Robust brain magnetic resonance image segmentation using modified rough-fuzzy C-means with spatial constraints

Abstract

The structure of human brain is complicated, usually overlapping, uncertain, vague, and indiscernible in nature. Moreover, brain magnetic resonance images (MRI) often suffer from outliers, noise and other artifacts. To deal those issues in this article, a novel robust clustering algorithm rough-fuzzy C-means with spatial constraints (RFCMSC) for brain MRI segmentation is proposed. The judicious amalgamation of fuzzy set and rough set theory in the clustering can better handle the inherent vagueness, uncertainties, overlapping, and indiscernibility present in brain MRI, whereas the concept of spatial constraints (in the form of contextual information) allows the pixel labeling to be influenced by the immediate neighboring pixels to handle the noise and other artifacts. The proposed method is simulated with a variety of benchmark brain MRI datasets as well as with synthetic images with added noise. The algorithm is tested using overall accuracy, precision, recall, macro F1, micro F1 and Kappa. The superiority and robustness of the algorithm is justified from the experimental results in comparison to other counterpart clustering based segmentation methods on both benchmark brain MRI and synthetic images with and without noise. Paired t-test confirms the statistical significance of the results in favor the proposed method compared to other algorithms.