Automatic liver detection and segmentation from 3D CT images: a hybrid method using statistical pose model and probabilistic atlas
文献类型：期刊论文
| 作者 | C. Huang; F. Jia; C. Fang; Y. Fan; Q.Hu |
| 刊名 | International Journal of Computer Assisted Radiology and Surgery
 |
| 出版日期 | 2013 |
| 英文摘要 | Purpose:Precise, robust and fully automatic segmentation of liver parenchyma from medical images is the first step and one of the major tasks in liver surgical planning. Fast and automatic liver detection is the first step to achieve good liver segmentation, due to wide variable scan- ning range of abdominal CT images, automatic liver detection is often unreliable which could cause total failure in the subsequent extraction of liver surface. Various kinds of approaches have been proposed to try to achieve robust liver segmentation. Here, a hybrid method is proposed that combines a novel statistical pose model (SPM) with probabilistic atlas, for fast, robust and fully automatic liver detection and segmentation.Methods:To detect liver in CT images, we build a SPM and align this model to the target image to be segmented. The term ‘‘pose’’ simply refers to the spatial position and rotation of liver. The variation of liver poses can be represented in a related fashion. To build a SPM, first, one reference image is chosen from training image datasets. Then all the other training image datasets are aligned to the reference image using rigid transform. Then a rigid transform vector is generated for each training data Ti = {tx, ty, tz, ra, rb, rc}, with tx, ty, and tz for translation and ra, rb, and rc for rotation angles, respectively. The transform vector is then normalized to be a standard pose vector Pi = {tx/Mx, ty/ My, tz/Mz, ra/2p, rb/2p, rc/2p}, with Mx, My, and Mz for the maximum distance along x, y and z direction of all training data respectively. The SPM is constructed by a principal component analysis (PCA) of the training poses. Eigenvectors p1, ..., pk corresponding to the first k maximum eigenvalues of the SPM’s eigen space are selected so that it covers 98 % of the variance of the observed poses. With SPM, liver is detected by aligning the reference image to the target image using mutual information (MI) metric [1], B-Spline registration [2] and evolution optimizer [3] in which the parameters to be optimized is the SPM parameters corresponding to the first k eigenvalues.Accurate liver surface extraction is done by fitting the reference image or the reference atlas, to the target image using non-rigid registration constrained by a probabilistic atlas constructed from manual segmented training data. Let Bi(x) be a binary image where liver region is assigned to 1 and others to 0, then value of the atlas P(x) is defined as the average of Bi(x) over n training datasets. Gaussian filtering is applied on P(x) to make its value smoother. Consequently, we introduce probabilistic constraint by designing the energy function as EðF;MÞ1⁄4Pfmiðfx;mxÞ pðxÞg, where mi(fx, mx) refers to a metric function and p(x) refers to the probabilistic value of the coordinate x, in this case, mutual information. Figure 1 demonstrates how the probabilistic atlas prevents over segmentation and improves results.Results:We build the SPM and probabilistic atlas over twenty CT volumes of training datasets and validate our method on ten testing datasets. Both datasets are provided by MICCAI 2007 liver segmentation challenge (http://www.sliver07.org), which have in-plane resolution of 512 9 512 pixels and inter-slice spacing varying from 0.5 to 5.0 mm. Figure 2 shows several evaluation measurements of the segmentation results. The overall segmentation measurement on the 10 testing datasets is as follows: volumetric overlap error 7.55 ± 1.77 %, rel- ative absolute volume difference -1.30 ± 3.62 %, symmetric root mean square surface distance 2.35 ± 0.95 mm, average surface dis- tance 1.28 ± 0.38 mm, and maximum surface distance 22.10 ± 10.27 mm. Our method reaches an average segmentation score of 72.40 with six of ten datasets scoring at least 75. |
| 收录类别 | 其他 |
| 原文出处 | http://link.springer.com/article/10.1007/s11548-013-0879-6#page-2 |
| 语种 | 英语 |
| 源URL | [http://ir.siat.ac.cn:8080/handle/172644/4852]  |
| 专题 | 深圳先进技术研究院_医工所 |
| 作者单位 | International Journal of Computer Assisted Radiology and Surgery |
| 推荐引用方式 GB/T 7714 | C. Huang,F. Jia,C. Fang,et al. Automatic liver detection and segmentation from 3D CT images: a hybrid method using statistical pose model and probabilistic atlas[J]. International Journal of Computer Assisted Radiology and Surgery,2013. |
| APA | C. Huang,F. Jia,C. Fang,Y. Fan,&Q.Hu.(2013).Automatic liver detection and segmentation from 3D CT images: a hybrid method using statistical pose model and probabilistic atlas.International Journal of Computer Assisted Radiology and Surgery. |
| MLA | C. Huang,et al."Automatic liver detection and segmentation from 3D CT images: a hybrid method using statistical pose model and probabilistic atlas".International Journal of Computer Assisted Radiology and Surgery (2013). |
入库方式： OAI收割
来源：深圳先进技术研究院
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。