【作者】 王坤；

【导师】 李三立；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2009， 博士

【摘要】 乳腺癌普查医学网格是美国和欧洲一种重要的医学网格,利用网格技术把异地、异构的计算、存储和医学设备等资源动态组合为虚拟机构,向网格用户提供动态共享这些资源的平台,同时利用网格的大规模计算资源向用户提供更高级的计算机辅助医学应用。在与上海市第二医科大学、上海市瑞金医院和上海市肿瘤医院等单位的共同跨学科研究中,本文的主要贡献包括:(1)提出了一个契合医学需求的大规模乳腺癌普查医学网格的总体设计,定义了三种符合实用的计算机辅助医学服务工作流,研究了中间件设计等问题;(2)提出了一种自适应的数据并发传输策略ADT,根据用户的数据需求权衡取舍TVQ值,自适应采用四种传输模式,提高数据的传输效率;(3)提出使用BI-RADS分级作为计算机辅助对病例影像学特征进行分类的结果。基于BI-RADS分级结果,提出了一种分布式动态分类算法DSMM,减少分类准确率低点的数目;(4)提出了一种用于复杂病例聚类的V3COCA算法,同时满足普查中计算机辅助诊断和流行病学分析研究两方面的需求;(5)将影像并行处理分为两种模式,分别满足计算机辅助诊断和辅助筛查的需求。在两种模式下,分别提出用于计算最优处理机数目的BNP公式和用于影像分割的BIP策略,提高了并行处理的加速比。在乳腺癌普查医学网格试验床中,本文采用了上海市瑞金医院和上海市肿瘤医院的2,937个病例,以及美国SEER数据库的10,000个病例,对所提出设计和算法进行了大量的实验验证,实验结果表明本文的研究贡献应用于大规模乳腺癌普查医学网格时,具有实用价值和理论意义。更多还原

【Abstract】 Breast cancer screening medical grid is an important grid research both in the United States and in the Europe. With high-speed internet, it integrates the geographically distributed computing nodes, storage nodes, as well as medical instruments, stores the large-scale and heterogeneous data generated by breast cancer screening, and provides the users with transparent access interfaces to these dynamical data and computer aided medical applications.As a multi-disciplinary research, this dissertation makes research on the overall design, middleware, data management and mining, as well as mammograms’processing for large scale breast cancer screening medical grid, which is cooperated by the Shanghai’s second Medical Insitute, Shanghai Ruijing Hospital, and Shanghai Tumor Hospital. The main contribution includes: (1) An overall design which meets the demands of practical medical applications is proposed, where three computer aided medical application procedures are defined. The grid services and middleware design topic is researched as well; (2) An adaptive data transfer strategy ADT is proposed, which dynamically adopts four concurrent transfer modes, according to the users’tradeoff value TVQ on the transfer speed and data quality; (3) The BI-RADS level is used as the classfication result, and a distributed, dynamical classification algorithm DSMM is proposed to reduce the number of low classification accuracy cases; (4) A V3COCA clustering algorithm is proposed for the complicated breast cancer cases’ clustering, which simultaneously satisfies two application modes: computer aided diagonosis and epidemiology analysis. It also has four performance advantages that traditional algorithms can not possess at one time; (5) The processing of mammograms is partitioned into two modes for computer aided diagnosis and screening. For the two modes, a BNP formula for finding a proper number of processors and a BIP image partition strategy are designed respectively to reduce the execution time.Experiments, which are conducted with 2,937 cases from the Shanghai Ruijing Hospital and Shanghai Tumor Hospital, and 10,000 cases from the American’s SEER database, validate the efficacy of the algorithms and methods proposed by this dissertation. The experimental results show that the research of this dissertation is practically useful for large scale breast cancer screening medical grid.更多还原