Semi-automatic liver segmentation based on probabilistic models and anatomical constraints

Abstract

Segmenting a liver and its peripherals from abdominal computed tomography is a crucial step toward computer aided diagnosis and therapeutic intervention. Despite the recent advances in computing methods, faithfully segmenting the liver has remained a challenging task, due to indefinite boundary, intensity inhomogeneity, and anatomical variations across subjects. In this paper, a semi-automatic segmentation method based on multivariable normal distribution of liver tissues and graph-cut sub-division is presented. Although it is not fully automated, the method minimally involves human interactions. Specifically, it consists of three main stages. Firstly, a subject specific probabilistic model was built from an interior patch, surrounding a seed point specified by the user. Secondly, an iterative assignment of pixel labels was applied to gradually update the probabilistic map of the tissues based on spatio-contextual information. Finally, the graph-cut model was optimized to extract the 3D liver from the image. During post-processing, overly segmented nodal regions due to fuzzy tissue separation were removed, maintaining its correct anatomy by using robust bottleneck detection with adjacent contour constraint. The proposed system was implemented and validated on the MICCAI SLIVER07 dataset. The experimental results were benchmarked against the state-of-the-art methods, based on major clinically relevant metrics. Both visual and numerical assessments reported herein indicated that the proposed system could improve the accuracy and reliability of asymptomatic liver segmentation.

Symmetric Reconstruction of Functional Liver Segments and Cross-Individual Correspondence of Hepatectomy

“…Accordingly, the main contributions of this study are threefold. Firstly, the probability distribution of abdominal tissues, vital to most existing methods, In this paper, the extension and improvements of a previous work [22] are presented. Specifically, the MND model was built from multiple seed points (instead of one) to better capture the inhomogeneity of pixel intensities.…”

“…Then, for each pixel q ∈ Ω p , local statistics, i.e., µ q and σ q , were computed within its local neighborhood, Ω q , and associated with q. Subsequently, a bivariate Gaussian model was created from these attributes around the seed point, Ω p , as given in Equation (1) [22]:…”

“…where F q is a feature vector at point q, i.e., F q = µ q σ q T , F is its average over a local patch Ω p , and Σ is a covariance matrix of these attributes. Thus, a total of (N + N − 1) 2 pixels were involved in building the model for a seed point, where N is the width (or height) of Ω p,q [22]. As pointed out in earlier studies, the liver interior in a typical CT exhibits inhomogeneity, and hence results in an uneven probability map.…”

“…In addition, to prevent mistaking fissures for ligamentum teres and venosum for BN, only pairs (p i , p j ) whose distances were greater than 60 pixels were considered. It differed from [5] that, we adopted an outer angle criterion (T c ) to limit the search space [22]. A point p whose outer angle was less than T c would be discarded from potential BN, and hence not detected by Equation (12).…”

“…Over the years, numerous techniques for segmenting a 3D liver from CT have been proposed. These techniques relied on different features being extracted from imaging data, and can be generally classified into those based on thresholding [ 11 , 15 ], region growing (RG) [ 13 , 16 , 17 ], graph processing [ 5 , 18 , 19 , 20 , 21 , 22 ], machine learning [ 23 , 24 ], level-set [ 25 ], and deformable model [ 26 ]. Detailed survey and discussion on the state of the art are presented in [ 27 , 28 ] and summarized as follows:…”