马圆,王风,韩勇,张凤,梁志刚,黄健,杨志,郭秀花.基于深度信念网络检测PET/CT图像肺结节良恶性[J].中国医学影像技术,2020,36(1):77~80
基于深度信念网络检测PET/CT图像肺结节良恶性
Classification of pulmonary nodules on PET/CT image based on deep belief network
投稿时间:2019-02-19  修订日期:2019-12-24
DOI:10.13929/j.issn.1003-3289.2020.01.021
中文关键词:  肺肿瘤  诊断  人工智能  正电子发射断层显像术
英文关键词:lung neoplasms  diagnosis  artificial intelligence  positron-emission tomography
基金项目:国家自然科学基金项目(81773542)、国家青年科学基金项目(81703318)。
作者单位E-mail
马圆 首都医科大学公共卫生学院流行病与卫生统计学系 北京市临床流行病学重点实验室, 北京 100069  
王风 北京大学肿瘤医院暨北京市肿瘤防治研究所核医学科 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100036  
韩勇 首都医科大学公共卫生学院流行病与卫生统计学系 北京市临床流行病学重点实验室, 北京 100069  
张凤 首都医科大学公共卫生学院流行病与卫生统计学系 北京市临床流行病学重点实验室, 北京 100069  
梁志刚 首都医科大学宣武医院核医学科, 北京 100053  
黄健 爱尔兰科克大学数学学院, 爱尔兰 科克 T12 K8AF  
杨志 北京大学肿瘤医院暨北京市肿瘤防治研究所核医学科 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100036 pekyz@163.com 
郭秀花 首都医科大学公共卫生学院流行病与卫生统计学系 北京市临床流行病学重点实验室, 北京 100069 statguo@ccmu.edu.cn 
摘要点击次数: 190
全文下载次数: 118
中文摘要:
      目的 观察深度信念网络(DBN)方法识别PET/CT图像良恶性肺结节的效果。方法 收集216例肺结节患者的PET/CT图像,共339个肺结节,其中良性190个、恶性149个;共截取2 055张ROI图像,良性1 069张,恶性986张。对ROI图像进行灰度、大小归一化处理后,采用DBN方法进行分类诊断。通过实验方法确定网络结构及训练参数,并以混淆矩阵、总体精度、Kappa系数等指标评价分类结果。提取同一批图像数据非下采样双树复轮廓波变换(NSDTCT)的小波纹理参数,构建支持向量机分类模型(SVM),对比分析其与DBN的检测结果。结果 DBN和SVM方法测试集检测结果分别为总体精度0.94和0.72、灵敏度0.96和0.66、特异度0.92和0.96及Kappa系数0.87和0.42。结论 DBN识别肺结节良恶性的准确性高于SVM方法。
英文摘要:
      Objective To observe classification effect of pulmonary nodules on PET/CT images with deep belief network (DBN). Methods PET/CT images of 216 patients with pulmonary nodules were collected, among them 339 pulmonary nodules were detected, including 190 benign and 149 malignant ones. Totally 2 055 ROI images were captured, incuding 1 069 of benign ones and 986 of malignant ones. Gray scale and size normalization were performed on ROI images, and then the lesions were detected with DBN. The network structure and training parameters were determined by experimental methods, and the results were evaluated by confusion matrix, overall accuracy, Kappa coefficient and other indicators. A support vector machine model (SVM) was also built with wavelet texture features based on nonsubsampled dual-tree complex contourlet transform (NSDTCT), using the same data as DBN. The results detected with DBN and SVM were compared. Results The results of DBN and SVM methods were 0.94 and 0.72 for overall accuracy, 0.96 and 0.66 for sensitivity, 0.92 and 0.96 for specificity, and 0.87 and 0.42 for Kappa coefficient, respectively. Conclusion The accuracy of DBN in identifying benign and malignant pulmonary nodules is better than that of SVM.
查看全文  查看/发表评论  下载PDF阅读器