| 题名 | 医学双目光谱成像多数据融合技术及其智能分类方法研究 |
| 作者 | 卜昌郁 |
| 学位类别 | 硕士 |
| 答辩日期 | 2014 |
| 授予单位 | 中国科学院上海光学精密机械研究所 |
| 导师 | 阮昊 |
| 关键词 | 数据融合,内窥镜检测,光谱成像,空间成像,智能聚类 |
| 其他题名 | Medical binocular and multi-spectral imaging using data fusion technology with intelligent classification |
| 中文摘要 | 在许多领域，由于数据集过度庞大，科学家经常在分析处理上遭遇限制和阻碍；这些领域包括气象学、基因组学、神经网络体学、复杂的物理模拟，以及生物和环境研究。这样的限制也对网络搜索、金融与经济信息学造成影响。数据集大小增长的部份原因来自于信息持续从各种来源被广泛收集。 本文主要是针对如今的内窥镜检测技术检测手段的单一、检测结果的不全面以及对医护人员的依赖性较强的缺点，以及医学检测领域图像检测的重要性以及医学检测手段需要向更加快速、更加精确以及需要结合大数据和云技术等发展需求，提出一种高效、低成本的双目空间成像与光谱成像内窥镜检测系统，并通过软件与算法上的深入研究，提出一种对硬件要求不高的情况下的实时性强、融合数据信息量多（包括待检测部位的物质成分分析、内窥镜移动路径图像全场景以及特定部位的立体信息）、且可以智能自动对数据进行分析分类且低成本的医学检测系统雏形。 1）首先针对如今内窥镜系统的不足，提出一种多数据融合内窥检测系统原型（其中包括用于光谱信息采集的单元以及基于双目传感器的空间环境重建和立体成像单元），可以提供实现单次检测即可获得检测部位的大场景彩色图像以及具有深度信息的图像，并同时获得物质光谱反射率信息的硬件基础；2）针对空间环境重建和立体成像的需求，提出一种基于双目视觉的图像空间数据融合和基于空间坐标的深度信息的获取和模拟显示。这种算法基于双目CCD图像信息，实现图像的拼接已获得更大的空间感官，从而在探测中获得更大的检测区域信息，同时结合图像特征，将普通图像转化为具有深度特性的RGB-D图像；3）针对光谱信息采集的需求，提出了一种光谱成像信息提取和融合算法。这种算法可以将光谱成像数据序列中分析并提取出感兴趣的区域，并重绘兴趣区域的光谱曲线，然后通过与先验知识的比对，得到异常区域的位置以及该区域的光谱特性。即利用成像的方法进行物质成分分析，并对光谱成像数据处理进行了算法和软件系统的优化；4）针对以上数据融合中需要处理海量的数据量，因此考虑有必要对待处理数据进行压缩处理以实现数据更快的处理：我们结合压缩感知理论，提出一种对多光谱数据进行欠采样压缩，并进行后稀疏化模拟实现，使得光谱成像系统实时性性能的进一步提高；5）最后对于内窥设备对医护人员的依赖性较强的问题，提出一种光谱检测智能算法：S-HTM（Spectral-Hierarchical Temporal Memory）算法，实现检测结果的自动分类。通过这种智能算法对于所设计的仪器采集到的光谱图像中光谱指纹进行智能分析和分类，通过对光谱数据库的提前分析，将光谱数据库中的数据进行聚类，对需检测的光谱数据进行自动特征提取和拟合，实现待检测光谱数据非人为干预的自动聚类检测。论文中将血红蛋白作为实验物质，用S-HTM算法自动对正常与非正常区域的血红蛋白进行自动分析，实现较高的基于血红蛋白的正常、非正常区域自动分类。实现空间信息和光谱信息的多数据融合以及自动检测。 |
| 英文摘要 | Scientists often encounter restrictions and obstacles in the analysis process because of the large amount of data sets in many areas, including meteorology , genomics ,neural network learning body, complex physical modeling,biological and environmental research. This phenomenon also affect on Web search , financial and economic information science. The growth of the size of the data set is partly influenced by the collection of a variety of sources . Endoscopic detection technology is not complete, cannot get enough information and strong dependence on doctors. The importance of image analysis in medical testing, the needs of more faster and more preciser of medical testing and Development of big data and cloud technologies, is the starting point of this project. We put forward an efficient, low-cost endoscopic spectral imaging detection system, with our research on related software and algorithm, set up a system, which can do strong case for real-time hardware, fuse more data ( including parts of the substance to be detected component analysis , three-dimensional image information endoscope movement path and the specific parts of the whole scene) and intelligently classify automatically analyzes the data and cost less. Firstly, the system, which is called “multi-spectral imaging endoscopic detection system” and its promotion system for 3D image and more image details, can get larger parts of the detected area, the scene image with RGB-D depth information and spectral reflectance ratio information by one time detection. Then, we come up with an algorithm based on binocular vision spatial information and depth information fusion, and we can get the related information and show to users. Because this algorithm is based on binocular CCD image information, this system cans sense greater space and turn common image to RGB-Depth（RGB-D） image with 3D information. Spectral imaging information extraction and fusion algorithm of our own can analyze the sequence of spectral imaging data and extract the region of what we want, and redraw the spectral curve of interest region. Comparing with the prior knowledge we can obtain position of the abnormal region and its spectral information. And we optimize spectral imaging data processing algorithms and software with normal algorithm and Compressed sensing, and proposed a multi-spectral data compression method due to sampling and post-simulation sparse , making further improve for spectral imaging system and its real-time performance. Finally, we propose one kind of spectroscopy intelligent algorithm: S-HTM（Spectral-Hierarchical Temporal Memory）for the design of the medical instrument which is used to collect the spectral fingerprints, analyse spectral image and classification intelligently. By analyzing the spectrum of the database in advance, clustering the spectral data in the database, extracting feature of spectral data and fitting automatically. We use hemoglobin as a test substance, automatic analyze the difference between normal and abnormal regions automatically analyze algorithms using S-HTM, improve the accuracy of classification of normal and abnormal area automatically. |
| 语种 | 中文 |
| 内容类型 | 学位论文 |
| 源URL | [http://ir.siom.ac.cn/handle/181231/16851]  |
| 专题 | 上海光学精密机械研究所_学位论文 |
| 推荐引用方式 GB/T 7714 | 卜昌郁. 医学双目光谱成像多数据融合技术及其智能分类方法研究[D]. 中国科学院上海光学精密机械研究所. 2014. |
| 个性服务 |
| 查看访问统计 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
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。
修改评论