Please wait a minute...
浙江大学学报(工学版)  2019, Vol. 53 Issue (6): 1164-1170    DOI: 10.3785/j.issn.1008-973X.2019.06.016
计算机与自动化技术     
基于DenseNet的低分辨CT影像肺腺癌组织学亚型分类
杨婧1,2(),耿辰2,王海林3,纪建松3,戴亚康2,*()
1. 中国科学技术大学,安徽 合肥 230026
2. 中国科学院 苏州生物医学工程技术研究所,江苏 苏州 215163
3. 丽水市中心医院,浙江 丽水 323000
Classification on histological subtypes of lung adenocarcinoma from low-resolution CT images based on DenseNet
Jing YANG1,2(),Chen GENG2,Hai-lin WANG3,Jian-song JI3,Ya-kang DAI2,*()
1. University of Science and Technology of China, Hefei 230026, China
2. Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, China
3. The Central Hospital of Lishui City, Lishui 323000, China
 全文: PDF(872 KB)   HTML
摘要:

为了实现在低剂量、低分辨率CT扫描影像中对肺腺癌组织学亚型的分类鉴别,提出一种基于DenseNet的深度学习方法,从混合性磨玻璃结节(mGGNs)5 mm层厚的低分辨率CT影像中预测IAC和MIA病理分类. 从丽水市中心医院105例患者的105个5 mm层厚低分辨率CT图像中选取样本,划分训练集和测试集后,对训练集进行数据扩展,构建深度学习2D和3D DenseNet模型,分类鉴别IAC和MIA. 2D DenseNet模型的分类准确度为76.67%,敏感性为63.33%,特异性为90.00%,受试者工作特征曲线下的区域面积为0.888 9,显著优于3D DenseNet模型和其他几种深度学习网络模型. 深度学习技术,尤其是2D DenseNet模型,可辅助并指导医生在肺癌CT筛查中对患者的肺腺癌组织学亚型进行预判,特别是在图像分辨率较低的情况下,仍能够快速提供较为准确的诊断.

关键词: 深度学习DenseNet混合性磨玻璃结节肺腺癌厚层CT    
Abstract:

A deep learning method based on DenseNet was proposed to distinguish between IAC and MIA from mixed ground glass nodules low-resolution CT images with 5 mm slice thickness, in order to classify histological subtypes of lung adenocarcinoma from low-dose CT images with low resolution. Samples were obtained from 105 low-resolution CT images with 5 mm slice thickness of 105 patients in the Central Hospital of Lishui City. The data was divided into training set and testing set. Then the training set was augmented; 2D and 3D DenseNet deep learning models were built to distinguish between IAC and MIA. The accuracy, sensitivity, specificity and the area under the receiver operating characteristic curve of the proposed 2D DenseNet method achieved 76.67%, 63.33%, 90.00% and 0.888 9, respectively, which was better than 3D DenseNet and other deep learning models. The deep learning method, especially the 2D DenseNet, may assist doctors in lung cancer screening to predict and guide histological subtypes of patients, which can quickly provide more accurate diagnosis results even under condition of low image resolution.

Key words: deep learning    DenseNet    mixed ground glass nodules    lung adenocarcinoma    thick slice CT
收稿日期: 2019-01-05 出版日期: 2019-05-22
CLC:  R 318.13  
通讯作者: 戴亚康     E-mail: 18366136246@163.com;daiyk@sibet.ac.cn
作者简介: 杨婧(1993—),女,硕士生,从事医学图像处理研究. orcid.org/0000-0003-0073-0735. E-mail: 18366136246@163.com
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章  
杨婧
耿辰
王海林
纪建松
戴亚康

引用本文:

杨婧,耿辰,王海林,纪建松,戴亚康. 基于DenseNet的低分辨CT影像肺腺癌组织学亚型分类[J]. 浙江大学学报(工学版), 2019, 53(6): 1164-1170.

Jing YANG,Chen GENG,Hai-lin WANG,Jian-song JI,Ya-kang DAI. Classification on histological subtypes of lung adenocarcinoma from low-resolution CT images based on DenseNet. Journal of ZheJiang University (Engineering Science), 2019, 53(6): 1164-1170.

