Selective segmentation of inhomogeneous images based on local clustering and global smoothness

Abstract

Selective segmentation of inhomogeneous images is a challenging task in computer vision. Most existing distance-based selective segmentation models usually study homogeneous images and seldomly consider intensity inhomogeneity. On the other hand, in order to segment inhomogeneous images, the local intensity clustering (LIC) property of bias field has been considered as an important prior in various segmentation models. In this paper, we propose a new selective segmentation model based on the LIC property and spatial smoothness prior of bias field. To our knowledge, it is the first time that the LIC property is applied in selective segmentation of inhomogeneous images, which is different from the existing distance-based models. In addition, existence of the minimizers to the proposed model is established. By using the heat kernel convolution approximation, we design an effective and fast convergent numerical algorithm for solving the model, which significantly improves the operational speed. Furthermore, the stability of the proposed algorithm is proven. Extensive experiments on synthetic images, brain MR images, liver CT images and LiTS dataset demonstrate the effectiveness of the proposed method, and our method outperforms state-of-the-art methods in terms of selective segmentation accuracy and running efficiency.