图1 提出方法的总流程图
纸质出版日期:2021-07-25,
网络出版日期:2021-06-02,
收稿日期:2020-11-04,
录用日期:2021-01-19
扫 描 看 全 文
引用本文
阅读全文PDF
基于医学影像的身体部位识别旨在确定特定医学影像所属的身体部位,是许多医学影像分析任务必不可少的预处理步骤。目前,计算机断层扫描(CT,Computed Tomography)技术是临床中最常用的医学影像技术之一,而基于CT图像的医学影像分析算法(例如,病灶检测、器官分割等)同样需要首先确定CT图像所属身体部位以获取先验知识,从而保证算法的速度及鲁棒性。然而CT图像是由被称为CT弦图(CT Sinogram)的CT原始数据重建得到,而图像重建过程有可能导致信息丢失。因此,相较于CT图像,CT弦图中应该包含更多适用于身体部位识别的有效信息。然而,目前基于CT弦图的身体部位识别研究仍比较少。鉴于此,本研究使用基于卷积神经网络(CNN,Convolutional Neural Network)的深度学习对CT弦图进行特征学习,并验证其在身体部位识别任务中的可用性。实验采用了一个公开数据集(DeepLesion)和来自著名医学机构的三个临床数据集来验证本研究提出方法的性能。具体而言,本研究通过Radon变换理论对CT图像进行数据仿真得到CT弦图,并以CT弦图作为输入,构造基于CNN的分类器(Sino-Net),从而对五个最常见的身体部位(头部、颈部、胸部、上腹部以及骨盆)进行识别。实验结果表明,使用CT弦图进行身体部位识别可以达到与使用CT图像进行身体部位识别相似的性能,甚至优于基于CT图像识别的结果。
Medical-image-based human bodypart recognition,which aims to accurately locate the bodypart of a specific medical image,is an essential preprocessing step for many medical image analysis tasks.Currently,computed tomography (CT) is one of the most available medical imaging techniques in clinic.Many CT-based medical image analysis algorithms (such as lesion detection,organ segmentation,etc.) need to first identify the bodypart information contained in the CT image to obtain prior knowledge,so as to ensure the speed and robustness of the algorithms.However,CT images are reconstructed from CT raw data,which is also known as CT sinogram.And the image reconstruction process may cause information loss.Therefore,compared with CT images,CT sinogram may contain more effective information suitable for bodypart recognition tasks.However,there are still relatively few researches on bodypart recognition based on CT sinogram.Therefore,the deep learning based convolutional neural network (CNN) technique is used to train on CT sinogram and its usability in bodypart recognition tasks is verified.A public dataset (i.e.,DeepLesion) and three clinical datasets from well-known medical institutions are adopted to verify the performance of our proposed method.Specifically,the Radon transform is used to perform data simulation on CT images to obtain CT sinogram,which is served as input to train a CNN-based classifier (Sino-Net) to recognize the five most common bodyparts (i.e.,head,neck,chest,upper abdomen and pelvis).The experimental results show that the use of CT sinogram for bodypart recognition can achieve similar performance to the use of CT images,and sometimes even better than the results based on CT images.
基于医学影像的身体部位识别(BPR,Bodypart Recognition)旨在准确定医学影像所属身体部位(例如,头部、胸部或腹部等),是许多医学影像分析算法的预处理步骤[
医学影像种类良多,其中包括计算机断层扫描(CT,Computed Tomography)[
