Noise and cluster size insensitive robust weighted fuzzy clustering for medical image segmentation

Fuzzy c-means clustering (FCM) is a popularly-known method used for image segmentation. To handle sensitivity due to noise, many variants are proposed in the literature. However, these methods do not perform well for data having non-spherical distribution and are also sensitive to unbalanced cluster size. Many of these methods suffer from noise sensitivity. To overcome these limitations, in this paper, we propose the FkPCM_S_UB method that utilizes a weighted sum of FCM and FkPC to segment images having non-spherical or a mixture of flat-shaped and spherical-shaped data. It includes both local spatial and gray-level information to control the effect of noise and a fuzzy exponential entropy term to tackle the problem of unbalanced cluster size. The effectiveness of the proposed FkPCM_S_UB method is evaluated on four medical image datasets using the average segmentation accuracy and dice score as performance measures. The proposed FkPCM_S_UB method outperforms all eighteen existing methods for ASA and DS for all four medical image datasets. Friedman statistical test is also utilized to establish the statistically significant better performance of the proposed FkPCM_S_UB method compared to the eighteen existing methods used for medical image segmentation.