«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《天津医科大学学报》[ISSN:1006-8147/CN:12-1259/R]

卷:

28卷

期数:

2022年03期

页码:

295-300

栏目:

临床医学

出版日期:

2022-05-20

文章信息/Info

Title:

Efficacy and stability of 3D ResNet for predicting nodule subtypes in lung adenocarcinoma

文章编号:

1006-8147(2022)03-0295-06

作者:

骆源¹; 徐启飞²; 吕泽政¹; 蔡娜¹; 郭丽¹

（1.天津医科大学医学技术学院医学图像处理教研室，天津300203；2.山东省临沂市人民医院影像科，临沂 276000）

Author(s):

LUO Yuan¹; XU Qi-fei²; LYU Ze-zheng¹; CAI Na¹; GUO Li¹

（1．Department of Medical Image Processing，School of Medical Technology，Tianjin Medical University，Tianjin 300203，China； 2 .Department of Imaging，Linyi People′s Hospital，Linyi 276000，China）

关键词:

人工智能; 深度学习; 卷积神经网络; 肺腺癌; 诊断

Keywords:

artificial intelligence; deep learning; convolutional neural network; lung adenocarcinoma; diagnosis

分类号:

R734.2；R730.44

DOI:

-

文献标志码:

A

摘要:

目的：探究ResNet模型对肺腺癌不同亚型结节的分类表现及稳定性。方法：回顾性收集2014 年2 月—2020 年10 月期 间的364 例肺腺癌结节CT 影像数据，以7∶3 的比例分为训练集和内部测试集，将2020 年4 月到2020 年11 月的58 例结节数 据作为外部测试集。使用基于ResNet的三维卷积神经网络在训练集中进行训练以及调参，并使用内部测试集和外部测试集对 模型的准确性及泛化性进行评估。使用随机中心移动和掩膜处理的方式分别以内部测试集和外部测试集为基础构造新的测试 集，新数据集对模型进行测试验证模型的稳定性。结果：模型在内部测试集AUC 为0.949 1（95%CI：0.910 8～0.987 4），模型在随 机中心移动以及掩膜处理之后的数据集的AUC 值分别为0.940 4 和0.918 1， 与其差异无统计学意义（P 值分别为0.425 3 和 0.239 3）。在外部测试集中模型AUC 为0.959 6（95%CI：0.901 2～1.000 0），在用于稳定性测试的随机中心移动以及掩膜处理之 后的数据集中，模型所得AUC 分别为0.948 5和0.947 3，与其同样差异无统计学意义（均P>0.05）。结论：ResNet 模型对肺腺癌 结节亚型有优异的鉴别能力，并且具有一定稳定性。

Abstract:

Objective：To investigate the classification performance and model stability of ResNet models for different subtypes of nodules in lung adenocarcinoma. Methods: The CT image of 364 lung adenocarcinoma nodules collected retrospectively between February 2014 and October 2020 were divided into a training set and an internal test set in a ratio of 7∶3，and data of 58 nodules from April 2020 to November 2020 were used as the external test set. The ResNet-based 3D convolutional neural network was trained and tuned in the training set， and the accuracy and generalization of themodel was evaluated using both internal and external test sets. To verify the stability of the model， two new test sets were constructed using random center shifts and masking process in both internal and external test set，and the model was tested using the new test set. Results：The model obtained an AUC of 0.949 1（95% CI：0.910 8-0.987 4）on the internal test set，and the AUC values of the model were not statistically different（P=0.425 3 and 0.239 3，respectively）from those measured on the data sets with the random center shift and after the masking process. The model AUC in the external test set was 0.959 6（95% CI ：0.901 2-1.000 0）. In the dataset after the random center shift and mask processing used for stability testing，the AUC obtained for the model（0.948 5 and 0.947 3，respectively）was again not statistically different from it（all P>0.05）. Conclusion:ResNet model has excellent ability to discriminate subtypes of lung adenocarcinoma，and the model has considerable stability.

