您的位置:山东大学 -> 科技期刊社 -> 《山东大学学报(工学版)》

山东大学学报 (工学版) ›› 2018, Vol. 48 ›› Issue (5): 77-84.doi: 10.6040/j.issn.1672-3961.0.2018.191

• 控制科学与工程 • 上一篇    下一篇

基于PSO-ConvK卷积神经网络的肺部肿瘤图像识别

梁蒙蒙1(),周涛1,2,*(),夏勇3,张飞飞1,杨健1   

  1. 1. 宁夏医科大学公共卫生与管理学院, 宁夏 银川 750004
    2. 宁夏医科大学理学院, 宁夏 银川 750004
    3. 西北工业大学计算机学院, 陕西 西安 710072
  • 收稿日期:2018-05-31 出版日期:2018-10-01 发布日期:2018-05-31
  • 通讯作者: 周涛 E-mail:1020881411@qq.com;zhoutaonxmu@126.com
  • 作者简介:梁蒙蒙(1992—),女,安徽临泉人,硕士研究生,主要研究方向为图像处理与分析. E-mail:1020881411@qq.com
  • 基金资助:
    国家自然科学基金资助项目(61561040);陕西省教育厅资助项目(2013JK1142)

Lung tumor images recognition based on PSO-ConvK convolutional neural network

Mengmeng LIANG1(),Tao ZHOU1,2,*(),Yong XIA3,Feifei ZHANG1,Jian YANG1   

  1. 1. School of Public Health and Management, Ningxia Medical University, Yinchuan 750004, Ningxia, China
    2. School of Science, Ningxia Medical University, Yinchuan 750004, Ningxia, China
    3. School of Computer Science, Northwestern Polytechnical University, Xi′an 710072, Shaanxi, China
  • Received:2018-05-31 Online:2018-10-01 Published:2018-05-31
  • Contact: Tao ZHOU E-mail:1020881411@qq.com;zhoutaonxmu@126.com
  • Supported by:
    国家自然科学基金资助项目(61561040);陕西省教育厅资助项目(2013JK1142)

摘要:

针对卷积核随机初始化以及梯度下降法训练卷积神经网络易陷入局部最值问题,提出粒子群算法优化卷积核(particle swarm optimization-convolution kernel, PSO-ConvK)的图像识别方法。使用参数迁移法构造卷积神经网络,并提取卷积核,利用PSO不断更新粒子的速度和位置,寻找全局最优值以初始化卷积核,将其传递到卷积神经网络,用肺部肿瘤数据训练卷积神经网络,结合梯度下降法修正网络权重,使得PSO算法的全局优化能力与梯度下降法的局部搜索能力相结合。试验通过批次大小、迭代次数以及网络层数3个角度验证方法的有效性,并与高斯函数优化卷积核进行对比。结果显示, PSO优化卷积核的识别率始终高于随机化卷积核和高斯卷积核,识别率最终达到98.3%,具有一定的可行性和优越性。

关键词: 粒子群算法, 卷积核, 卷积神经网络, 肺部肿瘤, 医学图像

Abstract:

In order to solve problems that convolution kernel was random initialization and the gradient descent method to train convolution neural network was easy to fall into local minimum, an image recognition method based on particle swarm optimization for convolution kernel was proposed. CNN(convolution neural network) was constructed by using the parameter migration method, and convolution kernel was extracted. The particle swarm algorithm was used to update the velocity and position of particles constantly and find the global optimal value to initialize convolution kernels. Convolution kernels were transferred to convolution neural network, and lung tumor images were used to train them. CNN model was trained by lung tumor images, and gradient descent method was used to modify network weights, hence global optimization ability of PSO algorithm was combined with local search ability of gradient descent method. The experiments verified effectiveness of method through three perspectives: batch sizes, iteration numbers, and network layer numbers. The particle swarm algorithm was compared with gauss function. The recognition rates of PSO optimized convolution kernel were always higher than that of randomized convolution kernel and gauss convolution kernel, its recognition rate reached 98.3%, which had certain feasibility and superiority.

