Overview

Accurate and robust brain tissue segmentation from MR images is a key issue in many applications of medical image analysis and, particularly, in the study of many brain disorders. Manual tracing of the three brain tissue types, white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF), in MR images by an expert is far too time consuming as the data involved in most studies is large. On the other hand, automated and reliable tissue classification is a demanding task as the intensity representation of the data normally does not allow a clear delimitation of the different tissue types present in a natural MRI. This is due to the partial volume (PV) effect (presence of more than one brain tissue type in a voxel), image noise and intensity non-uniformities caused by the in-homogeneities in the magnetic field of the MR scanner.

Validation of brain tissue classification is a complex issue in medical image processing. Visual inspection and comparison with manual segmentation are labor intensive and almost not reliable since the amount of data to deal with is usually large. Tissue classification methods can also be assessed by using synthetic data even if these kinds of images can hardly capture the complexity and the artifacts present in a MRI. There is however the possibility to validate brain tissue segmentation methods on a brain phantom. This phantom is very well-suited for this purpose since a ground-truth classification is known while different types of T1w MR modalities and image artifacts can be reproduced.

Figure 1. Digital brain phantom T1-weighted.

Our research

The goal of this project is to assess the robustness and accuracy of some of the most used unsupervised classification methods. In this comparative analysis and validation only T1-weighted MR brain image are considered. No enhancement of the image quality is done either before or during the classification process. This way robustness and accuracy of the methods is tested in front of the image artifacts.

First, we consider the finite Gaussian mixture model, noted by A:FGMM, with a Bayesian classification. The second method, B-HMRF, is closely related to the first one but it also assumes a hidden Markov random field (HMRF) model to account for spatial prior information as in [1]. For this model, the classification is done by the maximum a posteriori (MAP) criterion. Third, method C-GPV represents the image model proposed by Santago et al. (see [2]) where pure tissues are still modelled by a Gaussian distribution and mixture voxels are composed by two pure tissues. Consequently, a partial volume distribution for mixture tissues different from Gaussian is derived. Here again, Bayesian classification is used for the final classification. The fourth method, D-GPV-HMRF, uses the same image model as in method C-GPV, but it also encodes spatial information by a HMRF as in [3,4]. The fifth algorithm does not model the tissue classes by parametric probability densities, but rather by non-parametric models [5]. In this method, the probabilistic tissue model and the classification criterion can not be distinguished anymore, but are directly interdependent. The resulting algorithm minimizes an information theoretic quantity, called the error probability (E-EP). The final method is also non-parametric, but again adds to E-EP a HMRF to model spatial prior information (F-NPHMRF).

Figure 2. Validated methods.

Results

The best classification corresponds to the highest percentage of correct classified voxels. For low levels of noise (N={0,1,3}%), it is not obvious to determine a method that classifies better than others. These levels of noise are though not realistic in simulating the noise present in a MR image. For higher levels of noise (N={5,7,9}%), the method D:GPV-HMRF has almost always performed the best classification closely followed by the method B:GHMRF (their performance differs from less than 2%). It can be seen in Figure 3 that the classification based only on intensity information (methods A:FGMM, C:GPV, and E:EP) is much more noisy than classification that also encodes spatial information. Errors are due to the overlap between tissue distributions, and this overlap is larger for higher values of noise and in-homogeneities. On the contrary, when spatial information is also used in the classification process results are much less noisy: methods B:GHMRF, D:GPV-HMRF and F:NP-HMRF improve the percentage of voxels classified correctly, with respect to methods A-FGMM, C:GPV and E:EP, by a 7% in average. However, they still make some errors mostly in the mixtures classification because the partial volume distribution model is probably not well-suited but also because of the MRF. In fact, results show that MRF considerably increases the classification quality and that makes the algorithms more robust than the intensity-based approaches in presence of noise.

Figure 3. Results of the classification methods for the brain phantom with 7N and 20RF. First row, methods using intensity information only. Second row, the methods that add to the intensity the spatial prior.

Figure 4. Percentage of voxels correctly classified for the brain phantom with 20RF and all levels of noise.

Related publications

• M. Bach Cuadra, L. Cammoun, T. Butz, O. Cuisenaire and J. Thiran, "Comparison and Validation of Tissue Modelization and Statistical Classification Methods in T1-weighted MR Brain Images", IEEE Transactions on Medical Imaging, Vol. 24, No 12, pp. 1548- 1565, December 2005. [PDF].
• M. Bach Cuadra, J. Gomez, P. Hagmann, C. Pollo, J. Villemure, B. Dawant and J. Thiran, "Validation of Tissue Modelization and Classification Techniques in T1-Weighted MR Brain Images", MICCAI 2002; 290-297; Tokyo; Japan; 2002. [PDF].
• Meritxell Bach Cuadra, "Atlas-based Segmentation and Classification of Magnetic Resonance Brain Images", Ph. D. Thesis Number 2875, ITS-EPFL, November 2003. [PDF].
• Torsten Butz, "From error probability to information theoretic signal and image processing", Ph. D. Thesis Number 2798, ITS-EPFL, June 2003. [PDF].
• T. Butz, P. Hagmann, E. Tardif, R. Meuli and J. Thiran, A new brain segmentation framework, Lecture Notes in Computer Science, 6th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2003, Montreal, Canada, Vol. 2879, pp. 586-593, November 2003. [PDF].

