基于模糊模型和形状特征的CT序列图像分割方法研究

发布时间：2018-08-05 19:17

【摘要】：计算机层析成像(Computed Tomography,CT)作为一种先进的检测技术,以图像形式直观清晰地反映被测对象内部结构和性状,在医学诊断和工业无损检测领域的应用十分广泛。随着CT技术发展和应用需求的不断提高,通过CT图像处理技术实现对被测物体的定量化、自动化分析和测量,以克服定性的主观评价的不足是CT技术发展的重要方向之一。CT图像分割是实现图像量化分析、自动识别和测量的关键和难点。本文以典型的CT序列三维图像分割为研究内容,针对医学CT中一类边界模糊的图像分割难题和工业CT中一类体素尺度不同比的图像分割难题开展研究,提出了一种基于物体隶属度的医学全身正电子发射层析成像/计算机层析成像(PET/CT)图像自动解剖结构分割(AAR)方法和一种基于形状特征的工业CT图像裂缝分割方法。主要工作如下:1、提出了一种基于物体隶属度的PET/CT图像AAR方法。针对原有AAR方法在图像质量较好的诊断CT图像解剖结构分割中精度较高,但是在图像质量较差的PET图像(解剖结构较模糊)和低剂量CT图像(对比度较低)解剖结构分割中精度较低的问题,本文利用器官的灰度和纹理特征,提出了一种基于物体隶属度的AAR方法。该方法在建模过程中,提出了结合训练图像灰度和纹理信息的物体隶属度函数,用于估算各体素属于物体的概率;在分割过程中,利用物体隶属度函数得到测试图像的物体隶属度,再结合物体隶属度对物体模型进行初定位和阈值化最优姿势搜索,确定物体模型的最优姿势,最后得到物体的空间分布结果;采用定位误差和尺度误差两个指标,通过PET/CT图像进行了实验验证,结果表明改进方法可实现更高精度的解剖结构分割,平均定位误差仅1-2体素,平均尺度误差接近标准值1。2、改进了AAR方法的最优阈值训练方法。针对AAR方法中,原有最优阈值训练方法搜索空间维度高、适应性差、仅适用于灰度图像的问题,本文利用超掩模和累计灰度直方图,提出了改进的最优阈值训练方法。即在超掩模下计算目标和背景的累计灰度直方图,并在任意可能的阈值区间下求两个直方图面积绝对差,选取使绝对差值最大的阈值为最优阈值。改进方法将搜索空间从5维降为1维,实现了高效率的最优阈值搜索,避免了因限制阈值搜索范围而丢失可能的最优阈值。实验结果表明,改进方法能够适用于灰度、纹理和隶属度图像,输出合理的物体阈值以实现更精确的解剖结构分割。3、改进了AAR方法的层级结构。原有AAR方法只适用于胸腹部等局部身体区域图像,需要将全身图像手动地分割为局部身体区域。为了提高自动化程度,本文利用全身各个器官之间的解剖结构关系,提出了全身层级结构,即全身所有器官以一个树状层级结构表示,按广度优先遍历所有器官依次进行建模和分割。通过全身躯干PET/CT图像进行了实验验证,结果表明改进方法可实现全身躯干解剖结构精确分割,提高了自动化程度。4、提出了不同成像模式间的建模-初步分割方案。通常模型类方法中,成像模式内的建模-初步分割需要来自同一种成像模式的训练数据,未考虑建立通用于各种成像模式的快速原型的可能性。本论文利用模糊模型包含物体形状和空间位置信息而独立于成像模式的优点,在AAR方法建模和初步分割两个基本步骤上,提出了成像模式间的建模-初步分割方案。通过实验验证了用诊断CT图像建立的模糊模型在PET、低剂量CT和它们的物体隶属度图像上进行解剖结构分割的可行性。这为建立适用于各种成像模式的快速物体原型提供了一种有效途径。5、提出了一种基于形状特征的工业CT序列图像裂缝分割方法。被测工件内部的裂缝检测、自动显示与测量是工业CT需要解决的难点之一,而图像分割是关键。在工业CT系统中,获得的三维图像大多由序列断层图像组成,图像中的体素在断层平面内的等效尺寸当量与垂直于断层方向的等效尺寸当量存在很大差别,有时达到10倍以上,加之工业CT图像的各种伪影较严重,这都给裂缝分割及定量测量增加了难度。针对这一问题,本文研究了适用于体素尺寸各向不同性的工业CT序列图像的裂缝自动分割方法:首先采用基于Hessian矩阵的二维线状结构滤波增强图像线状区域;在此基础上,进一步提出了结合层间灰度和方向的连续度以及层内线状邻域灰度平均值的二维直方图,以抑制伪影对裂缝分割的影响;根据直方图的最大类熵确定阈值区间,得到裂缝的二值化分割结果。最后采用准确率、查全率和F1值三个指标,通过实际工件的工业CT序列图像进行了实验验证,结果表明本文提出的方法不仅与其它四种常用的方法相比较,能够得到更完整、更准确的分割结果,满足实际工业CT序列图像裂缝分割精度的要求,而且自动化程度更高。
[Abstract]:Computed Tomography (CT), as an advanced detection technology, reflects the internal structure and character of the object clearly and clearly in the form of image. It is widely used in the field of medical diagnosis and industrial nondestructive testing. With the development and application of CT technology, the application of CT image processing technology is realized. Quantitative, automated analysis and measurement of the measured objects to overcome the lack of qualitative subjective evaluation is one of the important directions of the development of CT technology..CT image segmentation is the key and difficult point to realize image quantization analysis, automatic recognition and measurement. This paper takes a typical CT sequence three-dimensional image segmentation as the research content, aiming at a class of boundary in medical CT The difficult problem of blurring image segmentation and the image segmentation problem of a kind of voxel scale in industrial CT is studied. A kind of automatic anatomical structure segmentation (AAR) method for medical systemic positron emission tomography / computerized tomography (PET/CT) image based on object membership and an industrial CT image crack based on shape feature are proposed. The main work of slit segmentation. The main work is as follows: 1, a PET/CT image AAR method based on object membership is proposed. In view of the original AAR method, the accuracy is higher in the segmentation of the anatomical structure of the diagnostic CT image with better image quality, but the PET image with poor image quality (the dissected structure is more obscure) and the low dose CT image (lower contrast) In this paper, a AAR method based on object membership is proposed in this paper, based on the gray and texture features of organs. In the process of modeling, the membership degree function of the object, which combines the gray and texture information of the training image, is proposed to estimate the probability of each body element belonging to the body, and the object is used in the segmentation process. The subjection function of body membership function obtains the membership degree of the object in the test image, and then combines the membership degree of the object to the initial position of the object model and the optimal position searching of the threshold, determines the optimal position of the object model, and finally obtains the result of the space distribution of the object, and uses the two indexes of the positioning error and the scale error, and the experimental verification is carried out through the PET/CT image. The results show that the improved method can achieve more accurate anatomical structure segmentation, the average positioning error is only 1-2 voxel, the average scale error is close to the standard value 