Reviews

James Miller

Thursday 31 August 2006

Summary: This paper describes an contribution to ITK to generate well composed images. In well composed images, all foreground and background pixels are 4-connected. This simplifies post-processing algorithms.

Hypothesis: Non applicable

Evidence: Example images of the results are provided. The motivating images in Figure 1 are difficult discern the benefit of well-composed. Perhaps this image can be annotated for the reader.

Open Science: The paper uses ITK and extends itkInplaceImageFilter

Reproducibility: Code built without issue

Use of Open Source Software: Code uses ITK.

Open Source Contributions: Source code provided

Code Quality: Code provides separate implementations for 2D and 3D processing. Can these two implementations be combined into one filter? The use of the InPlaceFilter is nice.

Applicability to other problems:

Suggestions for future work:

Requests for additional information from authors: Details on the specific algorithms should be provided in the paper. The paper describes the \\\“what\\\” but not the \\\“how\\\”. The reader is directed to the references for the \\\“how\\\'s\\\”.

Additional Comments: [This is a free-form field]

Torsten Rohlfing

Friday 22 September 2006

Summary: The paper describes the implementation of filters for repairing morphological defects in 2D and 3D binary images.

Hypothesis: Not applicable.

Evidence: 2D and 3D test data and results are provided.

Open Science: The algorithm is implemented in ITK; source code is included. Two test images (one 2d, one 3d) are also included.

Reproducibility: Code compiles and runs, but the name of one example file (brain3d.mha) does not match the name coded into the test program (brain3D.mha with capital \“D\”).

Use of Open Source Software: The algorithm is implemented using ITK and fits within the ITK design.

Open Source Contributions: Source code is provided. The paper does an excellent job of explaining how to use the implemented filter.

Code Quality: The code seems quite well written and the style conforms with ITKs coding style. Documentation is lacking (this is why I am rating 4 out of 5) - many fields in the filter class are undocumented. Critical sections of the actual implementation should also be documented in the txx file.

Applicability to other problems: See Suggestions for future work

Suggestions for future work: I am wondering - can these algorithms be extended to accomodate multi-label images? One could of course treat each label separately and then merge the resulting corrected binary images back into a corrected multi-label image. Is there anything in the algorithm that prevents ambiguities in such a merging step (it is not a shortcoming of the algorithm if there isn\'t such a property, but it would be immensely useful if it did).

Alternatively - is it conceivable to truly extend the method to multi-label situations and fixing all labels simultaneously?

I am asking because bad configurations like the ones described here often result when I combine multiple segmentations of a single image using label voting. We have recently proposed \“shape-based averaging\” based on the Euclidean distance transform (Rohlfing & Maurer, MICCAI 2005; IEEE-TIP paper in press December 2006). But that algorithm is extremely computationally expensive (one signed EDT per label and input image), so a morphological algorithm like the one proposed here might be a good alternative.

Requests for additional information from authors: One of the beautiful aspects of the proposed algorithms is that it seems to be completely parameter-free. So the description is naturally quite complete.

Additional Comments: This is an extremely well-written paper. Complete yet compact, to the point, and very carefully illustrated. Excellent work, and a pleasure to read!