【作者】 陈海燕；

【导师】 余日胜；

【作者基本信息】 浙江大学 ， 影像医学与核医学（专业学位）， 2020， 硕士

【摘要】 目的使用影像学特征联合增强CT门脉期纹理特征建立诊断模型来鉴别胰腺浆液性囊腺瘤(SCN)和黏液性囊腺瘤(MCN)。材料与方法回顾性分析2010年1月至2019年10月经病理证实的57例胰腺SCN和43例胰腺MCN。记录一般临床资料和影像学资料,如年龄、主诉、病灶大小、部位、中央瘢痕、钙化、囊壁增厚情况等。使用Ma Zda软件提取直方图特征、灰度共生矩阵等共计271个纹理特征。将病灶以7:3随机分成实验组和验证组,对实验组的数据采用最小绝对收缩与选择算子算法(LASSO)和结合分类与回归决策树(CART)算法进行特征筛选,用验证组数据进行验证。将影像特征和筛选出的纹理特征进行单因素和多因素Logistic回归分析。将影像学特征结合纹理特征的Logistic回归模型,与只有影像学特征和只有纹理特征所建立的Logistic回归模型三者进行比较。通过绘制接受者操作特征曲线(ROC),使用最大曲线下面积(AUC)量化三个模型的预测效能,并记录各个模型的敏感度和特异度。结果SCN患者平均年龄52.3±13.0岁,75.4%为女性;MCN患者平均年龄45.6±13.9岁,90.7%为女性,差异分别具有统计学意义(P<0.05)。两者在发病部位、直径、钙化分布、囊壁增厚情况上差别具有统计学意义(P<0.001);而在伴随症状、肿瘤指标升高情况上和肿瘤形态中未发现明显统计学差异(P>0.05)。通过LASSO算法中的10折交叉验证方法确定最小λ值=0.0106,筛选出23个特征,去除无关统计学变量最后纳入10个纹理特征到CART决策树模型中,根据最小“xerror”值=0.4667,CP值=0.0667为依据建立决策树模型,最后筛选出一个纹理特征(S(4,4)Sum Entrp)。实验组的准确性80%,AUC为0.804(95%CI:0.692-0.889),敏感度77.5%,特异度83.33%;验证组的准确性63.3%,AUC为0.658(95%CI:0.464-0.821),敏感度47.06%,特异度84.62%。影像结合纹理特征的多因素Logistic回归分析示,囊壁<3 mm(OR:21.52,95%CI:1.21-383.3)和S(4,4)Sum Entrp>1.243(OR:8.08,95%CI:2.78-23.47)是诊断SCN的关键因子(P<0.05);影像特征的多因素Logistic回归分析示,胰腺头颈部(OR:3.96,95%CI:1.25-12.5)、直径≤51.2 mm(OR:2.9,95%CI:1.13-7.46)和囊壁厚度<3 mm(OR:12.61,95%CI:1.03-153.8)是诊断SCN的因子(P<0.05)。影像结合纹理特征得出的Logistic回归模型的AUC为0.779(95%CI:0.685-0.856),敏感度68.42%,特异度83.72%;当只有影像学特征时建立的Logistic回归模型AUC为0.762(95%CI:0.666-0.841),敏感度84.21%,特异度58.14%;当只有纹理特征时建立的Logistic回归模型的AUC为0.758(95%CI:0.662-0.838),敏感度70.18%,特异度81.4%。结论本研究显示纹理特征在一定程度上可以鉴别胰腺SCN和MCN,而影像学特征也有助于鉴别;影像学特征结合纹理特征得出的诊断模型比单纯影像学特征或纹理特征建立的模型更能有效鉴别胰腺SCN和MCN。更多还原

【Abstract】 Objective A diagnostic model was established by combining imaging features with enhanced CT portal phase texture analysis to differentiate between pancreatic serous cystadenoma(SCN)and pancreatic mucinous cystadenoma(MCN).Materials and methods Fifty-seven patients with SCN and 43 patients with MCN confirmed by surgery or biopsy from January 2010 to October 2019 were retrospectively analyzed.Demographic and radiological information,such as age,main complaint,size,location,central scar,calcification,thickening of cyst wall and other situation were collected.Total 271 texture features like histogram features、the gray-level co-occurrence matrix features etc.were extracted through Ma Zda.The lesions were randomly divided into training and validation group with a ratio of 7:3.As for training group,least absolute shrinkage and selection operator and classification and regression trees algorithm were used to extract features.Validation group was used for verifying.Univariate analyses and binary logistic regression analyses including clinical,image features and texture features,wereperformed to identify independent factors and establish a diagnostic model.Comparison between the logistic regression model based on the combination of image features and texture features and the logistic regression model based on the combination of image features or texture features.Receiver operating characteristic curves were used,the prediction efficiency of the three models was quantified by area under the curve(AUC),and the sensitivity and specificity of each model were recorded.Results The average age of SCN patients was 52.3 ± 13.0 years old,and 75.4% of them were female;while the average age of MCN patients was 45.6 ± 13.9 years old,and 90.7% of them were female,the difference between them was statistically significant(P < 0.05).There was significant difference in the location,diameter,distribution of calcification and thickening of the cystic wall between the two groups(P<0.05).However,there was no significant difference in symptoms,tumor marker and tumor morphology(P > 0.05).Twenty-three texture features were extracted by the minimum λ value = 0.0106,which was determined through the 10 fold cross validation method in lasso algorithm.After remove irrelevant statistical variables,finally,10 texture features were included in cart decision tree model.According to the minimum "xerror" value = 0.4667,CP value =0.0667,the decision tree model was established,and finally one texture feature(s(4,4)sumentrp)was selected.The accuracy of the training group was 80%,AUC was 0.804(95% CI: 0.692-0.889),sensitivity was 77.5%,specificity was 83.33%;and the accuracy of the validation group was 63.3%,AUC was 0.658(95% CI: 0.464-0.821),sensitivity was 47.06%,specificity was 84.62%.Multivariate logistic regression analysis of image combined with texture features showed that cyst wall < 3 mm(OR: 21.52,95% CI: 1.21-383.3)and S(4,4)sumentrp >1.243(OR: 8.08,95% CI: 2.78-23.47)were the key factors for diagnosis of SCN(P <0.05).Multivariate logistic regression analysis of image features showed that location ofthe pancreatic head and neck(OR: 3.96,95% CI: 1.25-12.5),diameter ≤ 51.2 mm(OR:2.9,95% CI: 1.13-7.46)and cyst wall thickness < 3 mm(OR: 12.61,95% CI:1.03-153.8)were the diagnostic factors of SCN(P < 0.05).As for model combined imaging features with texture features,the AUC of the logistic regression model was0.779(95% CI: 0.685-0.856),the sensitivity was 68.42%,and the specificity was83.72%.As for model contained imaging features only,the AUC of the logistic regression model was 0.762(95% CI: 0.666-0.841),the sensitivity was 84.21%,and the specificity was 58.14%.As for model included texture feature only,the AUC of the logistic regression model was 0.758(95% CI: 0.662-0.838),the sensitivity was 70.18%,and the specificity was 81.4%.Conclusion This study showed that texture features can distinguish SCN from MCN of pancreas to some extent,and the imaging features can also distinguish them.Furthermore,the combination of imaging features and texture features is more effective in distinguishing SCN from MCN than that of imaging features or texture features alone.更多还原