Key words: PSO, convolution kernel, convolutional neural network, lung tumor, medical images

中图分类号: 

  • TP183

图1

卷积神经网络结构"

图2

算法框架图"

图3

PSO结合梯度下降法流程图"

图4

预处理后的试验数据"

表1

PSO优化卷积核随迭代次数变化的识别率和灵敏度比较"

迭代次数 随机化卷积核 高斯卷积核 PSO优化一层卷积核 PSO优化两层卷积核
训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/%
30 2 558.22 93.10 97.60 2 545.03 95.00 96.00 2 562.22 94.90 99.00 2 570.00 95.50 97.00
40 3 408.20 93.30 98.40 3 395.23 95.60 96.00 3 411.77 95.60 97.40 3 406.02 95.80 97.40
50 4 265.31 94.00 98.40 4 235.45 96.00 97.40 4 273.44 95.80 97.20 4 257.02 96.50 97.80
60 5 116.59 94.90 97.00 5 094.44 96.20 96.80 5 124.09 95.90 96.40 5 109.05 96.70 96.80
70 5 975.69 95.30 96.00 5 963.13 96.40 97.20 6 001.06 96.50 97.20 5 933.23 97.00 99.00

表2

PSO优化卷积核随迭代次数变化的特异度、MCC和F1Score比较"

迭代次数 随机化卷积核 高斯卷积核 PSO优化一层卷积核 PSO优化两层卷积核
特异度/% MCC F1Score 特异度/% MCC F1Score 特异度/% MCC F1Score 特异度/% MCC F1Score
30 88.60 0.87 0.93 94.00 0.90 0.95 90.80 0.90 0.95 94.00 0.91 0.96
40 88.20 0.87 0.93 95.20 0.91 0.96 93.80 0.91 0.96 94.20 0.92 0.96
50 89.60 0.88 0.94 94.60 0.92 0.96 94.40 0.92 0.96 95.20 0.93 0.97
60 92.80 0.90 0.95 95.60 0.92 0.96 95.40 0.92 0.96 96.60 0.93 0.97
70 94.60 0.91 0.95 95.60 0.93 0.96 95.80 0.93 0.97 95.00 0.94 0.97

表3

PSO优化卷积核随批次大小变化的识别率和灵敏度比较"

批次大小 随机化卷积核 高斯卷积核 PSO优化一层卷积核 PSO优化两层卷积核
训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/% 训练时间/s 识别率/% 灵敏度/%
50 3 404.27 94.40 97.40 3 420.74 95.30 97.20 3 549.56 95.00 97.80 3 608.60 96.10 97.00
100 3 321.73 91.40 99.20 3 329.66 93.10 98.20 3 428.76 92.50 93.20 3 569.05 94.30 98.00
150 3 297.91 90.40 97.60 3 298.88 92.80 97.40 3 324.80 91.10 96.60 3 455.64 93.00 93.00
200 3 199.72 89.10 99.40 3 186.80 90.50 98.60 3 251.77 89.50 100.00 3 358.09 91.60 95.00

表4

PSO优化卷积核随批次大小变化的特异度、MCC和F1Score比较"

批次大小 随机化卷积核 高斯卷积核 PSO优化一层卷积核 PSO优化两层卷积核
特异度/% MCC F1Score 特异度/% MCC F1Score 特异度/% MCC F1Score 特异度/% MCC F1Score
50 91.40 0.89 0.94 93.40 0.91 0.95 92.20 0.90 0.95 95.00 0.92 0.96
100 83.60 0.84 0.91 88.00 0.87 0.93 91.80 0.85 0.93 90.00 0.89 0.94
150 83.20 0.82 0.90 88.20 0.86 0.93 85.60 0.83 0.91 93.00 0.86 0.93
200 78.80 0.80 0.89 82.40 0.82 0.90 79.00 0.81 0.89 88.00 0.83 0.92

表5

PSO优化卷积核随网络层数变化的识别率比较"

网络层数 随机化卷积核 高斯卷积核 PSO一卷积核 PSO两卷积核 PSO三卷积核 PSO四卷积核
训练时间/s 识别率/% 训练时间/s 识别率/% 训练时间/s 识别率/% 训练时间/s 识别率/% 训练时间/s 识别率/% 训练时间/s 识别率/%
5 3 909.86 92.40 3 901.33 94.60 3 926.79 95.20 4 017.86 96.10
6 4 414.08 93.90 4 232.13 95.00 4 453.02 95.80 4 477.57 96.50 4 496.98 97.00
7 4 471.54 94.20 4 566.62 95.50 4 497.06 96.30 4 507.26 97.20 4 523.96 97.70
8 4 526.35 95.80 4 647.58 96.80 4 545.81 96.70 4 577.11 97.60 4 774.98 98.00 5 6491.26 98.30

