基于联合索引的医学图像检索研究

发布时间：2018-05-16 13:51

本文选题：融合医学图像特征 + 病灶区域特征　；参考：《吉林大学》2017年硕士论文

【摘要】：目前在医学领域,医学成像技术已经成为医生进行医学诊断最为信赖的辅助技术之一。各大医院每天都在借助如CT,核磁共振等各种成像技术来生成不同格式、大小的医学图像。这些海量的医学图像蕴含了大量的来自不同的病人的病理信息,因此借助现代信息科学技术来挖掘医生无法发现的病理规律和共性成为了一个非常有研究价值的课题。基于内容的图像检索技术能够利用现代计算机技术实现自动地在海量图像中高效、准确的识别出与被检索图像在视觉上类似的,具有相同语义的图像。这一技术在医学上可以帮助医生在数量巨大的图像中找到最有价值的相似医学图像,通过发掘更多的具有同样影像特征的病例来为医学临床诊断、医学教学和科研提供更为多方面的有价值的历史数据支持。因此,对于基于内容的医学图像检索的研究在临床诊断和医学研究领域都有重要的意义。本文研究了如下内容:1、提出了基于sift局部不变子特征的EVLAD图像全局特征与基于局部敏感哈希的灰度特征,基于Gabor小波变换算法图像纹理特征相融合的融合医学图像特征的提取和计算方法,实验证明基于本特征表示法的图像相似度匹配算法能够达到比较理想的精度。2、针对在医生临床诊断,教学和科研当中常常存在对特定病灶部位进行精准研究的需要,在本文提出了基于区域分割的针对特定病灶区域的候选集排序算法,通过医生指定的种子点可以对初步检索得到的候选集进行病灶区域的分割,并提取病灶区域特征并依据该特征对候选集进行重排序,实现可疑病灶部位的精确检索和排序。3、提出了联合索引结构,分别独立的对融合医学特征的三种特征进行聚类,并对不同层的聚类中心进行三个一组的全组合,得到多个索引节点。这种索引结构只需计算少数聚类中心就可得到多个索引节点,降低了索引构建的计算量。在索引节点与图像进行相似度计算时引入了权重计算,在提高检索效率的同时保证了检索精度。4、根据医学图像数据量巨大的特点,设计了一套基于本文检索算法的分布式计算架构,通过将每个大型检索任务分解为若干个小任务在不同的硬件上分布式的并行执行,大大提高了算法特征提取、索引构建以及在线查询的运行速度,很好的解决了检索过程中相关计算量巨大的问题。
[Abstract]:At present, medical imaging technology has become one of the most reliable assistant technology in medical diagnosis. Every day, hospitals use various imaging techniques such as CTand NMR to generate medical images in different formats and sizes. These massive medical images contain a large amount of pathological information from different patients, so it is a very valuable subject to explore the pathological rules and commonalities that doctors can not find by means of modern information science and technology. Content-Based Image Retrieval (CBIR) technology can use modern computer technology to automatically identify images with the same semantic semantics as the retrieved images in a large amount of images with high efficiency and accuracy. This technique can help doctors find the most valuable medical images in a large number of images, and diagnose the medical clinic by finding more cases with the same image characteristics. Medical teaching and research provide more valuable historical data support. Therefore, the research of content-based medical image retrieval is of great significance in the field of clinical diagnosis and medical research. In this paper, the following content: 1: 1 is studied. The global feature of EVLAD image based on sift local invariant feature and the gray level feature based on locally sensitive hash are proposed. Based on Gabor wavelet transform algorithm, the feature extraction and calculation method of fusion medical image based on image texture feature fusion. Experiments show that the image similarity matching algorithm based on this feature representation method can achieve an ideal accuracy of .2. aiming at the need of precise research on specific lesions in doctors' clinical diagnosis, teaching and scientific research. In this paper, a candidate set sorting algorithm based on region segmentation is proposed. The candidate set can be segmented by using the seed points specified by the doctor. The region feature of the lesion is extracted and the candidate set is reordered according to the feature. The precise retrieval and sorting of the suspected lesions are realized. A joint index structure is proposed to cluster the three features of the fusion medical features independently. The clustering centers of different layers are combined in three groups, and multiple index nodes are obtained. This kind of index structure only needs to calculate a few clustering centers to get multiple index nodes, which reduces the computation amount of index construction. The weight calculation is introduced in the similarity calculation between the index node and the image, which improves the retrieval efficiency and ensures the retrieval accuracy. A distributed computing architecture based on this retrieval algorithm is designed. By decomposing each large retrieval task into several small tasks distributed parallel execution on different hardware, the algorithm feature extraction is greatly improved. The running speed of index construction and online query solves the problem of huge computation in the retrieval process.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】