BRAINSFit prostate registration

Goal

We are developing registration module in Slicer version 4 (which is using ITKv4) for deformable registration of prostate MRI using BRAINSFit This Slicer4/ITKv4 registration tool should be functional, accurate and fast (i.e., comparable with the functionality we had in Slicer3/ITKv3, which has been evaluated previously [1]). We worked on implementing BRAINSFit registration in SimpleITK but due to several issues (see here) we are considering a C++ implementation that uses ITKv3 libraries and can be integrated in the latest Slicer versions.

Details on the registration approach

Registration is applied to align preprocedural and intraoperational MR T2 image volumes. We are using masks the prostate for both image data sets. Registration is done using MMI metric with rigid, affine and B-spline stages applied in sequence. In Slicer3/BRAINSFit we use gradient descent for rigid/affine, and LBFGS for B-spline.

Parameters we are using to call BRAINSFit in Slicer 3.6 and Slicer 4.4

Parameter	Slicer 3.6	Slicer 4.4
--fixedVolume	dir	dir
--movingVolume	dir	dir
--outputVolume	dir	dir
--bsplineTransform	dir	dir
--movingBinaryVolume	dir	dir
--fixedBinaryVolume	dir	dir
--useCenterOfROIAlign	True	not an option
--initializeTransformMode	not an option	useCenterOfROIAlign
--samplingPercentage	not an option	0.002
--useRigid	True	True
--useAffine	True	True
--useROIBSpline	True	True
--useScaleVersor3D	True	True
--useScaleSkewVersor3D	True	True
--useBSpline	not an option	True
--splineGridSize	3,3,3	3,3,3
--numberOfIterations	1500	1500
--maskProcessing	ROI	ROI
--outputVolumePixelType	float	float
--backgroundFillValue	0	0
--maskInferiorCutOffFromCenter	1000	1000
--interpolationMode	Linear	Linear
--minimumStepSize	0.005	not an option
--minimumStepLength	not an option	0.005
--translationScale	1000	1000
--reproportionScale	1	1
--skewScale	1	1
--numberOfHistogramBins	50	50
--numberOfMatchPoints	10	10
--numberOfSamples	100000	not an option
--fixedVolumeTimeIndex	0	0
--movingVolumeTimeIndex	0	0
--medianFilterSize	0,0,0	0,0,0
--ROIAutoDilateSize	0	0
--relaxationFactor	0.5	0.5
--maximumStepSize	0.2	not an option
--maximumStepLength	not an option	0.2
--failureExitCode	-1	-1
--debugNumberOfThreads	-1	not an option
--numberOfThreads	not an option	-1
--debugLevel	0	0
--costFunctionConvergenceFactor	1.00E+09	1.00E+09
--projectedGradientTolerance	1.00E-05	1.00E-05
--maxBSplineDisplacement	0	0
--maximumNumberOfEvaluations	not an option	900
--maximumNumberOfCorrections	not an option	25
--metricSamplingStrategy	not an option	Random
--costMetric	not an option	MMI
--removeIntensityOutliers	not an option	0
--ROIAutoClosingSize	not an option	9
--useExplicitPDFDerivativesMode	AUTO	not an option
--useCachingOfBSplineWeightsMode	ON	not an option

Results using the above parameters

We used this sample data with Reader 1 segmentations to visually evaluate differences in the registration results. All three results of BRAINSFit with ITKv3 in Slicer 3.6 look reasonable. Using BRAINSFit with ITKv4 in Slicer 4.4 creates a reasonable result for Case 7, but look strongly distorted towards the edges of the bounding box in Case 8 and 9.

Case 7

Reference intraop volume

Reg Result Slicer 3.6 with ITKv3

Reg Result Slicer 4.4 with ITKv4

Case 8

Reference intraop volume

Reg Result Slicer 3.6 with ITKv3

Reg Result Slicer 4.4 with ITKv4

Case 9

Reference intraop volume

Reg Result Slicer 3.6 with ITKv3

Reg Result Slicer 4.4 with ITKv4

Current status

SimpleITK registration code on github: here

Latest update: Mai 18, 2015

Things to do

• Set ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS to 1 as described here
• Write message to ITK mailing list, cc Nicholas Tustin: ask for regularization of the BSpline displacement field
• look at DICE comparison between Slicer3 and Slicer4 to quantify the problem
• displace the masks and run with Slicer3 and Slicer4 to quantify the impact of the mask in the registration result

Things that are working

• exhaustive search based initialization procedure implemented and produces good results (cross-correlation metric) in SimpleITK

Things that are implemented, but are not working

• SimpleITK code for rigid registration
  • the voxel sampling exception can be avoided by cropping the images sharply around the masks.

Fixed Volume as Reference

SimpleITK result after Initialization

SimpleITK result after rigid registration

BRAINSFit result after rigid without using masks and --useCenterOfGeometryAlign

• • The rigid registration result is too much rotated and translated. Also the number of iteration steps varies with every computation. Reason might be that seed is set randomized and not set as done in BRAINSFit here. There is no option at the SimpleITK::ImageRegistrationMethod to set a metric seed.

Things that are not implemented, but need to be implemented

• SimpleITK code for affine registration
  • setting affineOptimizer here
  • setting RegistrationMethod here
• SimpleITK code for BSpline registration (started here)

Unresolved issues out of our direct control

Standing SimpleITK issues

• no API for consistent initialization of the metric seed
  • The ITKv4 ImageRegistrationMethod random sampling utilizes a fixed seed (here)

Standing ITKv4 issues

• capability to sample voxels only within the mask (see topic under review on Gerrit: http://review.source.kitware.com/#/c/19684/)

References

[1] Fedorov A, Tuncali K, Fennessy FM, Tokuda J, Hata N, et al. (2012) Image registration for targeted MRI-guided transperineal prostate biopsy. J Magn Reson Imaging 36: 987–992. Available: http://dx.doi.org/10.1002/jmri.23688.