链接本文:

http://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2019.06.016        http://www.zjujournals.com/eng/CN/Y2019/V53/I6/1164

属性 各分类值
患者人数 33/72/105(男性/女性/总数)
患者年龄 39/84/62(最小值/最大值/平均值)
肺结节数目
(按病理结果分类)
71/34/105(IAC/MIA/总数)
肺结节数目
(按长轴直径分类)
23/57/20/5(≤1 cm/≤2 cm/≤3 cm/>3 cm)
表 1  患者和肺结节的属性
图 1  确诊为IAC和MIA的混合性磨玻璃结节CT图像示例
图 2  结节长轴直径实际所占像素数分布情况
图 3  DenseNet基础网络模型结构
图 4  2D和3D DenseNet网络模型实验流程
组成成分 2D DenseNet 3D DenseNet
卷积层 7×7 7×7×7
池化层 3×3 3×3×3
紧密块(1) $\left[\begin{array}{*{20}{c}} {1 \times 1\;{\rm{conv}}}\\ {3 \times 3\;{\rm{conv}}} \end{array}\right] \times 6$ $\left[\begin{array}{*{20}{c}} {1 \times 1 \times 1\;{\rm{conv}}}\\ {3 \times 3 \times 3\;{\rm{conv}}} \end{array}\right] \times 6$
过渡层(1) $\left[\begin{array}{*{20}{c}} {1 \times 1\;{\rm{conv}}}\\ {2 \times 2\;{\rm{pool}}} \end{array}\right] $ $\left[\begin{array}{*{20}{c}} {1 \times 1 \times 1\;{\rm{conv}}}\\ {2 \times 2 \times 2\;{\rm{pool}}} \end{array}\right] $
紧密块(2) $\left[\begin{array}{*{20}{c}} {1 \times 1\;{\rm{conv}}}\\ {3 \times 3\;{\rm{conv}}} \end{array}\right] \times 12$ $\left[\begin{array}{*{20}{c}} {1 \times 1 \times 1\;{\rm{conv}}}\\ {3 \times 3 \times 3\;{\rm{conv}}} \end{array}\right] \times 12$
过渡层(2) $\left[\begin{array}{*{20}{c}} {1 \times 1\;{\rm{conv}}}\\ {2 \times 2\;{\rm{pool}}} \end{array}\right]$ $\left[\begin{array}{*{20}{c}} {1 \times 1 \times 1\;{\rm{conv}}}\\ {2 \times 2 \times 2\;{\rm{pool}}} \end{array}\right] $
紧密块(3) $\left[\begin{array}{*{20}{c}} {1 \times 1\;{\rm{conv}}}\\ {3 \times 3\;{\rm{conv}}} \end{array}\right] \times 24$ $\left[\begin{array}{*{20}{c}} {1 \times 1 \times 1\;{\rm{conv}}}\\ {3 \times 3 \times 3\;{\rm{conv}}} \end{array}\right] \times 24$
全局池化层 3×3 3×3×3
表 2  2D和3D DenseNet模型结构细节
图 5  2D和3D DenseNet网络模型的性能对比
深度学习网络 a/% e/% p/% AUC
注:括号内是与2D DenseNet相应性能指标相比下降的百分比
2D DenseNet 76.67 63.33 90.00 0.888 9
2D DenseNet-无颈瓶层 54.17(↓22.5) 20.84(↓42.49) 87.50(↓2.50) 0.592 6(↓0.296 3)
2D DenseNet-无弃权 58.33(↓18.34) 33.33(↓30.00) 83.33(↓6.67) 0.683 3(↓0.205 6)
2D DenseNet-无数据扩展 58.33(↓18.34) 63.33(↓0.00) 53.34(↓36.66) 0.694 4(↓0.194 5)
表 3  2D DenseNet网络模型不同参数的性能对比
深度学习网络 a/% e/% p/% AUC
LeNet 60.00 66.67 53.33 0.750 0
AlexNet 53.33 56.67 50.00 0.722 2
AgileCNN 53.33 63.33 43.33 0.638 9
Multi-channel CNN 56.66 63.33 50.00 0.777 8
2D DenseNet 76.67 63.33 90.00 0.888 9
表 4  2D DenseNet与其他深度学习网络模型的性能对比
1 SIEGEL R L, MILLER K D, JEMAL A Cancer statistics, 2018[J]. CA: a Cancer Journal for Clinicians, 2018, 68 (1): 7- 30
doi: 10.3322/caac.21442
2 YANG J, WANG H, GENG C, et al Advances in intelligent diagnosis methods for pulmonary ground-glass opacity nodules[J]. Biomedical Engineering Online, 2018, 17 (1): 20
doi: 10.1186/s12938-018-0435-2
3 HENSCHKE C I, YANKELEVITZ D F, MIRTCHEVA R, et al CT screening for lung cancer: frequency and significance of part-solid and nonsolid nodules[J]. AJR American Journal of Roentgenology, 2002, 178 (5): 1053- 1057
doi: 10.2214/ajr.178.5.1781053