表6

PSO优化卷积核随网络层数变化的灵敏度和特异度比较"

网络层数 随机化卷积核 高斯卷积核 PSO一卷积核 PSO两卷积核 PSO三卷积核 PSO四卷积核
灵敏度/% 特异度/% 灵敏度/% 特异度/% 灵敏度/% 特异度/% 灵敏度/% 特异度/% 灵敏度/% 特异度/% 灵敏度/% 特异度/%
5 89.40 95.40 99.20 90.00 98.40 92.00 96.80 95.40
6 99.40 88.40 93.60 96.40 98.60 93.00 96.00 97.00 98.20 95.80
7 92.20 96.20 96.40 94.60 99.80 92.80 98.00 96.40 99.00 96.40
8 94.80 96.80 99.00 94.40 98.40 95.00 98.00 97.20 98.60 97.40 100.00 96.60

表7

PSO优化卷积核随网络层数变化的MCC和F1Score比较"

网络层数 随机化卷积核 高斯卷积核 PSO一卷积核 PSO两卷积核 PSO三卷积核 PSO四卷积核
MCC F1Score MCC F1Score MCC F1Score MCC F1Score MCC F1Score MCC F1Score
5 0.85 0.92 0.90 0.95 0.91 0.95 0.92 0.96
6 0.88 0.94 0.90 0.95 0.92 0.96 0.93 0.97 0.94 0.97
7 0.88 0.94 0.91 0.96 0.93 0.96 0.94 0.97 0.95 0.98
8 0.92 0.96 0.94 0.97 0.93 0.97 0.95 0.98 0.96 0.98 0.97 0.98
1 陈万青, 郑荣寿, 张思维, 等. 2012年中国恶性肿瘤发病和死亡分析[J]. 中国肿瘤, 2016, 25 (1): 1- 8.
doi: 10.3969/j.issn.1674-4136.2016.01.001
CHEN Wanqing , ZHENG Rongshou , ZHANG Siwei , et al. Analysis of the incidence and death of malignant tumors in China in 2012[J]. Bulletin of China, 2016, 25 (1): 1- 8.
doi: 10.3969/j.issn.1674-4136.2016.01.001
2 张忠凤, 张春. 多层螺旋CT对肺癌的诊断价值[J]. 实用临床医学, 2016, 17 (1): 53- 54.
ZHANG Zhongfeng , ZHANG Chun . Multi-slice spiral CT in diagnosis of lung cancer[J]. Practical Clinical Medicine, 2016, 17 (1): 53- 54.
3 常亮, 邓小明, 周明全, 等. 图像理解中的卷积神经网络[J]. 自动化学报, 2016, 42 (9): 1300- 1312.
CHANG Liang , DENG Xiaoming , ZHOU Mingquan , et al. Convolutional neural networks in image understanding[J]. Acta Automatica Sinica, 2016, 42 (9): 1300- 1312.
4 HU Hong , PANG Liang , SHI Zhongzhi . Image matting in the perception granular deep learning[J]. Knowledge-Based Systems, 2016, 102, 51- 63.
doi: 10.1016/j.knosys.2016.03.018
5 余萍, 赵继生. 基于矩阵2-范数池化的卷积神经网络图像识别算法[J]. 图学学报, 2016, 39 (5): 694- 701.
YU Ping , ZHAO Jisheng . Convolution neural network image recognition algorithm based on matrix 2-norm pooling[J]. Journal of Graphics, 2016, 39 (5): 694- 701.
6 GRINBLAT Guillermo , UZAL Lucas , LARESE Mónica , et al. Deep learning for plant identification using vein morphological patterns[J]. Computers & Electronics in Agriculture, 2016, 127, 418- 424.
7 ISLAM Jyoti, ZHANG Yanqing. Visual sentiment analysis for social images using transfer learning approach[C]// Processdings of IEEE International Conferences on Big Data and Cloud Computing. Washington, USA: IEEE, 2016: 124-130.
