| 尹柯,张久权,伍建林,巴文娟,林琳,沈晶,熊婧彤,张伟杰.对比卷积神经网络分类模型与放射科医师鉴别浸润性肺腺癌的效能[J].中国医学影像技术,2021,37(9):1338~1342 |
| 对比卷积神经网络分类模型与放射科医师鉴别浸润性肺腺癌的效能 |
| Comparison on convolutional neural network classification model and radiologists in differentiating invasive lung adenocarcinoma |
| 投稿时间：2020-12-21 修订日期：2021-06-20 |
| DOI：10.13929/j.issn.1003-3289.2021.09.015 |
| 中文关键词: 肺肿瘤 肺腺癌 肿瘤侵袭性 人工智能 |
| 英文关键词:lung neoplasms adenocarcinoma of lung neoplasm invasiveness artificial intelligence |
| 基金项目:重庆市科卫联合医学科研项目（2019ZDXM007）。 |
|
| 摘要点击次数: 1463 |
| 全文下载次数: 358 |
| 中文摘要: |
| 目的 比较基于胸部CT建立的卷积神经网络（CNN）分类模型与放射科主治医师分类磨玻璃结节（GGN）中的浸润前病变（PIL）与浸润性腺癌（IA）的效能。方法 回顾性分析1 086例经病理确诊PIL或IA患者的胸部CT，共1 214枚GGN，按9 ∶ 1比例将其分为训练组（n=1 092）及验证组（n=122）。对训练组数据进行各向同性预处理和3D图像块随机采样和填充，建立CNN分类模型，将训练组GGN分类为PIL或IA，并于验证组进行验证。由2名放射科主治医师（医师1和2）对验证组GGN进行分类，鉴别PIL与IA。采用受试者工作特征（ROC）曲线比较CNN分类模型与2名放射科主治医师对验证组进行分类的效能。结果 ROC曲线显示，CNN分类模型、医师1及2鉴别PIL与IA的曲线下面积（AUC）分别为0.866、0.742及0.769；CNN分类模型诊断敏感度（84.81%）显著高于医师1（67.09%，χ2=11.352，P<0.001）、2（74.68%，χ2=18.473，P<0.001），而特异度与医师1、2差异均无统计学意义（P均>0.05）。结论 CNN分类模型鉴别GGN中的PIL与IA的效能优于放射科医师。 |
| 英文摘要: |
| Objective To compare the diagnostic efficiency of convolutional neural network (CNN) classification model based on chest CT and attending physicians of radiology in distinguishing pre-invasive lesions (PIL) and invasive adenocarcinoma (IA) manifested as lung ground glass nodule (GGN). Methods Chest CT of 1 086 patients with 1 214 pathologically confirmed PIL or IA were retrospectively analyzed. GGN were divided into training group (n=1 092) and verification group (n=122) at the ratio of 9 ∶ 1. Isotropic preprocessing of the training data and random sampling and filling of 3D image blocks were performed, so as to establish CNN classification model to identify PIL and IA in GGN and verify using verification group. The verification group was classified by 2 radiologists (physician 1 and 2) to identify PIL or IA manifested as lung GGN. Receiver operating characteristic (ROC) curve was used to compare the efficacy of CNN classification model and 2 radiologists in classifying the verification group. Results ROC curve showed that the area under the curve (AUC) of CNN classification model and physician 1 and 2 was 0.866, 0.742 and 0.769, respectively. The diagnostic sensitivity of CNN classification model was 84.81%, significantly higher than that of physician 1 (67.09%, χ2=11.352, P<0.001)and physician 2 (74.68%, χ2=18.473, P<0.001), while the specificity was not statistically different between physician 1 and 2 (both P>0.05). Conclusion The efficacy of CNN classification model for classifying PIL and IA manifested as lung GGN was higher than that of attending physicians of radiology. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
