Fully automatic liver segmentation in CT images using modified graph cuts and feature detection

Highlights

• A fully automatic and fast liver segmentation procedure for CT images is proposed.

• A contrast term is integrated in the graph cuts to enhance its segmentation ability in weak boundary and prevent leakage.

• A feature detection method is proposed to identify and remove redundant vessels after graph cuts.

• The procedure can effectively prevent leakage and proved to be accurate and fast for liver segmentation in clinical.

Abstract

Purpose

Liver segmentation from CT images is a fundamental step in trajectory planning for computer-assisted interventional surgery. In this paper, we present a fully automatic procedure using modified graph cuts and feature detection for accurate and fast liver segmentation.

Methods

The initial slice and seeds of graph cuts are automatically determined using an intensity-based method with prior position information. A contrast term based on the similarities and differences of local organs across multi-slices is proposed to enhance the weak boundaries of soft organs and to prevent over-segmentation. The term is then integrated into the graph cuts for automatic slice segmentation. Patient-specific intensity and shape constraints of neighboring slices are also used to prevent leakage. Finally, a feature detection method based on vessel anatomical information is proposed to eliminate the adjacent inferior vena cava with similar intensities.

Results

We performed experiments on 20 Sliver07, 20 3Dircadb datasets and local clinical datasets. The average volumetric overlap error, volume difference, symmetric surface distance and volume processing time were 5.3%, −0.6%, 1.0 mm, 17.8 s for the Sliver07 dataset and 8.6%, 0.7%, 1.6 mm, 12.7 s for the 3Dircadb dataset, respectively.

Conclusions

The proposed method can effectively extract the liver from low contrast and complex backgrounds without training samples. It is fully automatic, accurate and fast for liver segmentation in clinical settings.

Introduction

Accurate trajectory planning of liver tumors is important in computer-assisted interventional liver surgeries, such as biopsy or ablation [1,2]. In surgical planning, liver contour, tumor size and the relative positions of the liver and tumor to certain vital structures (such as liver vessels, bone and the lungs) should be considered to determine a safe access route and precise treatment. Liver segmentation from CT images is a fundamental step of automatic liver tumor and vessel segmentation, 3D visualization and other additional steps of interventional surgery planning, such as multiphase registration and automatic path planning [2,3]. Manual delineation of liver contour on each slice is tedious and time-consuming, and the amount of time available for surgical planning is sometimes restricted to only a few minutes [4]. Therefore, there is a high demand for fast and accurate liver segmentation in clinical settings [5,6].

Liver segmentation remains challenging due to the low contrast of the liver against surrounding organs, blurred boundaries, adjacent vessels that have various appearances and pathologies with heterogeneous densities. Additionally, different liver shapes, sizes and positions increase the difficulty of automatic segmentation.

Current liver segmentation methods can be roughly classified as pixel-based methods [7,8], shape model-based methods [[9], [10], [11]], deformable model-based methods [[12], [13], [14], [15]], graph cuts-based methods [[16], [17], [18]] and machine learning-based methods [[19], [20], [21]]. Because boundary properties and shape constraints are not utilized in pixel-based methods, such methods may fail in liver tumor division or leak into adjacent organs [5,8]. Shape model-based methods (statistical shape models and probabilities atlases, etc.) can effectively realize the initialization or refinement of liver segmentation, but they depend on the shapes of training samples, and their registration step is time-consuming [5,9]. Deformable model-based methods (level set, etc.) are usually sensitive to initial contours or parameters [[12], [13], [14]]. Recently reported machine learning-based methods, especially deep learning methods, have efficiently extracted useful features and achieved good segmentation. However, they still heavily depend on the training sample type and some demand expensive hardware setups [20,21].

Graph cuts-based methods can combine boundary regularization with the regional and shape properties of images. They have good global optimization properties and topology variability and are popular for use in image segmentation [22]. Beichel et al. [16] acquired an initial liver contour by means of graph cuts and refined it by interactive segmentation methods. Afifi et al. [17] manually defined the contour of an initial liver slice and embedded the shape and intensity constraints estimated from previously segmented results into graph cuts for segmentation of the next slice. However, human interventions are needed in these algorithms. For automatic initialization, a liver probability map learned from a convolutional neural network [20] or deformable shape model [23] was applied to initially locate the liver region. Liao et al. chose a slice located around one-third of the way through volume as the initial slice and segmented it via density peak clustering [18]. Some additional organ division and operations were needed to prevent leakage into nearby tissues with similar intensity. Wu et al. [24] extracted background seeds of the kidney and the heart to prevent over-segmentation. Ribs were previously segmented and connected to avoid leakage into subcostal tissue [25]. A previous study removed the inferior vena cava (IVC) with a similar intensity to that of the liver by detecting circle-like objects with a predefined radius [8], whereas another study compensated for the IVC with a different intensity to that of the liver by border marching [18]. These methods achieved satisfactory results in most situations, although researchers still encounter problems with automatic segmentation [16,17,26], over-segmentation of adjacent organs [20] and under-segmentation of liver tumors [18,24].

In this paper, we propose a liver segmentation procedure using modified graph cuts and feature detection. The algorithm first locates an initial liver slice by k-means clustering and thresholding with a position constraint, and then segments the slice by thresholding and concave filling. The intensity and shape variations of different organs in multi-slices are integrated into graph cuts for precise slice segmentation and to prevent leakage into nearby organs. Finally, the adjacent IVC with a similar intensity to that of the liver is eliminated by an anatomical-based feature detection method. The proposed method is more accurate and faster than most previously reported algorithms. It is fully automatic and can be used as a replacement for manual segmentation in clinical settings.