The author describe a new filter to perform the convolution of an image by another.

The convolution is well supported in ITK, but could be easier to setup.

I've been able to download, build and validate the result with the test suite. It has required a few changes though: I replace the end of the CMakeLists.txt by

ADD_TEST( ConvolutionTest ${CurrentExe} ${CMAKE_SOURCE_DIR}/Images/BrainT1SliceBorder20.nii ${CMAKE_BINARY_DIR}/ )
ADD_TEST( AverageImageCompare ${IMAGE_COMPARE} ${CMAKE_SOURCE_DIR}/Images/output0.nii ${CMAKE_BINARY_DIR}/output0.nii )
ADD_TEST( Sobel1ImageCompare ${IMAGE_COMPARE} ${CMAKE_SOURCE_DIR}/Images/output1.nii ${CMAKE_BINARY_DIR}/output1.nii )
ADD_TEST( Sobel2ImageCompare ${IMAGE_COMPARE} ${CMAKE_SOURCE_DIR}/Images/output2.nii ${CMAKE_BINARY_DIR}/output2.nii )

Convolution is one of the basis transform of image analysis. It is useful almost everywhere.

The filter makes the convolution a lot easier in ITK, and makes it possible to integrate it in the pipeline architecture.

A few suggestions:

• You should show an example of convolution without your filter in the article, and highlight the differences.


• You should say in the article that the pipeline integration is way better with your contribution.


• To make the experimentation easier, you should provide a test program which accept the kernel image as input.


• To enhance the pipeline architecture, I would suggest changing the way how the kernel image is managed: I think it should be managed as a second input. That way, the test program of the previous point would require a single Update() call.


• Finally, to make the life of the user easier, you may want to add an option in the filter to normalize the values of the kernel. The filter would change the values of the pixel to make the sum of the pixels be 1.0. That way, the user would be able to pass integer images as kernel, images which are easier to edit by hand.