8 ZHAO Wenzhi , DU Shihong . Learning multiscale and deep representations for classifying remotely sensed imagery[J]. Isprs Journal of Photogrammetry & Remote Sensing, 2016, 113, 155- 165.
9 GAO X W , HUI R , Tian Z . Classification of CT brain images based on deep learning networks[J]. Computer Methods & Programs in Biomedicine, 2017, 138, 49- 56.
10 CHEN Jie, QI Xianbiao, TERVONEN O, et al. Thorax disease diagnosis using deep convolutional neural network[C] // Processdings of Engineering in Medicine and Biology Society. Ottawa, Canada: IEEE, 2016: 2287- 2290.
11 MISHRA Manish, SCHMITT Sabine, WANG Lichao, et al. Structure-based assessment of cancerous mitochondria using deep networks[C]// Processdings of IEEE, International Symposium on Biomedical Imaging. Prague, Czech Republic: IEEE, 2016: 545-548.
12 王媛媛, 周涛, 陆惠玲, 等. 基于集成卷积神经网络的肺部肿瘤计算机辅助诊断模型[J]. 生物医学工程学杂志, 2017, 34 (4): 543- 551.
WANG Yuanyuan , ZHOU Tao , LU Huiling , et al. Computer-aided diagnosis model of lung cancer based on integrated convolutional neural network[J]. Journal of biomedical engineering, 2017, 34 (4): 543- 551.
13 赵志刚, 林玉娇, 尹兆远. 基于自适应惯性权重的均值粒子群优化算法[J]. 计算机工程与科学, 2016, 38 (3): 501- 506.
doi: 10.3969/j.issn.1007-130X.2016.03.016
ZHAO Zhigang , LIN Yujiao , YIN Zhaoyuan . A mean particle swarm optimization algorithm based on adaptive inertia weight[J]. Computer Engineering & Science, 2016, 38 (3): 501- 506.
doi: 10.3969/j.issn.1007-130X.2016.03.016
14 LAN Zhenzhong , YU Shooul , LIN Ming , et al. Handcrafted local features are convolutional neural networks[J]. Computer Science, 2015, (3): 92411- 92419.
15 OH Beonseok , OH Kangrok . A gabor-based network for heterogeneous face recognition[J]. Neurocomputing, 2017, 261, 253- 265.
doi: 10.1016/j.neucom.2015.11.137
16 BALASARASWATHI M , KALPANA B . Fast and effective classification using parallel and multi-start PSO[J]. Journal of Information Technology Research, 2018, 11 (2): 13- 30.
doi: 10.4018/JITR
17 YE Fei . Particle swarm optimization-based automatic parameter selection for deep neural networks and its applications in large-scale and high-dimensional data[J]. Plos One, 2017, 12 (12): 152- 188.
18 王媛媛.基于卷积神经网络的肺部肿瘤PET/CT计算机辅助诊断研究[D].宁夏医科大学, 2017.
WANG Yuanyuan. Computer aided diagnosis of lung tumors PET/CT based on convolution neural network[D]. Ningxia Medical University, 2017.
