Kernelized Fuzzy c-means Method in Fast Segmentation of Demyelination Plaques in Multiple Sclerosis

Abstract

Fuzzy c-means method (FCM) is a popular tool for a fuzzy data processing. In the current study, a FCM-based method of fuzzy clustering in a kernel space has been implemented. First, a "kernel trick" is applied to the fuzzy c-means algorithm. Then, the new method is employed for a fast automated segmentation of demyelination plaques in Multiple Sclerosis (MS). The clusters in a Gaussian kernel space are analysed in the histogram context and used during the initial classification of the brain tissue. Received classification masks are then used to detect the region of interest, eliminate false positives and label MS lesions.

Review of automatic segmentation methods of multiple sclerosis white matter lesions on conventional magnetic resonance imaging

“…Some authors extended the previous methods for healthy brains to use several sequences at the same time and added an extra class to include the WML in their classification (Forbes et al, 2010;Kawa and Pietka, 2007;Kikinis et al, 1999;Shahar and Greenspan, 2004;Shiee et al, 2009). …”

“…A hierarchical segmentation method was proposed whereby the fuzzy C-means is used twice in PDw images: first, to detect the lesion+CSF cluster with respect to the GM and WM and, second, to differentiate lesions from CSF (Boudraa et al, 2000). Kawa and Pietka (Kawa and Pietka, 2007) proposed to include spatial information in the clustering using the kernel fuzzy C-means combined with fuzzy connectedness theory (Udupa et al, 1997). Shiee et al (Shiee et al, 2009) modified the fuzzy C-means algorithm to include anatomical information using both anatomical and topological atlases.…”

Kernelized Fuzzy C-Means Method and Gaussian Mixture Model in Unsupervised Cascade Clustering

“…The kernelised clustering methods constitute the development of conventional FCM clustering algorithm, discussed in detail in [7,8,9].…”

“…The algorithm described in [7] uses a kernel-induced distance metric to replace the square one (||·||) applied in FCM. In this modification, the objective function is given as [7] …”

“…The improvement of clusters shape in the feature space can be implemented using KFCM clustering. Two different approaches are used [7,12,13,8]: (1) the cluster centers (prototypes) are calculated in the original feature space and the clustering process is transformed to the kernel space, (2) the whole clustering process is transformed to the multidimensional space -described as clustering in the kernel space with implicit prototypes [7].…”