An unsupervised orthogonal rotation invariant moment based fuzzy C-means approach for the segmentation of brain magnetic resonance images

Brain magnetic resonance images (MRI) suffer from many artifacts such as noise and intensity inhomogeneity. Moreover, they contain an abundant amount of fine image structures, edges, and corners in various areas of the image. These anomalies and structural complexities affect the segmentation process of the brain MRI which is required by physicians for the diagnosis purpose. Recently, we have proposed a local Zernike moment (LZM)-based unbiased nonlocal means fuzzy C-means (LZM-UNLM-FCM) approach that has dealt with the noise artifact in the moment domain using the LZM approach. The method provides high segmentation results for the MR images corrupted with Rician noise. However, the method does not deal with the intensity inhomogeneity artifact effectively. Moreover, the method uses a regularization parameter that needs to be adjusted to obtain effective segmentation results. This paper presents an unsupervised local Zernike moment and unbiased nonlocal means-based bias corrected fuzzy C-means (LZM-UNLM-BCFCM) approach that deals with both noise and intensity inhomogeneity artifacts. The main concept behind the proposed method is to use the attractive properties of the LZMs to effectively filter the image by determining a large number of similar regions in an MR image which is mostly corrupted by Rician noise and intensity inhomogeneity. The ability of the LZMs to determine such regions in MR images consisting of fine tissue structures in any orientation is well utilized for dealing with the high levels of noise. The intensity inhomogeneity is removed by estimating the bias field pixel-by-pixel during the segmentation process using the filtered image without the use of regularization parameter. The bias field is estimated as a linear combination of the orthogonal polynomials in which the weights are obtained by minimizing the fuzzy objective function. Experimental results on both simulated and real MR images show the superiority of the proposed method as compared to other unsupervised state-of-the-art approaches. © 2020 Elsevier Ltd