Serial slice image segmentation of digital human based on adaptive geometric active contour tracking

Abstract

Segmentation is one of the crucial problems for the digital human research, as currently digital human datasets are manually segmented by experts with anatomy knowledge. Due to the thin slice thickness of digital human data, the static slices can be regarded as a sequence of temporal deformation of the same slice. This gives light to the method of object contour tracking for the segmentation task for the digital human data. In this paper, we present an adaptive geometric active contour tracking method, based on a feature image of object contour, to segment tissues in digital human data. The feature image is constructed according to the matching degree of object contour points, image variance and gradient, and statistical models of the object and background colors. Utilizing the characteristics of the feature image, the traditional edge-based geometric active contour model is improved to adaptively evolve curve in any direction instead of the single direction. Experimental results demonstrate that the proposed method is robust to automatically handle the topological changes, and is effective for the segmentation of digital human data.

Introduction

The digital human is important to promote the medical research and the understanding of human being. The Visible Human Project (VHP) was firstly run by the U.S. National Library of Medicine to build a data set of cross-sectional photographs of the human body, including magnetic resonance (MR) images, computed tomography (CT) images, and anatomical images. In addition, the Chinese Visible Human (CVH) and Visible Korean Human (VKH) were also released to represent different populations of the world.

As one of the critical problems of the digital human research, image segmentation is the foundation of many operations, such as three-dimensional (3D) visualization and image measurement. Although a great deal of literature [1], [2], [3] on medical image segmentation are available, only a small number of methods have been proposed for digital human dataset segmentation, such as color similarity based methods [4], [5], [6], [7], Voronoi diagram based method [8], hybrid method [9] and image analogy based method [10]. In RGB feature space, an interactive ellipsoidal classification was presented by Schiemann et al. [4], while the detailed segmentation for a complete 3D anatomical atlas is still necessary. The user's contextual and interpretive knowledge in [5] were utilized in the segmentation and visualization of volumetric data sets. In addition, a spatial/symbolic model of inner organs [6] and a CVH brain image segmentation method [7] were also presented mainly depending on color information of tissues. Due to the non-uniform lighting between digital human slices and the complex composing of color information of tissues, the segmentation precision of color similarity based methods [4], [5], [6], [7] will be influenced. The basic idea of the Voronoi diagram based method [8] is repeatedly to partition and classify an image based on Voronoi diagrams and experimental classification statistics, respectively. Imielinska et al. [9] segmented internal organs by integrating deformable model and region-based segmentation method. In the interactive segmentation process of Voronoi diagram based method [8] and hybrid method [9], human–computer interaction is needed during each iteration, because seeds should be added to the interior of object regions. The image analogy based method [10] required supervised training data, which can only automatically segment neighboring four slices. Therefore, at present the digital human data is mainly segmented by hand or with software, such as Photoshop software [11] and VOXEL-MAN software [12], [13], [14]. In order to improve the automatic level of digital human segmentation, this paper proposes an object tracking based method.

Since digital human datasets contain thin slices (0.1–1 mm for CVH [15]) and most of tissues are soft, the level of continuity between consecutive slices is high. The continuity is taken as the deformation of temporal sequential images on an identical slice, so the serial slices of digital human can be segmented with the geometric active contour tracking method. A parametric active contour model [16], proposed for object tracking, was applied to brain stem segmentation of Chinese visible human datasets. Since the parametric active contour model tackles the segmentation problem by considering an object boundary as a single and connected structure, the previous model cannot segment several objects simultaneously or the objects with topological changes. In addition, the matching degree image was constructed according to the information of object contour points and image gradient, without region statistical information.

To overcome the deficiencies of the model [16], we propose an adaptive geometric active contour model for the segmentation of digital human data in this paper. The construction of the feature image is improved by incorporating the region statistical information into the matching degree image. The feature image means a transformed image generated with image features, such as color and gradient. To the best of our knowledge, there is no previous work about geometric active contour tracking into the digital human segmentation. Fig. 1 shows the flow chart of our method. The main contributions of this paper are: (1) the geometric active contour tracking is firstly used for the segmentation of digital human data; (2) a novel construction of the feature image is presented by improving the matching degree image and incorporating the color models of the object and background colors. In addition, the traditional edge-based geometric active contour model is improved to adaptively evolve the curve in any direction, not in the traditional single direction (inward or outward). The proposed model builds considerably on the previous model [16], and extends and improves the model. By comparing with the model [16], the new model is more robust (such as Fig. 13, Fig. 14) and can handle the topological changes automatically (such as Fig. 8, Fig. 10).