参考文献/References:

[1] MEZA R，MEERNIK C，JEON J，et al. Lung cancer incidence trends by gender，race and histology in the United States，1973-2010 [J]. PLoS One，2015，10（3）：e0121323.
[2] 李媛，谢惠康，武春燕. WHO 胸部肿瘤分类（第5 版）中肺肿瘤部 分解读[J]. 中国癌症杂志，2021，31（7）：574-580.
[3] YOSHIZAWA A，MOTOI N，RIELY G J，et al. Impact of proposed IASLC/ATS/ERS classification of lung adenocarcinoma：prognostic subgroups and implications for further revision of staging based on analysis of 514 stage I cases [J]. Mod Pathol，2011，24（5）：653-664.
[4] YANAGAWA N，SHIONO S，ABIKO M，et al. New IASLC/ATS/ERS classification and invasive tumor size are predictive of disease recurrence in stage I lung adenocarcinoma[J]. J Thorac Oncol，2013，8（5）： 612-618.
[5] 陶雪敏，方瑞，吴重重，等. 深度学习模型对纯磨玻璃结节肺腺 癌病理亚型的预测分析[J]. 中国医学科学院学报，2020，42（4）： 477-484.
[6] XIA X，GONG J，HAO W，et al. Comparison and fusion of deep learning and radiomics features of ground-glass nodules to predict the invasiveness risk of stage-I lung adenocarcinomas in CT scan [J]. Front Oncol，2020，10：418.
[7] HE K，ZHANG X，REN S，et al. Deep residual learning for image recognition; proceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition，F，2016 [C].
[8] IOFFE S，SZEGEDY C. Batch normalization：accelerating deep network training by reducing internal covariate shift[J]. arXiv preprint arXiv:1502.03167，2015.
[9] PRECHELT L. Early stopping -but when? [M]. Neural Networks： Tricks of the trade. Springer，1998：55-69.
[10] KINGMA D P，BA J. Adam：a method for stochastic optimization [J]. arXiv preprint arXiv:1412.6980，2014.
[11] SMITH L N. Cyclical learning rates for training neural networks; proceedings of the 2017 IEEE Winter Conference on Applications of Computer Vision（WACV），F，2017 [C]. IEEE.
[12] PASZKE A，GROSS S，MASSA F，et al. Pytorch：an imperative style， high-performance deep learning library[J]. Adv Neural Inf Process Syst，2019，32：8026-8037.
[13] DELONG E R，DELONG D M，CLARKE-PEARSON D L. Comparing the areas under two or more correlated receiver operating characteristic curves：a nonparametric approach[J]. Biometrics，1988，44（3）： 837-845.
[14] WANG X，LI Q，CAI J，et al. Predicting the invasiveness of lung adenocarcinomas appearing as ground-glass nodule on CT scan using multi-task learning and deep radiomics[J]. Transl Lung Cancer Res， 2020，9（4）：1397-1406.
[15] WANG S，WANG R，ZHANG S，et al. 3D convolutional neural network for differentiating pre-invasive lesions from invasive adenocarcinomas appearing as ground-glass nodules with diameters≤3 cm using HRCT [J]. Quant Imag Med Surg，2018，8（5）：491-499.

相似文献/References:

[1]孟祥虹,吴迪嘉,马信龙,等.深度卷积神经网络技术自动诊断肋骨骨折的CT应用初探[J].天津医科大学学报,2022,28(02):205.
　MENG Xiang-hong,WU Di-jia,MA Xin-long,et al.Preliminary study on CT application of deep convolutional neural network in automatic diagnosis of rib fractures[J].Journal of Tianjin Medical University,2022,28(03):205.

备注/Memo

备注/Memo:

基金项目国家自然科学基金（8197070539）；天津市自然科学基 金(*18JCYBJC95600*)
作者简介:骆源（1997-），男，硕士在读，研究方向：医学影像技术；
通信 作者：郭丽，E-mail：yxgl@tmu.edu.cn；蔡娜，E-mail：caina302@ aliyun.com。

常用功能

更新日期/Last Update: 2022-06-01