【作者】 孙涛；

【导师】 冯焕清； 李传富；

【作者基本信息】 中国科学技术大学 ， 生物医学工程， 2009， 硕士

【摘要】 随着医学影像技术的飞速发展,包括多排螺旋CT在内的先进的成像设备开始在临床上广泛使用。它们产生的影像资料清晰度越来越高,数据量也越来越大,一方面为影像医师提供了更清晰的医学图像,但另一方面也给他们带来了很大的读片负荷。在此背景下,计算机辅助诊断成为解决此问题的一个有效途径。在颅脑病变的检查中,许多MRI图像难以应用的场合,高分辨率CT图像是一种十分重要的替代手段,针对脑出血和脑肿瘤等病变更是具有自己独特的优势,具有不可替代的地位。开展基于颅脑CT图像的计算机辅助诊断的研究,需要针对CT图像的特点研究诸如特征提取、图像配准和图像分割在内的一系列相关的技术。本文利用医学图像纹理分析技术提出一种基于局部直方图的几何不变矩。通过提取不同尺度下的局部直方图,计算能够反映图像纹理特征的几何不变矩,并融合了图像像素的边界信息,为图像中的每个像素建立了属于其自身的特征向量作为其形态学签名。实验证明,特征向量将图像的灰度信息映射到特征空间,从而增大了属于同一组织的像素的相似性和分属不同组织的像素的差异性,具有良好的纹理特征区分性能。在实现计算机辅助诊断的研究中,图像的非刚性配准是关键的一步。本文利用具有相同解剖位置的像素的特征向量的相似性,实现了对应特征点的自动搜索和准确定位,解决了基于特征点的非刚性配准算法需要手动选择标志点的问题。更进一步,利用基于特征向量的对应特征点自动搜索算法的良好性能,通过将特征提取与近似薄板样条非刚性配准相结合,提出了一种特征向量驱动的颅脑CT图像配准算法。该配准算法保证了颅脑CT图像上具有关键解剖意义的点的一一对应,不仅适用于正常图像的配准,而且也适用于病变图像的配准。图像分割在基于数字化图谱的病变检出中,是一种重要的定量分析手段。特征向量具有良好的组织区分能力,能够区分出颅脑CT图像上分属不同组织但具有相似或相同灰度值的像素。用改进的模糊C均值聚类的方法将特征空间的点聚集成对应不同组织区域的类团,并将分类结果映射回图像空间,在颅脑CT图像颅内空间的组织分割中取得了良好的效果,能够分割出完整的白质、灰质和脑脊液。课题研究工作的成果较好的实现了计算机辅助诊断研究中所需的特征提取、非刚性配准和图像分割等技术,具有实际应用价值。本论文的工作得到了国家自然科学基金项目(60771007)的资助。更多还原

【Abstract】 With the rapid development of medical imaging technology, some advanced imaging equipments, such as Multi-slice Spiral CT, are clinically widely used. The volume data of high quality and large quantity, which is produced by these equipments, can provide the doctors with clearer medical images, while at the same time, also put heavy reading burden on them. The Computer Aided Diagnosis (CAD) system can be an effective means to solve the problem.The high resolution CT (HRCT) image is an important substitute for MRI image in some occasions where MRI cannot be used. In particular, the HRCT image has a unique advantage for the detection of brain tumors and cerebral hemorrhage. In order to perform the research of the CT image-based Computer Aided Diagnosis (CAD), many related algorithms, like feature extraction, image registration and image segmentation should be designed and studied.In this paper, the local histogram-based Geometrical Moment Invariants (GMI) is proposed on the basis of medical image texture analysis algorithm. Eigenvector is constructed by both calculating the moments based on the local histogram in different scale regions of interest (ROI) and incorporating the boundary information of every pixel. Eigenvector can be regarded as the morphology signature of the pixel. The experiment results show that the eigenvectors, mapping the information of CT image from intensity space into feature space, aggrandize both the similarity of the pixels in the same tissue and discrepancy of the pixels belonging to different tissues, thus demonstrate its good performance in discriminating the medical image texture feature.The non-rigid registration is a key process in the CAD system. An eigenvector-based corresponding point automatic detection algorithm is proposed to determine the corresponding landmark points between moving images and fixed images automatically and correctly. The algorithm successfully solved the problem that corresponding point must be labeled manually in the point-based non-rigid registration. Furthermore, this advantage is used in the Approximating Thin-Plate Splines (ATPS) scheme to design an eigenvector-based non-rigid registration algorithm for brain CT images. The registration algorithm can guarantee the one-to-one mapping between the corresponding points in the key anatomical positions and can be applied for the registration of both normal medical images and the abnormal ones.In the Atlas-based pathology automatic detection, image segmentation is an important quantitatively analysis method. Eigenvectors can discriminate the pixels of which the intensities are same or similar but belong to different tissues. Therefore, they can be classified by Modified Fuzzy C-mean Method (MFCM) in the feature space and the classification result will be mapped into intensity space. The cerebrum area on brain CT images, in this way, can be segmented into white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF).In this research, the algorithms of texture feature extraction, medical non-rigid registration and medical segmentation has been realized. These algorithms can be applied in the Computer-Aided Diagnosis System, so to some extent, they are valuable.This research was sponsored by Nature and Science Foundation of China (Project No. 60771007).更多还原