A fast kernel-based clustering algorithm with application in MRI image segmentation

In kernel-based algorithms, Mercer kernel techniques have been used for improving the separability of input patterns. Although designed to tackle the problem of curse of dimensionality, non-accelerated kernel-based clustering algorithms fail to provide enough time efficiency for practical applications, such as medical image segmentation. For improving the time efficiency of kernel-based clustering, different speed-up schemes can be adopted. Different from the approximate pre-image technique used in KFCM-II, this work proposes a novel approach based on the technique of reduced set for speeding up kernel clustering process. This algorithm, called KFCM-III, balancing the power of KFCM-I and the efficiency of KFCM-II, has potentials to outperform its rivals and has been applied in segmenting MRI images. The experiments on synthetic images and MRI brain phantom have shown the effectiveness of the proposed algorithm, which can not only reduce computational complexity of kernel clustering but also retain better segmentation results when using well-tuned parameters of reduced set.