Current research

Emerging classification methods add atlas information to the intensity and local spatial priors [3,7,8,9]. One main line of our current research is to quantify the importance of this kind of information. Some preliminary results [6] have shown that the performance of such methods is very sensitive to registration errors and to the precision of the atlas prior. Actually, mixture tissues are particularly affected by prior class template errors while pure tissue classification has been almost always improved.
Our current research aims also to quantify the sensitivity to the algorithm parameters like in HMRF methods as well the the effect of pre-processing the images (by an anisotropic filter or a bias corrector) or adding a bias field estimation model. We expect both the pre-processing and bias model (as in [4,10,11]) to make the classification more robust faced with noise and inhomogeneities.

Source Code

All parametric methods are implemented in Matlab (we are implementing some routines in C++). Currently, only the methods A-FGMM and B-GHMRF are public available. If you are interested just contact me by email. The code is not extremely well documented but it sticks quite close to the algorithm described in the paper [1]. Note that we will not provide any kind of technical support and that you can use/modify the programs for any use you wish, provided you cite [6] in any publication about it.

Some important references

• [1] Y. Zhang et al., "Segmentation of Brain MR Images Through a Hidden Markov Random Field Model and the Expectation-Maximization Algorithm", IEEE Trans. Medical Imaging, vol. 20, 45-57, 2001.
• [2] P. Santago and H.D. Gage, "Quantification of MR Brain Images by Mixture Density and Partial Volume Modeling", IEEE Trans. Medical Imaging, vol. 12, 566-574, 1993.
• [3] A. Noe, S. Kovacic and J.C. Gee, "Segmentation of Cerebral MRI Scans Using a Partial Volume Model, Shading Correction, and an Anatomical Prior", SPIE Medical Image Processing 2001.
• [4] K. Van Leemput, F. Maes, D. Vandermeulen and P. Suetens, "Automated Model-Based Bias Field Correction of MR Images of the Brain", IEEE Transactions on Medical Imaging, vol. 18(10), 885-896, 1999.
• [5] T. Butz, "From error probability to information theoretic signal and image processing", June 2003, Signal Processing Institut, Swiss Federal Institut of Technology, Switzerland.
• [6] M. Bach Cuadra, L. Cammoun, T. Butz, O. Cuisenaire and J. Thiran, "Comparison and Validation of Tissue Modelization and Statistical Classification Methods in T1-weighted MR Brain Images", IEEE Transactions on Medical Imaging, Vol. 24(12), 1548-1565, 2005.
• [7] K. Van Leemput, F. Maes, D. Vandermeulen and P. Suetens, "Automated Model-Based Tissue Classification of MR Images of the Brain", IEEE Transactions on Medical Imaging, vol. 18(10), 897-908, 1999.
• [8] K.M. Pohl, W.M. Wells, A. Guimond, K. Kasai, M.E. Shenton, R. Kikinis, W. Eric, L. Grimson and S. K. Warfield, "Incorporating Non-rigid Registration into Expectation Maximization Algorithm to Segment MR Images", Medical Image Computing and Computer-Assisted Intervention, MICCAI, 2002.
• [9] J.L. Marroquin, B. C. Vemuri, S. Botello, F. Calderon and A. Fernandez-Bouzas, "An Accurate and Efficient Bayesian Method for Automatic Segmentation of Brain MRI", IEEE Transactions on Medical Imaging, vol.21(8), 934-945, 2002.
• [10] W.M. Wells, R. Kikinis, W.E.L. Grimson and F. Jolesz, "Adaptive Segmentation of MRI data", IEEE Transactions on Medical Imaging, vol. 15, 429-442, 1996.
• [11] D.W. Shattuck, S.R. Sandor-Leahy, K.A. Schaper, D.A. Rottenberg and R.M. Leahy, "Magnetic resonance image tissue classification using a partial volume model", Neuroimage, vol. 13, 856-876, 2001.

Involved people

• Prof. Jean-Philippe Thiran
• Dr Olivier Cuiseanire
• Dr. Meritxell Bach Cuadra
• Leila Ben Hamida Cammoun
• Dr Patric Hagmann