19 CHEN Junwen , LIU Zhigang , WANG Hongrui , et al. Automatic defect detection of fasteners on the catenary support device using deep convolutional neural network[J]. IEEE Transactions on Instrumentation & Measurement, 2018, (99): 1- 13.
20 ZHOU Honglu , SONG Mingli , PEDRYCZ Witold . A comparative study of improved GA and PSO in solving multiple traveling salesmen problem[J]. Applied Soft Computing, 2018, 64, 564- 580.
doi: 10.1016/j.asoc.2017.12.031
[1] 张璞,刘畅,王永. 基于特征融合和集成学习的建议语句分类模型[J]. 山东大学学报 (工学版), 2018, 48(5): 47-54.
[2] 何正义,曾宪华,郭姜. 一种集成卷积神经网络和深信网的步态识别与模拟方法[J]. 山东大学学报(工学版), 2018, 48(3): 88-95.
[3] 谢志峰,吴佳萍,马利庄. 基于卷积神经网络的中文财经新闻分类方法[J]. 山东大学学报(工学版), 2018, 48(3): 34-39.
[4] 赵彦霞, 王熙照. 基于SVD和DCNN的彩色图像多功能零水印算法[J]. 山东大学学报(工学版), 2018, 48(3): 25-33.
[5] 马汉杰,林霞,胥晓晖,张健,张智晟. 基于自适应粒子群算法的智能家居管理系统负荷优化模型[J]. 山东大学学报(工学版), 2017, 47(6): 57-62.
[6] 易云飞,苗剑,林郭隆,殷智. 基于牛顿力学和博弈论模型的粒子网络优化算法[J]. 山东大学学报(工学版), 2017, 47(1): 28-36.
[7] 荆业飞1,张承慧1*,徐蓓蓓2,李珂1,褚晓广1. 基于阻抗匹配的小型风电系统功率输出优化方法[J]. 山东大学学报(工学版), 2013, 43(5): 39-43.
[8] 张飞,耿红琴. 基于混沌粒子群算法的车间作业调度优化[J]. 山东大学学报(工学版), 2013, 43(3): 19-22.
[9] 徐姗姗,刘应安*,徐昇. 基于卷积神经网络的木材缺陷识别[J]. 山东大学学报(工学版), 2013, 43(2): 23-28.
[10] 刘建华1,2, 黄添强2, 严晓明2. 融合PSO算法思想的进化算法[J]. 山东大学学报(工学版), 2010, 40(5): 34-40.
[11] 王新沛1,刘常春1*,白曈2. 基于均值距离的图像分割方法[J]. 山东大学学报(工学版), 2010, 40(4): 36-41.
[12] 王振树 李林川 李波. 基于粒子群与模拟退火相结合的无功优化算法[J]. 山东大学学报(工学版), 2008, 38(6): 15-20.
[13] 陈爽,姜威 . 基于DT-DWT的医学图像融合算法[J]. 山东大学学报(工学版), 2008, 38(1): 32-35 .
[14] 翟海亭,吴晓娟,彭彰 . 一种改进的基于互信息的三维医学图像配准算法[J]. 山东大学学报(工学版), 2006, 36(4): 33-39 .
[15] 李可,刘常春,李同磊 . 一种改进的最大互信息医学图像配准算法[J]. 山东大学学报(工学版), 2006, 36(2): 107-110 .
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] 何东之, 张吉沣, 赵鹏飞. 不确定性传播算法的MapReduce并行化实现[J]. 山东大学学报(工学版), 0, (): 22 -28 .
[2] 黄劲潮. 基于快速区域建议网络的图像多目标分割算法[J]. 山东大学学报(工学版), 2018, 48(4): 20 -26 .
[3] 唐庆顺,金璐,李国栋,吴春富. 基于自适应终端滑模控制器的机械手跟踪控制[J]. 山东大学学报(工学版), 2016, 46(5): 45 -53 .
[4] 张建明, 刘泉声, 唐志成, 占婷, 蒋亚龙. 考虑剪切变形历史影响的节理峰值剪切强度准则[J]. 山东大学学报(工学版), 0, (): 77 -81 .
[5] 肖乔, 裴继红, 王荔霞, 龚志成. 基于多通道Gabor滤波模糊融合的遥感图像舰船检测[J]. 山东大学学报(工学版), 0, (): 29 -35 .
[6] 马相明, 孙霞, 张强. 轮式装载机典型作业工况构建与分析[J]. 山东大学学报(工学版), 0, (): 82 -87 .
[7] 梁泽华,崔耀东,张雨. 有顺序依赖损耗的一维下料问题[J]. 山东大学学报(工学版), 2018, 48(3): 75 -80 .
[8] 吴树芳,徐建民. 基于HITS算法的微博用户可信度评估[J]. 山东大学学报(工学版), 2016, 46(5): 7 -12 .
[9] 代时雨,刘淑琴. 状态观测器对磁悬浮平台速度与加速度的估算[J]. 山东大学学报(工学版), 2018, 48(2): 114 -120 .
[10] 钟智彦,文志强, 张潇云,叶德刚. 基于半色调图像的邻域相似性描述子方法[J]. 山东大学学报(工学版), 2016, 46(3): 58 -64 .