Construction of Digital Statistical Atlases of the Liver and their Applications to Computer-Aided Diagnosis

Introduction

The atlas of human anatomy is an important teaching tool in the medical community (Netter, 2006). In the recent years, digital atlases of human anatomy have become a popular topic in the field of medical image analysis (Hoene, Pflesser, & Pommert, 1995; Hunter & Borg, 2003; Sato, 2007). The digital atlas can be categorized as a probabilistic atlas, a statistical shape atlas (statistical shape model) and a statistical texture atlas (statistical texture model). The basic idea of the digital atlas is to capture the variability of the organ’s location, shape and voxel intensity (texture) from a training set (either data from different individuals (inter-patient variability) or from the same individual (intra-patient variability)). The atlas is also very useful for organ segmentation and computer-aided diagnosis (CAD) because it can be used as a priori information of human anatomy. To date, only a few researchers have constructed probabilistic atlases and statistical shape models of anatomical organs, such as the brain (Tompson et al., 2000), heart (Peyrat et al., 2007), liver (Okada et al., 2008), and spleen (Tateyama, Foruzan, & Chen, 2009). The probabilistic atlas and the statistical shape model have also been applied to automatic segmentation of medical images (Okada et al., 2008; Linguraru et al., 2009).

In this chapter, we present a framework for constructing probabilistic atlases, statistical shape atlases and statistical texture atlases of the liver from a set of training samples and discuss their applications to liver segmentation and diagnosis of hepatic disease. Our framework is shown in Figure1.

Figure 1.

A framework for constructing digital atlases

Probabilistic Atlases Of The Liver

A probabilistic atlas of the liver is a mean model of the liver’s location in the abdomen. We used 10 patients’ abdominal CT scans and manually segmented the livers. A 3D affine registration was used to align the data. The 3D affine registration had seven parameters: three for translations, three for rotation angles and one for the scaling factor. This kind of registration can eliminate global differences while keeping local differences for the modeling. The probabilistic atlas is constructed by calculating the mean volume of the aligned livers. Typical slice images of the liver probabilistic atlas are shown in Figure 2.

Figure 2.

Probabilistic atlas of the liver

The probabilistic atlas of the liver can be used as the a priori probability of the liver for liver segmentation. In conventional liver segmentations, the probability of the intensity in the liver region (likelihood) is used for liver segmentation. Figure 3(a) shows an example of liver segmentation using the conventional method (maximum likelihood). This method detects many false positives because other tissues, like the kidney, have similar intensity distributions (likelihood) as the liver. So the intensity information alone is not enough to separate the liver and the kidney. However, by a priori knowledge, we know the liver is located in a different part of the body than the kidney. Because the probabilistic atlas includes the anatomical location information, it can be used as the a priori probability of the liver for liver segmentation by maximizing the following equation:, where ω represents liver class, x represents the organ’s intensity, is the a priori probability (probabilistic atlas) and is the likelihood of the liver. The segmentation results from the probabilistic atlas are shown in Figure 3(b). The probabilistic atlas removes a significant amount of false positives and accurately segments the liver (Yasuda et al., 2008).