【作者】 卢又燃；

【导师】 冯晓源；

【作者基本信息】 复旦大学 ， 医学信息学， 2009， 博士

【摘要】 目的：(1)建立放射科随访数据库,数据库中随访记录的诊断报告予以标准化,使其满足科研及CAD数据挖掘的要求；(2)比较粗糙集、分类回归树及二元Logistic回归法等三种方法依据常规MRI影像特点提取的诊断规则辅助分级诊断胶质瘤的性能,探讨粗糙集方法辅助影像诊断的价值。材料和方法：(1)利用sql server 2000数据库分别建立ACR编码库及ICD-10编码库对数据库中随访病例的诊断进行编码,实现诊断的标准化。ACR编码库包括解剖表和病理表,表中对每个解剖部位及病理诊断使用数字进行编码,其中解剖编码最长为4位,病理最长为6位；ICD-10编码库包括分类表和诊断编码表,每个表中分类或者诊断用字母+数字进行编码。(2)数据库中建立中枢神经系统肿瘤影像诊断模板,包括病变部位、病灶范围、大小、形态、边缘、水肿、占位效应、出血、钙化、坏死、脑积水、CT密度、MRI信号(包括T1WI、T2WI及Flair)及增强特点(包括常规增强、动脉期强化、静脉期强化、延迟期强化)等共19项中枢神经系统肿瘤常规CT/MRI的影像特征,并建立每项影像特征的标准化选项,用于中枢神经系统肿瘤诊断报告的标准化；(3)利用sql server 2000建立放射科随访数据库数据表存储随访记录,包括随访病例的临床资料(姓名、性别、年龄、病史、手术记录等)、病理资料(病理诊断、病理报告、病理图像)、影像资料(影像检查项目、影像报告、报告医生、核片医生、诊断符合、影像图片及影像相关特征等资料),其中文本部分的记录储存在数据表中,病理及影像图像储存在本地硬盘,通过SQL server 2000建立图像索引与数据记录相关联；(4)使用Visual C++开发数据库客户端程序,实现人机对话式的随访信息维护、记录查询、病种统计、诊断符合率统计、数据导入／导出、ACR编码及ICD-10编码修改及用户权限设置等操作功能；(5)275例确诊胶质瘤病例(低级别胶质瘤151例,高级别胶质瘤124例),术前常规MRI平扫及增强检查,提取包括病灶数目、形态、边缘、水肿、坏死、占位效应、钙化、出血、T1WI、T2WI及增强特点等MRI征象,粗糙集基于Rosetta软件使用遗传算法进行属性约简并产生诊断规则,分类回归树使用CRT算法建立预测胶质瘤分级诊断模型,回归法使用二元Logistic回归法建立回归方程。结果：(1)随访数据库实现设计的各项功能,运行稳定；并且诊断标准化的使用提高了数据检索的精确性；(2)粗糙集、决策树及回归法的诊断准确性分别为84.4%、83.3%、83.6%：敏感度分别为75%、74.2%、79.8%；特异度分别为92.1%、91.3%、86.8%,三种方法的ROC曲线下面积分别为0.92、0.907和0.902, ROC曲线下面积之间无明显差异性。相比其他两种方法,粗糙集可以得到更多确定性诊断规则。结论：(1)放射科随访数据库能够实现随访记录诊断报告的标准化管理,其标准化数据格式可以满足临床科研和数据挖掘的要求。(2)粗糙集具有与其他两种方法一样的诊断性能,但是粗糙集的诊断规则更加明确及清晰,具有更好的临床应用价值。更多还原

【Abstract】 Objective:(1) Considering the radiological research and CAD, the records in follow-up database of radiology with standard radiological diagnosis and reports was established. (2) To study the value of Rough set principle in assistant radiological diagnosis by compared Rough set with Decision tree and two-binary-Logistic-Regression in the diagnosis of glioma grade based on routine MR examination.Methods:(1)The ACR code library and ICD-10 library were used to standardized the diagnosis in radiological follow-up database which be build with SQL sever 2000.The ACR code library included anatomical table and pathological table in which the anatomy and pathology was encoded by number. The anatomical codes had 4 ranks and the pathological codes had 6 ranks at most.ICD-10 code library included class table and diagnosis table in which the classes and diagnosis were encoded with alphabet and number.(2)The diagnosis mould of CNS tumor was used to normalize the reports of CNS tumor which included the radiological features in CT and MRI like location, spread, size, shape, margin, edema, mass effect, hemorrhage, necrosis, calcification, hydrocephalus, CT density, MRI intensity (include T1WI, T2WI,Flair),enhancement style (Routine enhancement, enhancement in arterial phase, enhancement in venous phase, enhancement in delay phase). (3) The data tables was build in SQL server 2000 used to store the clinical data (include name, gender, age, history, operation record) and pathological data (pathological diagnosis and picture) and radiological data (include radiological examination method, primitive report, reporter, superior, diagnose accordance, radiological picture, radiological features) of follow-up records among which the pathological pictures and radiological picture were restored in local hardware and connected to the record by table of index.(4)The client program of man-computer interface was developed by using Visual C++ which used to carry out the functions to the database like follow-up record maintainm, inquiry records, statistics of the diagnose accordance rate and entity in database, records import or export, ACR and ICD-10 codes revision, users authoritative. (5)MR images of 275 patients with gliomas (151 low-grade gliomas,124 high-grade gliomas) examined before surgery were collected. The features of MRI with gliomas included numbers, shape, margin, edema, necrosis, mass effect, calcification, hemorrhage, intensity of T1WI and T2WI, enhancement style after administration of contrast agent. The attributes of glioma was reduced by genetic algorithm with software Rosetta and then the diagnostic rules about the grade diagnosis came from reduced attributes. The CR&T algorithm was used to build grade diagnostic model in decision tree. On the other hand the two-binary-Logistic-Regression constructed an equation to predict the grade of glioma.Result:(1)The follow-up database implement the functions of data entry, data inquiry, statistic analysis, system management and data input/output. And the use of encode made the inquiry of data in the database more accurate. (2) The accuracy, sensitivity, Specificity and the averaged area under the ROC curve for Rough set, Decision tree and two binary Logistic Regression were 84.4%,75%,92.1%,0.92; 83.3%,74.2%,91.3%,0.907 and 83.6%,79.8%,86.8%,0.902 respectively. There were no significant differences among the AUC of ROCs of the three methods. Among all features, mass effect, edema and enhancement style were three important features in differentiate grade of glioma.Conclusion:(1) The follow-up database made the management of information of follow-up more standard and its data can be used for clinical research and data mining more effectively. (2) All of three methods had good diagnostic performance. The importance of every feature in diagnostic model can get from decision tree and two-binary-Logistic-Regression. For the explicit and valuable and understandable rules, Rough set was more suitable for practice in diagnosis by experts. The Rough set should be study further to improve the accurate and cover rate of rules.更多还原