4 COHEN J G, REYMOND E, MEDICI M, et al CT-texture analysis of subsolid nodules for differentiating invasive from in-situ and minimally invasive lung adenocarcinoma subtypes[J]. Diagnostic and Interventional Imaging, 2018, 99 (5): 291- 299
doi: 10.1016/j.diii.2017.12.013
5 TRAVIS W D, BRAMBILLA E, NOGUCHI M, et al International association for the study of lung cancer/American thoracic society/European respiratory society international multidisciplinary classification of lung adenocarcinoma[J]. Journal of Thoracic Oncology Official Publication of the International Association for the Study of Lung Cancer, 2011, 6 (2): 244- 285
doi: 10.1097/JTO.0b013e318206a221
6 TRAVIS W D, BRAMBILLA E, NICHOLSON A G, et al The 2015 world health organization classification of lung tumors: impact of genetic, clinical and radiologic advances since the 2004 classification[J]. Journal of Thoracic Oncology Official Publication of the International Association for the Study of Lung Cancer, 2015, 10 (9): 1243- 1260
doi: 10.1097/JTO.0000000000000630
7 NAIDICH D P, BANKIER A A, MACMAHON H, et al Recommendations for the management of subsolid pulmonary nodules detected at CT: a statement from the fleischner society[J]. Radiology, 2013, 266 (1): 304- 317
doi: 10.1148/radiol.12120628
8 YUE X, LIU S, LIU S, et al HRCT morphological characteristics distinguishing minimally invasive pulmonary adenocarcinoma from invasive pulmonary adenocarcinoma appearing as subsolid nodules with a diameter of </=3 cm[J]. Clinical Radiology, 2018, 73 (4): 411. e7- 411. e15
doi: 10.1016/j.crad.2017.11.014
9 VAN SCHIL P E, ASAMURA H, RUSCH V W, et al Surgical implications of the new IASLC/ATS/ERS adenocarcinoma classification[J]. European Respiratory Journal, 2012, 39 (2): 478- 486
doi: 10.1183/09031936.00027511
10 LIU S, WANG R, ZHANG Y, et al Precise diagnosis of intraoperative frozen section is an effective method to guide resection strategy for peripheral small-sized lung adenocarcinoma[J]. Journal of Clinical Oncology, 2016, 34 (4): 307- 313
doi: 10.1200/JCO.2015.63.4907
11 ZHANG J, WU J, TAN Q, et al Why do pathological stage IA lung adenocarcinomas vary from prognosis?: a clinicopathologic study of 176 patients with pathological stage IA lung adenocarcinoma based on the IASLC/ATS/ERS classification[J]. Journal of Thoracic Oncology Official Publication of the International Association for the Study of Lung Cancer, 2013, 8 (9): 1196- 1202
doi: 10.1097/JTO.0b013e31829f09a7
12 涂文婷, 范丽, 顾亚峰, 等 计算机辅助定量分析对磨玻璃密度型肺腺癌浸润性的诊断价值[J]. 临床放射学杂志, 2018, 37 (3): 497- 502
TU Wen-ting, FAN Li, GU Ya-feng, et al The value of computer-aided quantitative analysis in the diagnosis of invasiveness of lung adenocarcinoma manifesting as ground glass nodule[J]. Journal of Clinical Radiology, 2018, 37 (3): 497- 502
13 SON J Y, LEE H Y, LEE K S, et al Quantitative CT analysis of pulmonary ground-glass opacity nodules for the distinction of invasive adenocarcinoma from pre-invasive or minimally invasive adenocarcinoma[J]. PLoS One, 2014, 9 (8): e104066
doi: 10.1371/journal.pone.0104066
14 左玉强, 冯平勇, 孟庆春, 等 肺纯磨玻璃结节微浸润腺癌与浸润性腺癌的CT鉴别诊断[J]. 临床放射学杂志, 2017, 36 (4): 495- 498
