examples

library( keras )
library( ANTsR )
nP1 = 5
nP2 = 20
psz = 32
img <- ri( 1 ) %>% iMath( "Normalize" )
img2 <- ri( 2 ) %>% iMath( "Normalize" )
mask = randomMask( getMask( img ), nP1 )
mask2 = randomMask( getMask( img2 ), nP2 )
match = deepPatchMatch( img2, img, mask, mask2 )
## Not run: 
library( ANTsR )
img <- ri( 1 ) %>% iMath( "Normalize" ) %>% resampleImage( c( 2, 2 ) )
nP1 = 10
nP2 = 40
psz = 32
mask = randomMask( getMask( img ), nP1 )
features = deepFeatures( img, mask, patchSize = psz )
img2 <- ri( 5 ) %>% iMath( "Normalize" ) %>% resampleImage( c( 2, 2 ) )
txStretch = createAntsrTransform( "AffineTransform", dim=2 )
params = getAntsrTransformParameters( txStretch )
params[1] = 0.8
setAntsrTransformParameters(txStretch, params)
cos45 = cos(pi*45/180)
sin45 = sin(pi*45/180)
txRotate <- createAntsrTransform( precision="float", type="AffineTransform", dim=2 )
setAntsrTransformParameters(txRotate, c(cos45,-sin45,sin45,cos45,0,0) )
setAntsrTransformFixedParameters(txRotate, c(128,128))
rotateFirst = composeAntsrTransforms(list(txStretch, txRotate))
# img2 = applyAntsrTransform(rotateFirst, img2, img2)
mask2 = randomMask( getMask( img2 ), nP2 )
match = deepPatchMatch( img2, img, mask2, mask, 64, 64 )

for ( k in 1:nrow( match$matches ) ) {
  if ( ! is.na( match$matches[k,1] ) ) {
    layout( matrix(1:2,nrow=1) )
    plot( as.antsImage( match$ffeatures$patches[k,,] ) )
    plot( as.antsImage( match$mfeatures$patches[match$matches[k,1],,] ) )
    print( k )
    print( match$ffeatures$patchCoords[k,] )
    print( match$mfeatures$patchCoords[match$matches[k,1],] )
    Sys.sleep(1)
    }
}

## End(Not run)

stnava/patchMatchR documentation built on March 23, 2022, 6:47 a.m.