虚拟结肠镜（virtual colonoscopy, VC）计算机辅助检测（computer-aided detection, CAD）是利用息肉与正常组织形态以及其他特征的差异由计算机实现结肠息肉自动检测的新的技术手段，与医师直接使用虚拟结肠镜进行结肠病变检查相比，基于CAD的VC系统有望提供更客观一致的检测结果，提高检测速度，从而促进了VC在普查与体检方面的临床应用。在VC过去几年的发展中，出现了许多基于形态特征的CAD息肉检测方法，在息肉检测方面敏感度较高。但是由于在CT影像上结肠息肉的密度与周围组织的差别较小，而且结肠内部存在着各种与息肉形态相似的结构，基于几何形态的方法检测结果中有许多的假阳性检测点（false positives, FPs）。除几何形态存在差异外息肉的纹理分布模式与正常的结肠壁组织及结肠内粪便等杂质存在着差异。为减少CAD检测的假阳性率，本文提出了一种基于3D形态/纹理分析的虚拟结肠镜辅助检测方法。为了获取可以区域的体素用于纹理分析，使用考虑部分容积效应的混合分割算法提取出结肠壁的粘膜层。使用息肉几何特征与3D纹理特征相结合用于息肉检测。首先使用曲度（curvedness,CV）形状指数（shape index, SI）等几何特征定位疑似区域。然后通过寻找疑似息肉区域的外边界以及侵入肠壁内的内边界建立了3DeROI模型。使用了灰度共生矩阵（GLCM）和灰度-梯度共生矩阵（GLGCM）两种常用的纹理分析法对3DeROI模型结构的疑似息肉区域提取纹理特征。灰度共生矩阵作为一种常用的纹理分析方法通常用于二维图像纹理分析。而由于已获取的三维eROI模型结构的疑似息肉区域为三维图像数据，为了获取三维纹理信息将灰度共生矩阵直接扩展到三维图像分析。传统的基于二维图像的灰度共生矩阵主要从0°、45°、90°、135°四个方向建立，而在三维空间中这四个方向无法全面反映纹理变化。因此我们在三维空间中采用(0°,90°), (45°,45°), (45°,135°), (90°,90°), (135°,45°), (135°,135°), (-°,0°)7个方向建立灰度共生矩阵。为每一个疑似息肉区域建立灰度共生矩阵后可以从中计算出若干纹理特征。除灰度变化外图像密度的梯度变化也可以反映纹理变化趋势，因此将灰度-梯度共生矩阵应用于三维疑似息肉区域图像并从中提取出一些纹理特征。使用这两种纹理分析方法可以从中计算出许多纹理特征，通过实验分析选择了能量、相关、熵、小梯度优势、梯度方差等12个有效的纹理特征量构成特征向量用于后续的分类判别。为了有效地利用已提取的纹理特征对疑似息肉区域进行分类，我们设计了BP神经网路（back propagation neural network，BPNN）和SVM（support vector machine）两种分类器用于息肉的分类判别。分类器的输入为通过灰度共生矩阵和灰度-梯度共生矩阵计算出的12个纹理特征量组成的特征向量。其中BP神经网络的结构采用一个隐藏层和8个隐藏节点，使用sigmoid函数作为激励函数。而SVM的设计使用径向基函数（Radial Basis Function，RBF）为其核函数。使用包含仰卧和侧卧体位的7个病人的CT扫描数据对分类器进行训练和测试，结果表明本文所提出的CAD方法能够有效检测出结肠中的息肉，并除去77.5%的假阳性检测点。而SVM分类器在检测直径大于5mm息肉方面性能优于BP神经网络。

As a new technique, computer-aided detection (CAD) for virtual colonoscopy (VC) utilizes the difference in morphology and other features between polyps and normal tissues to detect colonic polyps automatically. Compared with physicians’ performance using VC directly, CAD-based system can provide objective and consistent results, facilitate interpretation process, and be anticipated a promising mass screening for colorectal cancer in clinics. With the development of VC in past several years, many CAD schemes based on morphological features have been studied and a relatively high sensitivity has been achieved for polyp detection. However, surface geometry-based CAD may cause many false positives in detection results because the densities of the CT image voxels inside a polyp differ only subtly from those of the surrounding tissues and the shape of the colon surface contains a variety of structures that can mimic polyp shapes. Besides geometric morphology，a polyp also has different texture distribution patterns from normal colon tissues and other things in colon, such as stool. To low false negatives, a 3D texture-base CAD for VC is proposed in this paper. To obtain the entire volume of each suspicious patch for texture analysis, the mucosa layer was first extracted by tissue-mixture image segmentation mitigating the partial volume effect. The geometric and texture features are all utilized for detection. Firstly, geometric features i.e., curvedness(CV) and shape index (SI) are employed for extracting the suspicious patch. Then via searching the outer border and inner border which usually invading into the colon wall, a 3DeROI model was established for each suspicious patch. Two traditional texture analyses were performed to extract 3D texture information from suspicious patch with 3DeROI model, i.e., grey level cooccurrence matrix (GLCM) and grey level-gradient cooccurrence matrix (GLGCM).As one of the most widely-used texture analysis methods, GLCM is usually used for analysis of 2D images. Due to the 3D nature of eROI model for each polyp candidate, we expanded GLCM to 3D situation, getting 3D features directly. Traditional GLCM for 2D images usually uses four directions of 0°, 45°, 90°, and 135°. In 3D space, these directions could not cover the texture variance of the entire volume. Therefore, seven directions in 3D space were adopted for the construction of GLCM, they are (0°,90°), (45°,45°), (45°,135°), (90°,90°), (135°,45°), (135°,135°), (-°,0°). Once 3D GLCM was computed from a polyp candidate, a number of texture features could obtained directly from the GLCM. In addition to image density, density gradient can also reflect the variation trend of image texture. To get more valid texture features, GLGCM was also expanded to 3D situation and computed for each suspicious candidate. With these two texture analysis methods, we got a lot of derivative features and compared their performance on polyp detection. Finally 12 of them, such as Energy, Correlation, Entropy and GradsVariance, were selected to comprise the feature vector of texture for further classification.To develop a weighted classification scheme that can use the most appropriate classifier for extracted features, two kinds of classifiers, i.e., the back propagation neural network (BPNN) and the support vector machine (SVM), were used to classify polyps from normal tissues. The input of classifier was the feature vector integrating the 12 texture features computed from GLCM and GLGCM. In this study, the BPNN classifier used was multi-layer neural network with one hidden layer and 6 hidden neurons, using a sigmoid function as the activation function. SVM classifier was built with RBF(Radial Basis Function) as its kernel function. These two classifiers were trained and tested by 7 patients’ CT data acquired at both prone and supine position. Experimental result shows that the CAD method proposed in this paper could detect polyps in colon effectively and eliminate 77.5% false positives while the SVM classifier outperforms the BPNN classifier in polyp detection larger than 5mm.