Abstract:
The aim of this study was to minimize the execution cost and maximize the noise-robustness of
automatic segmentation of brain Magnetic Resonance Images (MRI). Partitioning brain tissues in
MRI is considerably tough as they present low contrast and intensity inhomogeneity. The case is
further aggravated due to inherent noise from imaging environment. Conventionally, Fuzzy CMeans (FCM) clustering is employed for tissue classification; however, this technique is costly
and sensitive to noise. Moreover, FCM can best detect spherical clusters only due to which this
technique becomes less accurate for brain images. Therefore, Gustafson-Kessel (G-K) clustering
technique is employed which has the capability to detect clusters of dissimilar shapes and sizes.
Nonetheless, G-K is also expensive and sensitive to noise. This technique has yet not achieved a
satisfactory accuracy level either, mainly because of high sensitivity to initialization. Hence, a
few modifications are presented in an effort to produce an efficient and robust G-K algorithm for
tissue segmentation. First, a histogram-based method is proposed for G-K initialization. Second,
cluster volume estimates are inputted to the algorithm for utilizing its ability of detecting variable
sized clusters. Third, intensity information is fused with textural information in G-K objective
function for better classification. The algorithm is used to segment test MRI of brain into three
classes: Gray Matter (GM), White Matter (WM) and Cerebrospinal Fluid (CSF). Experimental
results show that the proposed algorithm performs better than conventional FCM and GK in
terms of both execution cost and robustness.