ZUO Yu-qiang, FENG Ping-yong, MENG Qing-chun, et al CT differential diagnoses of pulmonary minimally invasive adenocarcinoma and invasive adenocarcinoma presenting as pure ground glass nodule[J]. Journal of Clinical Radiology, 2017, 36 (4): 495- 498
15 WANG H, ZHAO T, LI L C, et al A hybrid CNN feature model for pulmonary nodule malignancy risk differentiation[J]. Journal of X-ray Science and Technology, 2018, 26 (2): 171- 187
doi: 10.3233/XST-17302
16 ZHAO X, LIU L, QI S, et al Agile convolutional neural network for pulmonary nodule classification using CT images[J]. International Journal of Computer Assisted Radiology and Surgery, 2018, 13 (4): 585- 595
doi: 10.1007/s11548-017-1696-0
17 HUANG G, LIU Z, MAATEN L V D, et al. Densely connected convolutional networks [C] // 2017 IEEE Conference on Computer Vision and Pattern Recognition. Hawaii: IEEE, 2017: 2261–2269.
18 SRIVASTAVA N, HINTON G, KRIZHEVSKY A, et al Dropout: a simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014, 15 (1): 1929- 1958
19 BAHRAMPOUR S, RAMAKRISHNAN N, SCHOTT L, et al Comparative study of deep learning software frameworks[J]. Computer Science, 2016,
20 LING C X, HUANG J, ZHANG H AUC: a better measure than accuracy in comparing learning algorithms[J]. Lecture Notes in Computer Science, 2003, 329- 341
21 LECUN Y, BOTTOU L, BENGIO Y, et al Gradient-based learning applied to document recognition[J]. P Ieee, 1998, 86 (11): 2278- 2324
doi: 10.1109/5.726791
22 KRIZHEVSKY A, SUTSKEVER I, HINTON G E ImageNet classification with deep convolutional neural networks[J]. Commun ACM, 2017, 60 (6): 84- 90
doi: 10.1145/3098997
23 HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition [C] // 2016 IEEE Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2016: 770–778.
[1] 许佳辉,王敬昌,陈岭,吴勇. 基于图神经网络的地表水水质预测模型[J]. 浙江大学学报(工学版), 2021, 55(4): 601-607.
[2] 王虹力,郭斌,刘思聪,刘佳琪,仵允港,於志文. 边端融合的终端情境自适应深度感知模型[J]. 浙江大学学报(工学版), 2021, 55(4): 626-638.
[3] 张腾,蒋鑫龙,陈益强,陈前,米涛免,陈彪. 基于腕部姿态的帕金森病用药后开-关期检测[J]. 浙江大学学报(工学版), 2021, 55(4): 639-647.
[4] 徐利锋,黄海帆,丁维龙,范玉雷. 基于改进DenseNet的水果小目标检测[J]. 浙江大学学报(工学版), 2021, 55(2): 377-385.
[5] 许豪灿,李基拓,陆国栋. 由LeNet-5从单张着装图像重建三维人体[J]. 浙江大学学报(工学版), 2021, 55(1): 153-161.
[6] 黄毅鹏,胡冀苏,钱旭升,周志勇,赵文露,马麒,沈钧康,戴亚康. SE-Mask-RCNN:多参数MRI前列腺癌分割方法[J]. 浙江大学学报(工学版), 2021, 55(1): 203-212.
[7] 陈巧红,陈翊,李文书,贾宇波. 多尺度SE-Xception服装图像分类[J]. 浙江大学学报(工学版), 2020, 54(9): 1727-1735.
[8] 郑浦,白宏阳,李伟,郭宏伟. 复杂背景下的小目标检测算法[J]. 浙江大学学报(工学版), 2020, 54(9): 1777-1784.
[9] 周登文,田金月,马路遥,孙秀秀. 基于多级特征并联的轻量级图像语义分割[J]. 浙江大学学报(工学版), 2020, 54(8): 1516-1524.
[10] 明涛,王丹,郭继昌,李锵. 基于多尺度通道重校准的乳腺癌病理图像分类[J]. 浙江大学学报(工学版), 2020, 54(7): 1289-1297.
[11] 闫旭,范晓亮,郑传潘,臧彧,王程,程明,陈龙彪. 基于图卷积神经网络的城市交通态势预测算法[J]. 浙江大学学报(工学版), 2020, 54(6): 1147-1155.
[12] 汪周飞,袁伟娜. 基于深度学习的多载波系统信道估计与检测[J]. 浙江大学学报(工学版), 2020, 54(4): 732-738.
[13] 杨冰,莫文博,姚金良. 融合局部特征与深度学习的三维掌纹识别[J]. 浙江大学学报(工学版), 2020, 54(3): 540-545.
[14] 洪炎佳,孟铁豹,黎浩江,刘立志,李立,徐硕瑀,郭圣文. 多模态多维信息融合的鼻咽癌MR图像肿瘤深度分割方法[J]. 浙江大学学报(工学版), 2020, 54(3): 566-573.
[15] 贾子钰,林友芳,张宏钧,王晶. 基于深度卷积神经网络的睡眠分期模型[J]. 浙江大学学报(工学版), 2020, 54(10): 1899-1905.