SpatialObject.IsInsideInWorldSpace(point) much faster now!

Niels_Dekker · February 2, 2024, 10:35am

For your information, we have just achieved a significant performance improvement!

With the latest ITK master branch revision (commit PERF: Make `m_ObjectToWorldTransformInverse->TransformPoint` non-virtual · InsightSoftwareConsortium/ITK@4ae04b3 · GitHub) a SpatialObject.IsInsideInWorldSpace(point) call will be much faster than before! I observed that it has become approximately twice as fast, compared to the last ITK release (v5.3.0), or even the latest release candidate (v5.4rc02)

The performance improvement was achieved by three pull requests that I did, which were reviewed by @Stephen_Aylward @dzenanz and @hjmjohnson, and merged in the last few days:

As SpatialObject.IsInsideInWorldSpace(point) is often called repeatedly over a large range of points (typically inside a for or a while loop), I expect a significant performance improvement for various filters, metric calculations, and registration algorithms. Especially when they involve image masks. Specifically, I would expect a performance improvement for:

SpatialObjectToImageStatisticsCalculator
DiffusionTensor3DReconstructionImageFilter
ImageMomentsCalculator
HistogramImageToImageMetric
ImageToImageMetric
KappaStatisticImageToImageMetric
MatchCardinalityImageToImageMetric
MattesMutualInformationImageToImageMetric
MeanReciprocalSquareDifferenceImageToImageMetric
MutualInformationImageToImageMetric
NormalizedCorrelationImageToImageMetric
JointHistogramMutualInformationImageToImageMetricv4
MattesMutualInformationImageToImageMetricv4
ImageRegistrationMethodv4

Please have a look and try it out! I’m interested to hear if you see any performance improvement in your use case!

Niels_Dekker · February 7, 2024, 5:22pm

For your information, both spatialObject.IsInsideInWorldSpace(point) and spatialObject.IsInsideInObjectSpace(point) have become even faster for masks (that is, when the spatial object is an itk::ImageMaskSpatialObject), now that the following pull request is merged:

github.com/InsightSoftwareConsortium/ITK

Let `ImageSpatialObject` update the image regions of its base class, let `ImageMaskSpatialObject` use these image regions

InsightSoftwareConsortium:master ← N-Dekker:ImageSpatialObject-UpdateImageRegions

opened 07:26PM - 06 Feb 24 UTC

N-Dekker

+82 -2

`SpatialObject` stores the image regions of the spatial object. It appears reaso…nable for its derived class `ImageSpatialObject` to keep them up-to-date, in accordance with its image. A significant performance improvement of `ImageMaskSpatialObject` is then achieved by using the buffered region and the requested region as cached in its base class, instead of calling the virtual functions `GetBufferedRegion()` and `GetRequestedRegion()` from its image. ---- It's interesting to see that the common base class `itk::SpatialObject` already has those data members to cache the image regions (`m_LargestPossibleRegion`, `m_RequestedRegion`, `m_BufferedRegion`), but it looks like they weren't used anywhere, as far as I can see: https://github.com/InsightSoftwareConsortium/ITK/blob/436423d798e63e72444dd4a9d8787ffd34c5eb60/Modules/Core/SpatialObjects/include/itkSpatialObject.h#L731-L733 For the record, those data members were added to `itk::SpatialObject` with commit 433987062ccaee80c9a7d2f6c1ff857fe96cf929 "ENH: Removed NDimensionalSpatialObject", Julien Jomier (@jjomier), 10 February 2003.

commit: PERF: ImageMaskSpatialObject use image regions from its base class · InsightSoftwareConsortium/ITK@671d758 · GitHub

For ImageMaskSpatialObject::IsInsideInObjectSpace, I observed a performance improvement of more than 8%, with this commit.

cookpa · February 9, 2024, 7:53pm

Thanks for letting us know - I updated ANTs. It does seem to run a bit faster, maybe 5%, with a brain mask and the Mattes metric.

Niels_Dekker · April 9, 2024, 8:27am

FYI, the SpatialObject.IsInsideInWorldSpace(point) performance improvements that I presented here are included with the latest ITK tag: v5.4rc03.