CT纹理分析鉴别诊断胰腺导管腺癌、胰腺神经内分泌肿瘤及实性假乳头状肿瘤

目的 探讨CT纹理特征诊断及鉴别诊断胰腺导管腺癌（PDAC）、胰腺神经内分泌肿瘤（PNET）及实性假乳头状瘤（SPTP）的可行性。方法 回顾性分析经病理证实的98例PDAC、62例SPTP及39例PNET患者的CT资料，于肿瘤横断面最大层面沿肿瘤边界手动勾画ROI，提取46个CT纹理特征。按二分类（PDAC vs rest；SPTP vs rest；PNET vs rest）和三分类（PDAC vs SPTP vs PNET）分组方式将数据分组。以单因素回归分析每个纹理特征鉴别二分类各组的诊断效能，并计算AUC；基于随机森林算法选择特征后，采用6种机器学习分类器（LDA、K-NN、RF、Adabost、NB、NN）对二分类和三分类分组进行分类，以多因素回归分析分类器的诊断效能，基于十折交叉验证标准计算AUC。结果 采用单个纹理特征鉴别胰腺肿瘤时，低密度短域补偿和灰度不均匀性分别对PDAC vs rest和SPTP vs rest有较好鉴别能力（AUC=0.73、0.79，P<0.01），而总和均值对PNET vs rest具有极好鉴别能力（AUC=0.90，P<0.01）。分类器鉴别PDAC vs rest、SPTP vs rest、PNET vs rest的诊断效能很好或极好，最大AUC分别为0.88（RF）、0.86（RF）和0.94（Adaboost）。分类器鉴别三分类分组的准确率均较好，以RF最高（0.80）。结论 CT纹理分析可鉴别PDAC、SPTP和PNET；采用机器学习算法可进一步提高鉴别诊断效能。

CT Texture analysis in differential diagnosis of pancreatic ductal adenocarcinoma, pancreatic neuroendocrine tumor and solid pseudopapillary tumor of pancreas

Objective To explore the feasibility of CT texture analysis in diagnosis and differential diagnosis of pancreatic ductal adenocarcinoma (PDAC), pancreatic neuroendocrine tumors (PNET) and solid pseudopapillary tumor of pancreas (SPTP). Methods CT data of 98 patients with PDAC, 62 patients with SPTP and 39 patients with PNET proved by pathologically were retrospectively analyzed. ROI was manually delineated along tumor boundary at the largest level of tumor cross-section, and 46 texture features were extracted. All data were categorized according to dichotomies (PDAC vs rest; SPTP vs rest; PNET vs rest) and tri classification (PDAC vs SPTP vs PNET) methods. The single factor regression was used to analyze the diagnostic efficiency of each texture feature in differentiating dichotomy groups, and the AUC was calculated. After feature selection based on random forest algorithm, 6 machine learning classificators (LDA, K-NN, RF, Adabost, NB, NN) were used to classify dichotomous and triadic groups. The diagnostic efficiency of the classifier was analyzed using multi-factor regression analysis, and the AUC was calculated based on ten-fold cross validation. Results For single texture feature identifying pancreatic tumors, low intensity small area emphasis and grey level nonuniformity were good for identifying PDAC vs rest and SPTP vs rest, respectively (AUC=0.73, 0.79, both P<0.01), while sum average was excellent for differentiating PNET vs rest (AUC=0.90, P<0.01). The diagnostic efficiency of classificator identifying PDAC vs rest, SPTP vs rest and PNET vs rest were very good or excellent, and the maximum AUC was 0.88 (RF), 0.86 (RF) and 0.94 (Adaboost), respectively. The classification accuracy of all classifiers for classifying PDAC vs SPTP vs PNET was good, and that of RF was the highest (0.80). Conclusion CT texture analysis can be used to differentiate PDAC, SPTP and PNET. Machine learning algorithm can further improve the performance.