【作者】 王淼淼；

【导师】 伍建林；

【作者基本信息】 大连医科大学 ， 影像医学与核医学， 2003， 硕士

【摘要】 目的：分析孤立性肺结节(solitary pulmonary nodule，SPN)HRCT上的各种征象，采用最大似然法分析各种征象在周围型肺癌、肺错构瘤及肺结核球三种SPN中的诊断及鉴别诊断价值；并对常规阅片法、最大似然法、人工神经网络(artificial neural network，ANN)系统的诊断结果进行比较，分析其诊断效能。 材料与方法：收集经临床和手术病理证实的SPN(直径≤3cm)150例(包括周围型肺癌、肺错构瘤、肺结核球各50例)，全部病例行常规CT扫描，病灶行HRCT扫描(层厚1.25～3mm，采用骨重建算法)。选取10种SPN在HRCT上的征象(钙化、脂肪、空泡征、空洞、细支气管气象、分叶征、毛刺征、血管集束征、胸膜凹陷征及卫星灶)进行分类统计，运用最大似然法，将结节各种征象出现的概率转化为记分值，然后根据SPN的特点计算累计判别分数值，依据数值大小来判定结节所属类型。 在三类SPN中随机选择各25例作为训练数据，与MATLAB6.1中的神经网络工具箱共同搭建一个初级的人工神经网络(artificial neural network，ANN)测试系统。再将三类SPN中未做训练的余下病例作为测试数据，利用已训练好的ANN对其进行分析诊断。 结果：最大似然法对周围型肺癌、肺错构瘤、肺结核球的诊断正确率分别为86％、92％、90％，平均诊断正确率为89.3％，高于常规阅片法的82％，但统计学上无显著差异(P＞0.05)。最能提示为周围型肺癌的征象依次为空泡征、分叶征、细支气管气象、血管集束征，显示百分比分别为95％、70.3％、66.7％、55.7％。最能提示为肺错构瘤的征象为脂肪成份、钙化，显示百分比分别为100％、31.2％。最能提示为肺结核球的征象依次为空洞、卫星灶、钙化成份、胸膜凹陷征，显示百分比分别为100％、91.7％、55.7％、52.1％。ANN对三类SPN(各25例)的诊断正确率分别为80％、80％、84％，平均诊断正确率为81.3％。三种诊断方法(常规阅片法、最大似然法及ANN)采用两两配对x~2检验，P值无显著性差异(P＞0.05)。 结论：最大似然法是分析SPN很有价值的数理诊断方法。与常规阅片法比较，最大似然法对三类常见SPN的判别正确率均有提高，可用于指导日常阅片，提高诊断正确率。ANN.是计算机辅助诊断(eomputer aided diagnosis，e幼)中人工智能的最前沿领域，它的数字化概率诊断结果将为临床影像学诊断的智能化开辟广阔的前景。更多还原

【Abstract】 Purpose: To analyze various features of solitary pulmonary nodule (SPN) on HRCT, and with the application of the maximum likehood method, study the diagnosis of 3 kinds of SPNs (namely the peripheral lung cancer, hamartoma, and tuberculoma) using these features, and hence the usefulness of them in differential diagnosis. Comparisons are made on the diagnostic results from the using of the traditional method, the maximum likehood method, and the artificial neural network (ANN) system to investigate the diagnostic performance of these methods.Materials and methods: From clinically and pathologically proved cases, collect 150 cases of SPNs (including for each of the 3 kinds of SPNs 50 cases, and with their diameters all not more than 3cm). All are treated with ordinary CT, with the focus being scanned using HRCT (with the scan thickness of 1.25-3mm, and the bone construction). For a systematic statistic analysis, 10 features of SPN on HRCT are selected, including namely calcification, fat, vacuole sign, cavication, air bronchograms, lobulation, spiculation, vascular convergence, pleural indentation, and satellite. In the application of the maximum likehood method, first transform the occurrence probabilities for all the features of a nodule into scores. Then counting the judging scores according to the nodule characteristics of the patient, and finally determine the type of the nodule based on this score value.With all of the 3 SPNs, randomly chose 25 cases from each to collectively form the training data set for the ANN testing system established with the ANN Toolbox in MATLAB6.1. Taking the rest of the data (which has not been used as training data) as testing data, conduct the diagnosis analysis using the ANN already trained.Results: The correct rate of the maximum likehood method for the diagnosis of peripheral lung cancer, hamartoma and tuberculoma is 86%, 92%, and 90% respectively, with theaverage accurate diagnosis rate being 89.3% which is higher than the traditional method being 82%, but there is no significant difference in statistical study (P>0.05). The most suggestive features for peripheral lung cancer are vacuole sign, lobulation, air bronchograms, and vascular convergence, of which the revealing percentage is 95%, 70.3%, 66.7% and 55.7% respectively; while the dominant features for hamartoma are fat composition, and calcification, with their revealing percentage being 100% and 31.2% respectively; and those for tuberculoma are cavication, satellite, calcification composition, and pleural indentation, with the revealing percentage being 100%, 91.7%, 55.7% and 52.1% respectively. The correct rate of ANN for the diagnosis of the 3 kinds SPNs (each having 25 cases) is 80%, 82% and 84% respectively, with an average accurate diagnosis rate being 82%. There is no significant difference in statistical study (P>0.05).Conclusions: The maximum likehood method is an useful instrument for the statistical diagnosis of SPNs. Comparing with the traditional method, it results in higher accurate rate for all the 3 kinds of SPNs usually seen, and can be made use of in the guiding of daily method. ANN is a forefront in the application of AI in computer aided diagnosis (CAD). By using digitalized statistical diagnosis, it provides the intelligentized foreground of radiology.更多还原