Summary:
The authors provide a time and memory efficient implementation in ITK framework for computing Voronoi diagrams of a set of objects, as well as a k Nearest Neighbor transform that can be used for MRI segmentations. The kNN implementation is based on the output of the provided voronoiFilter. The implementation is based on an algorithm presented in [3], but improves its computational efficiency.

Evidence:
This paper is presenting a faster implementation for the problems it addresses, and convincing computation time plots are provided. A brief mathematical proof of better computational complexity could have been useful, too.

Open Science:
Source code is provided, and it seems these filters can be integrated to the ITK.

Reproducibility:
The authors provide their implementations. I have not run the code.

Code Quality:
Some comments in the source code is in Spanish (I think), which is confusing.

Applicability to other problems:
The Voronoi diagram computation and kNN segmentation methods are useful for many applications beside medical imaging.

Requests for additional information from authors:
It is not explained why the current implementation is limited to only 2 channels for data. How difficult would it be for the user to extend this code to an arbitrary number of channels?

Additional Comments:
The figure 4 is missing axis labels, and it took me a while to understand what was being shown. Also, one of the strongest arguments of the paper is that it provides a faster solution than existing algoritms; therefore I think the time efficiency comparison with the Danielsson DT implementation should be communicated more clearly and convincingly.

Summary:
This paper describes methods and implementations for non-parametric classification based on k nearest neighbors.

Evidence:
Timing plots comparing the execution time of this kNN implementation against a stock implementation.

Open Science:
Source code is provided.

Reproducibility:
I downloaded the code and it built on Visual Studio .NET 2003 without issue.

Use of Open Source Software:
This is based on ITK.

Open Source Contributions:
Source code is provided.

Code Quality:
Code does not adhere to the ITK coding conventions. Will need to be cleaned up.

It is unclear how to use the filter. The paper would benefit from an illustrative example of how to tie together the filters to perform a classification. Perhaps the architecture needs some work in order to tie into the ITK pipeline better. In particular, the specification of the prototypes could fit into ITK better.

Applicability to other problems:
kNN classification is an important non-parametric technique in medical and biological image analysis (as well as other domains).

Suggestions for future work:
[Suggest to authors future directions for improving their methods, or other domains from which they could learn technique that could help them advance in their research.]

Requests for additional information from authors:
The figures could use more description. Particularly Figure 4 which shows multiple timing responses that are not labeled.

The authors indicate the methods are N-dimensional. But the authors also state that the they provide 1 and 2 channel methods. Does this mean the implementation is not N-dimensional?

Additional Comments:
Minor language issues in the text.

Summary:
This paper presents a nice implementation of KNN classification within the ITK framework.

Hypothesis:
- Accurate Distance Transform that is more efficient than Danielson DT
- Voronoi based segmentation filter for 2D and 3D

Evidence:

- The results show that the filter classifies images, but not validation is preformed
- The plots do not prove, but suggest that the classification filter is indeed efficient with O(N*k) computation time
- Image 7 is from the Brainweb database, which is a simulated healthy control data. It is thus unclear, what the dark blue label represents computed classification. They look like segmented lesions, but since this is a normal, I can only assume that this a clear misclassification.
- No comparison to other methods using the Brainweb database or the IBSR database

- There is no evalutation/comparison of the proposed DT to Danielson or other DT\'s in regard to accuracy

Open Science:
The source code is provided with examples. The authors follow the standards of open source. Currently no KNN filter in ITK

Reproducibility:
Building on MacOSX was not problem of both source code parts, as well as reproducing the tests provided with the source code.

Use of Open Source Software:
Builds on ITK.

Code Quality:
Code seems to follow general ITK standards, though I did not look at the complete code. I was not able to understand the spanish comments in the code.

Applicability to other problems:
It is designed for segmentation, but the enhanced DT is applicable to a wide set of problems.

Minor Comments:
- Page 3, 1st paragraph: $k^th$ level should be $k$-th level or rathe $k^{th}$?
- training samples are called prototype patterns

Summary:
This paper describes the implementation of a new segmentation technique into ITK and its associated algorithms. Specifically,
distance transform (DT), Voroni diagram, and k Nearest Neighbor (kNN) transform. The results of these transforms are used to efficiently perform kNN
classification using the KNN rule.

Hypothesis:
Though not the main points of the paper, the authors make a few claims:

1) Their segmentation method is fast and accurate for 2D and 3D medical data sets
2) Their propagation based DT is faster then the Daneilsson DT implemented in ITK
3) Their "algorithms can be used for many other applications..."
4) The kNN classifier has O(kN) complexity

Evidence:
1) Fast: Table 1 provides running times showing quite reasonable results for a segmentation of a 3D volume (~6 seconds), but comparison to running times of similar algorithms is not provided. Accurate: Only qualitative validation is provided. No comparison (qualitative or quantitative) to the accuracy of other methods is provided.

2) Figure 4 shows a comparison between the Daneilsson DT and the proposed method. However, a key describing which plot is the Daneilsson and which is the proposed method is not provided. The accuracy of the DT is not mentioned, though I did not read reference [3].

3) The authors do not demonstrate their algorithm on other applications.

4) I did not see any support for this claim.

Open Science:
This paper adheres to the principles of Open Science. Source code and data for the programs used in the experiments is provided. Furthermore, the output images are
also provided.

Reproducibility:
I was unable to build the project using WindowsXP with MS Visual Studio 2005, therefore I could not test the algorithms. Specifically,
I encountered errors with "error C2027: use of undefined type 'timeval'" and "gettimeofday". Including "winsock2.h" solved the timeval problem.

Use of Open Source Software:
The authors use ITK. However, as noted I was unable to build the project.

Open Source Contributions:
The source code is provided, as are sample inputs. Though the code does contain comments, I believe doxygen complaint documentation in the header files
would be usefull.

Code Quality:
The code is fairly straightforward, though I only took a passing glance at it. Again, doxygen compliant documentation would be helpful.

I was unable to build the code on my windows machine due to what appears to be a system dependent function call: "gettimeofday"

Applicability to other problems:
None come to mind immediately.

Suggestions for future work:
See additional comments section.

Requests for additional information from authors:
A key is required for Figure 4 describing which plots correspond to which DT method.

Additional Comments:

I believe the new faster DT in itself can be a valuable addition to ITK, as many classes make use of a DT. However, it was not immediately
clear to me how or if I can use their classes to compute only a DT. Doxygen compliant documentation would be useful here.

Along those lines, I would have liked to see each geometric algorithm implemented as a separate ITK class which one could then string
together to create the final kNN segmentation filter. Though this would obviously induce some overhead, it would allow further development
of the method in a plug and play fashion. For example, one could then replace the new DT with a newer not yet existing but faster method.