Bias correction of intensity inhomogeneous images hybridized with superpixel segmentation

Abstract

Image intensity bias correction is still an open problem. Fuzzy c-means based bias correction methods are popular. However, they are time consuming and the exact clustering number has to be known in advance. To address the above problems, we propose a bias field correction embedded FCM (BCEFCM) method which is hybridized with superpixel clustering. Furthermore, the proposed method is extended to address similar problems existing in colourful images and multiple modality images. Moreover, a histogram based strategy is proposed to estimate the clustering number in advance, which is generally greater than the exact clustering number. The proposed methods have been extensively tested on representative images from BrainWeb and the MICCAI SATA challenge. Results demonstrate advantages of the proposed methods in no need to know the exact clustering number in advance and producing better bias correction and image segmentation results than representative methods in terms of coefficient of variations (CV), coefficient of joint variation (CJV), root mean square error (RMSE), and structural similarity index (SSI).