计算机辅助判别超声内镜图像诊断胰腺癌的实验研究

Computer aided endoscopic ultrasonography in diagnosis of pancreatic cancer

目的观察利用数字图像处理技术提取超声内镜图像纹理特征，并运用于胰腺癌诊断的价值。方法随机选择2005年2月-2007年2月间行胰腺EUS检查的216名患者。其中胰腺癌153例，非胰腺癌患者（包括正常胰腺与慢性胰腺炎）63例，所有胰腺癌病例均经EUS-FNA细胞学检查确诊。选择EUS图像并提取纹理特征。根据最优特征组合，通过支撑向量机将病例进行自动分类为胰腺癌和非胰腺癌病例，并计算该诊断方法的敏感性、特异性和准确率。结果根据EUS图像共提取9大类，69个特征用于模式分类特征，其中类间距最大的25个特征被选取作为初始特征。将现有216例病例，随机划分为训练集和测试集，训练集108例（癌症76例，非癌症32例）、测试集108例（癌症77例，非癌症31例），用训练集训练分类器，测试集进行测试。共进行50次随机实验，最终得出胰腺癌分类的准确性为（97．98±1．237）％，敏感性为（94．32±0．0354）％，特异性为（99．45±0．0102）％。结论数字图像处理技术与计算机辅助EUS图像判别法准确率高，无创伤性，为胰腺癌的临床诊断提供了一个新的、有价值的研究方向。

Objective To process the image of endoscopic uhrasonography （EUS） by digital imaging processing （DIP） and pattern recognition, and to evaluate its efficacy in diagnosis of pancreatic adenocarcinoma. Methods Two hundreds and sixteen patients, who underwent EUS between Feb 2005 and Feb 2007, were randomly recruited to the study. The cohort included 153 cases of pancreatic cancer, which were confirmed by cytological findings after fine-needle aspiration, and 63 cases of non-pancreatic cancer （ normal pancreas and chronic panereatitis）. The texture features of the EUS image were selected and extracted, and cases were automatically divided into cancer and non-cancer based on findings of support vector machine （SVM）. Sensitivity, specificity and accuracy of the technique were calculated. Results From each region of interest （ROI）, a total of 69 texture features vest in 9 sets were extracted, and 25 features with most set interval were taken as initial. The images of 216 cases were divided randomly into training set （108 cases, 76 cancer and 32 non cancer） and testing set （ 108 cases, 77 cancer and 31 non cancer）. After 50 times of random tests, the average accuracy, sensitivity and specificity of the diagnosis of pancreatic cancer were （97.98±1.237）%, （94.32±0.0354）%, and （99.45 ±0.0102）% respectively. Conclusion DIP, combined with computer aided EUS imaging, is an accurate and noninvasive technique in diagnosis of pancreatic cancer, which warrants novel and further researches.