An improved soft subspace clustering for magnetic resonance images segmentation

Magnetic resonance (MR) images with a large amount of information, intensity inhomogeneities and weak boundary may cause serious misclassification in image segmentation when the existing soft subspace clustering (SSC) algorithms are used. As for this problem, this paper presents an improved SSC algorithm based on brain storm optimization (BSO) algorithm, called BSOSSC algorithm to segment MR images. In BSOSSC, we introduce an objective function in combination of relaxation boundary constraint and generalized noise clustering (GNC) to remedy the defect that the existing algorithms are sensitive to noisy data firstly. Then, we apply a new calculation method of membership degree to seek the subspace where the cluster class is located. Finally, we use the BSO algorithm to equipoise the local search and the global search and prevent the SSC algorithm from immersing in local optimum in the clustering process. The experimental results of BSOSSC are compared with those of algorithms such as fuzzy C-means (FCM), entropy weighting K-means clustering (EWKM), Minkowski metric weighted K-means clustering (MWK-means), fireworks algorithm-based soft subspace clustering (FWASSC) and subspace clustering via graph filtering based thresholding ridge regression (FTRR) on the UCI datasets, Berkeley Computer Vision images and Simulated Brain Database (SBD) show the superiority of BSOSSC in terms of anti-noise, anti-intensity inhomogeneities and precision. Furthermore, the segmentation results of clinical breast MR images have validated the robustness of the BSOSSC algorithm.