Topic: 219 Processing and Quantification: Imaging
D. García-Lorenzo1, S. Prima1, L. Collins2, D. L. Arnold3, C. Barillot1, S. P. Morrissey4;
1Unit/Project VisAGeS U746, INSERM/INRIA, IRISA, UMR CNRS 6074, University of Rennes I, Rennes, France, 2McConnell Brain Imaging Centre,, Montreal Neurological Institute and Hospital,, Montreal,, QC, Canada, 3McConnell Brain Imaging Centre,, Montreal Neurological Institute and Hospital,, WB 321, Montreal,, QC, Canada, 4Department of Neurology, University hospital of Rennes, Unit/Project VisAGeS U746, INSERM/INRIA, IRISA, UMR CNRS 6074, University of Rennes I, Rennes, France.
Presentation Number: 712
Purpose/Introduction:Automatic segmentation of MS white matter lesions (WML) and normal appearing brain tissue (NABT) in MRI is a challenging task for image processing. Most automatic segmentation algorithms can be roughly divided into two groups: Global and local methods. The challenge of the latter methods is to combine these local regions to build a global and meaningful segmentation: Recently MeS, an unsupervised and non-parametric gradient density estimation algorithm, was introduced to segment brain MRI of healthy volunteers [1, 2]. The first method uses an atlas to label the regions given by the MeS in three classes (white matter (WM), grey matter (GM) and cerebrospinal fluid (CSF)). The second method uses an EM algorithm [3] over the image intensities to estimate a 3-class Finite Gaussian Mixture Model, and then assigns each MeS region to the class with highest probability. Here, we combine MeS with a modified (m)EM from [4] to improve MS WML segmentation (Fig.1).
Subjects and Methods:Synthetic images: T1-w, T2-w and PD-w of the simulated MS brain from Brainweb [5], with 3% noise (n) and 0%, 20% and 40% inhomogeneity (rf). Validation: We compared the ground truth with the automatic segmentation (Dice Similarity Coefficient (DSC) [6] is used for MS WML). MS patients (n=7) were acquired on a 1.5T (Philips): 3-mm axial slice thickness T1-w, T2-w and PD-w. Images were denoised, corrected for intensity inhomogeneity, normalized in the stereotaxic space and skull-stripped. WML were manually segmented by an expert and validated using DSC.
Results:Synthetic images (Table 1) show that MeS improves WML segmentation for all levels of inhomogeneity and EM algorithm outperforms mEM for all levels of inhomogeneity.
MS patients’ MRIs: EM algorithm fails to estimate the NABT model in all patients (algorithms B and D), which shows the limitation of the phantom studies. Algorithm A slightly improves the results of algorithm B and shows less variance in its results (Table 1). Regarding robustness of algorithm A Mean Shift regions are estimated with local information, and therefore their classification is facilitated even if the model is less well estimated.
Discussion/Conclusion:A new algorithm for MS WML and NABT segmentation is presented combining global and local information. Merging global and local information improves segmentation results instead of using global information. We also show that a robust estimation of parameters with mEM is crucial to correctly segment real images with this method [2].
| BW n3rf O% (BW: results of Brainweb synthetic images) | BW n3rf20% | BW n3rf40% | Average DSC of MS MRI | ||
| A (MeS combined with mEM) | 0.87 | 0.85 | 0.63 | 0.55+/- 0.05 | |
| B (MeS combined with classical EM, similar to Mayer et al. [2]) | 0.87 | 0.84 | 0.79 | ---- | |
| C (instead of classifying the MeS regions, each voxel is classified independently with the mEM algorithm, similar to Ait-Ali et al. [4]) | 0.72 | 0.77 | 0.41 | 0.52+/- 0.07 | |
| D (as algorithm C but using the classical EM algorithm instead of the mEM) | 0.79 | 0.80 | 0.78 | ---- |
| 1 | Jimenez-Alaniz, J. et al.: IEEE TMI 25(1) (Jan. 2006) 74-83 |
| 2 | Mayer, A., Greenspan, H.: ISBI’06 319-322 |
| 3 | Dempster, A.P. et al.: Journal of the Royal Statistical Society 39(1) (1977) 1-38 |
| 4 | Ait-Ali, L.S. et al.: MICCAI’05 8(Pt 1) 409-416 |
| 5 | Collins, D. et al.: IEEE TMI 17(3) (Jun 1998) 463-468 |
| 6 | Zijdenbos, A. et al.: IEEE TMI 13(4) (Dec 1994) 716-724 |
Figure 1: Workflow of Robust Expectation Maximization with Mean Shift
