Insight Journal - N4ITK: Nick's N3 ITK Implementation For MRI Bias Field Correction

N4ITK: Nick's N3 ITK Implementation For MRI Bias Field Correction

Tustison N., Gee J.

PICSL, University of Pennsylvania

Please use this identifier to cite or link to this publication: http://hdl.handle.net/10380/3053

Published in The Insight Journal - 2009 January - June.

Submitted by Nick Tustison on 04-30-2009.

Several algorithms exist for correcting the nonuniform intensity in magnetic resonance images caused by field inhomogeneities. These algorithms constitute important preprocessing steps for subsequent image analysis tasks. One such algorithm, known as parametric bias field correction (PABIC), has already been implemented in ITK. Another popular algorithm is the nonuniform intensity normalization (N3) approach. A particularly salient advantage of this algorithm is that it does not require a prior tissue model for its application. In addition, the source code for N3 is publicly available at the McConnell Brain Imaging Centre (Montreal Neurological Institute, McGill University) which includes source code and the coordinating set of perl scripts. This submission describes an implementation of the N3 algorithm for the Insight Toolkit given as a single class, viz. itk::N3MRIBiasFieldCorrectionImageFilter. We tried to maintain minimal difference between the publicly available MNI N3 implementation and our ITK im- plementation. The only intentional variation is the substitution of an earlier contribution, i.e. the class itk::BSplineScatteredDataPointSetToImageFilter, for the originally proposed least-squares approach for B-spline fitting used to model the bias field. In addition, we include a more extensive modification to the original N3 algorithm found in the class itk::N4MRIBiasFieldCorrectionImageFilter. The latter algorithm employs a multi-resolution approach, similar to FFD image registration strategies, and has a slightly modified iterative update scheme.

Code

Automatic Testing Results by Insight-Journal Dashboard on Fri Jan 22 22:26:54 2010 for revision #9

expertise: 5 sensitivity: 5

This project passed all of its tests.
Click here for more details.

Go here to access the main testing dashboard.

Automatic Testing Results by Insight-Journal Dashboard on Tue Oct 27 14:44:12 2009 for revision #6

expertise: 5 sensitivity: 5

Automatic Testing Results by Insight-Journal Dashboard on Tue Oct 20 11:18:04 2009 for revision #5

expertise: 5 sensitivity: 5

Automatic Testing Results by Insight-Journal Dashboard on Fri Jul 17 07:23:06 2009 for revision #4

expertise: 5 sensitivity: 5

Reviews

Very nice contribution for open source community by Andriy Fedorov on 2009-06-17 16:05:40 for revision #3

expertise: 2 sensitivity: 5

Summary:

The authors present an implementation of a published method for MRI bias correction. The code follows ITK style, is accompanied by guidelines on parameter initialization and references to the literature. The authors supply the code with the data they used to test the code.

Hypothesis:

n/a

Evidence:

n/a

Open Science:

The code uses ITK, the authors do provide images, and sufficient details to replicate the results.

Reproducibility:

The code is compiled with review enabled. We were able to reproduce the results authors present, and successfully apply the implementation to our data.

Use of Open Source Software:

The code promotes best practices of open source and ITK programming.

Open source Contributions:

The code is usable, the authors provide flexibility to work with either 2d or 3d images, the code includes not only filters, but also the test driver program to simplify usage.

Code Quality :

The coding style seems to be very good, the code is readable and follows ITK design. I did not observe any features that would make this code non-portable, but personally have compiled it only on linux.

Quality of the data :

The authors provide reasonable testing data sets.

Interest:

CORRECTION: the accompanying document does not stress the meaning of the shrink factor. My original experiments we using shrink factor 1, which results in no resampling. With the shrink factor 4, the code takes under 1 minute for the same data. At the same time, I cannot distinguish visually the bias field from the one produced with shrink factor 1. I withdraw my complaint about high execution time.

Free comment :

The code is rather slow on 3d data: 256x256x124 image takes about 50 min to process. The code is not multi-threaded.

Mask generation can be challenging due to the nature of input data -- field inhomogeneities may not (and do not, in our case) allow to generate the mask with simple Otsu threshold.

The algorithm requires setting of a relatively large number of parameters, and due to significant execution time, parameter tuning is not easy. We were able to achieve visually pleasing results by increasing the m_BiasFieldFullWidthAtHalfMaximum parameter from its default value.

A wrapper module for 3D Slicer (http://www.slicer.org) has been implemented by Nicolas Rannou (Brigham and Women's Hospital), and can be downloaded from NAMIC SandBox here: http://svn.na-mic.org/NAMICSandBox/trunk/MRIBiasCorrectionFilter/. This module has preview mode, which allows to apply correction to a single slice, making parameter selection easier. The use of Slicer allows to use additional filters to construct and edit mask that can be used in bias correction.

Cannot Compile the program by Sylvain Bouix on 2009-05-21 15:29:36 for revision #1

expertise: 2 sensitivity: 5

Add a new review

Statistics


Global rating:
Review rating:	[review]
Code rating:
Views:	2766
Downloads:	1227

Send a message to the author

Information


Paper Id:	640
Categories:	Filtering, IO
Keywords:	MRI, bias field, inhomogeneity, ninjas,
Toolkit:	ITK
Revision:	9 (01-22-2010)
See revision:
Status:	Open for public review
View license Loading license...

Data


Full download:	.zip
Paper:	view, .pdf
Source code :	Download

Share

Associated Publications


Camino: Diffusion MRI reconstruction and processing
3D Segmentation in the Clinic: A Grand Challenge II: MS lesion segmentation