1.2, and the optimal threshold training method of the AAR method is improved. In the AAR method, the original optimal threshold training method has high spatial dimension and poor adaptability, which is only applicable to the problem of gray image. In this paper, an improved optimal threshold training method is proposed by using the hyper mask and cumulative gray histogram. That is, the cumulative gray histogram of the target and the background is calculated under the super mask, and the absolute difference of the two histogram areas is calculated under any possible threshold range, and the optimal threshold is selected to make the maximum absolute difference as the optimal threshold. The improved method will search the search method. The space is reduced from 5 dimension to 1 dimension, and the optimal threshold search is achieved with high efficiency. It avoids the loss of possible optimal threshold by limiting the threshold search range. The experimental results show that the improved method can be applied to gray, texture and membership images, output reasonable object threshold to achieve more accurate anatomical structure segmentation.3, and improve the AAR method. The original AAR method is only suitable for local body area images such as chest and abdomen. It is necessary to manually divide the whole body image into a local body area. In order to improve the degree of automation, this paper uses the anatomical structure of various organs of the body to put forward a whole body structure, that is, the whole body is a tree structure with a tree structure. It is shown that all organs are modeled and segmented in sequence according to the breadth priority. The experimental verification through the PET/CT image of the whole body shows that the improved method can realize the precise segmentation of the body anatomy structure and improve the automation degree.4. The modeling and preliminary segmentation scheme between different imaging modes is proposed. The model class method is usually used. In the imaging mode, modeling - preliminary segmentation requires training data from the same imaging mode, without considering the possibility of building fast prototypes that are commonly used in various imaging modes. This paper uses the fuzzy model to contain the advantages of object shape and spatial location information, which is independent of the imaging mode, modeling and segmenting two preliminary methods in the AAR method. In the basic step, the modeling and preliminary segmentation scheme between imaging modes is proposed. The feasibility of dissecting the anatomical structure of the fuzzy model established by the diagnostic CT image on PET, low dose CT and their object membership image is verified by experiments. This provides a kind of rapid object prototype for various imaging modes. .5, an industrial CT sequence image segmentation method based on shape features is proposed. The detection of cracks inside the workpiece, automatic display and measurement is one of the difficulties that industrial CT needs to solve, and image segmentation is the key. In industrial CT system, the obtained 3D images are mostly composed of sequence fault images and the body in the image. The equivalent size equivalent in the plane of the fault and the equivalent size equivalent perpendicular to the fault direction are very different, sometimes more than 10 times, and the various artifacts of the industrial CT image are more serious. This adds to the difficulty of the fracture segmentation and quantitative measurement. In this paper, the application of the voxel dimension anisotropy is studied in this paper. The method of automatic segmentation of cracks in industrial CT sequence images: first, the two-dimensional linear structure filtering based on Hessian matrix is used to enhance the linear region of the image. On this basis, a two-dimensional histogram which combines the continuity of the interlayer gray scale and direction and the gray mean value of the inner line neighborhood is further proposed to suppress the segmentation of the artifacts. According to the maximum class entropy of the histogram, the threshold interval is determined and the two value segmentation results are obtained. Finally, the accuracy rate, the recall rate and the F1 value are used to verify the experimental results through the industrial CT sequence images of the actual workpiece. The results show that the proposed method is not only compared with the other four commonly used methods, and can be obtained. More complete and more accurate segmentation results meet the actual industrial CT sequence image crack segmentation accuracy requirements, and the degree of automation is higher.
【学位授予单位】：重庆大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】