基于CT图像的身体部位识别研究本质上是一个多分类问题。在过去十几年中,学者们已经提出了许多基于CT图像的身体部位识别方法。例如,田野等[
近年来,具有强大端到端学习能力的深度学习(DL,Deep Learning)技术,特别是卷积神经网络技术(CNN,Convolutional Neural Network)被广泛用于CT图像分析任务以及身体部位识别研究中,取得了一定的成功[
考虑到上述影响因素,本研究提出利用深度学习对更广泛的CT数据类型进行CT弦图学习,并用于身体部位识别。具体而言,本文方法以CT弦图作为CNN分类器的输入,从而构造基于CNN的五分类器(称为Sino-Net),对五个身体部位(头部、颈部、胸部、上腹部以及骨盆)进行识别。为了评估本文提出方法的有效性,三种常见的CNN结构(残差网络(ResNet)[
本文采用公开数据集(DeepLesion)来训练和五折交叉验证本文提出的方法,即使用CNN分类器对CT弦图进行学习,从而识别五个身体部位。DeepLesion数据集由美国国立卫生研究院临床中心(NIHCC)创建,是目前世界上最大的CT影像数据集。该数据集由4 427名患者的CT图像组成,涵盖了大多数人体解剖结构[
本文使用的三个临床数据集仅用于独立测试以评估所提出方法的性能。(a)胸部数据集:该数据集在美国德克萨斯大学西南医学中心(UTSW)收集,由100例早期(IA和IB)非小细胞肺癌(NSCLC)患者组成。共包括579张胸部CT图像,其图像大小为512×512,使用通用电气(GE)CT扫描设备扫描。(b)上腹部数据集:该数据集由2010年1月至2019年5月在南方医科大学附属珠江医院肝胆外科二科就诊的135例胰腺疾病患者组成。共包括1 537张上腹部CT图像,其图像大小为512×512,使用飞利浦(Philips)CT扫描设备扫描。(c)骨盆数据集:该数据集在美国德克萨斯西南医学中心收集,由15例 IB-IVA子宫颈癌患者组成。共包括836张骨盆CT图像,其图像大小为512×512,使用通用电气(GE)CT扫描设备扫描。
由于真实CT弦图为各厂商的商业机密,难以获取。因此,本研究通过数据仿真的方式对CT图像进行仿真从而获得弦图数据。此仿真方法已广泛应用于需要用到CT弦图数据的研究[
bi=Poisson{I0e-yi}+Normal(0,σ2e) , | (1) |
其中bi是第i个探测器的仿真数据(弦图),yi是对应正常剂量条件下的线积分值,σ2e是背景电子噪声方差,σ2e的值被假设用于商用CT扫描设备是稳定的,I0为X射线入射光子数量,本文设置I0=3×106为参考值,即正常剂量条件下仿真的数据。
本文提出算法的总流程图如
pi=exp (yi)∑5j=1exp(yj), i=1,2,…,5. | (2) |
图1 提出方法的总流程图
Fig.1 Illustration of the proposed method
最后,使用最大概率原则确定Sino-Net预测的类别L,即
L=argmax1≤i≤5 (pi) . | (3) |
当L为1,2,3,4或5时,则Sino-Net预测的身体部位分别为头部、颈部、胸部、上腹部或骨盆。
本实验使用RMSProp算法(http://www.cs.toronto.edu/∼tijmen/csc321/slides/lecture_slides_lec6.pdf)对Sino-Net进行优化训练,损失函数为平均交叉熵(Mean Cross Entropy),其定义为
Loss=-1N∑5i=1lilog pi, i=1,2,…,5 , | (4) |
其中N是训练样本的数量,li是相应样本的标签,满足li∈{0,1}且∑5i=1li=1。
实验使用一个具有24 GB内存容量的NVIDIA Tesla P40图形处理器(GPU)的PyTorch工具包进行实验。参数动量(Momentum)可加速学习过程,根据实践经验,我们将其设置为0.9。批大小(Batch Size)为150,对所有网络进行了300个周期的训练,学习率为10-5。
为了验证Sino-Net对广泛数据类型的CT弦图的学习能力和对身体部位的识别性能,本文对三种最常用的CNN结构进行修改,从而执行本文的实验。
2.2.1 基于ResNet修改的CNN结构
残差网络(ResNet)于2015年被首次提出,它在ImageNet大规模视觉识别挑战赛(ILSVRC)中获得了图像分类的优胜。残差网络易于优化,并且其内部跳跃连接(Shortcut)结构可以缓解由于神经网络深度增加而导致的梯度消失问题。因此,本文首先基于残差网络的跳跃连接构建了CNN模型,此模型被称为Res-BPR,如
图2 Res-BPR,它由八个卷积层、四个最大池化层和一个全连接层组成
Fig.2 Res-BPR, which consists of eight convolution layers, four max-pooling layers and one FC layer
2.2.2 基于DenseNet修改的CNN结构
密集连接卷积网络(DenseNet)于2017年被首次提出。自此之后,DenseNet被广泛用于各种图像分析任务中并取得了上佳的效果。DenseNet具有较强的泛化能力,它的成功得益于它内部的密集连接块(DB,Dense Block)结构,该结构能减轻网络训练过程中出现的过拟合问题。鉴于密集连接块的优点,本文构建了CNN网络结构Dense-BPR,如
图3 Dense-BPR,它由一个卷积层、四个密集连接卷积块、三个过渡层、一个平均池化层和一个全连接层组成。两个相邻的密集连接块由过渡层连接
Fig.3 Dense-BPR, which consists of one convolution layer, four Dense Blocks, three transition layers (TLs), one average-pooling layer, and one FC layer. Two adjoining Dense Blocks are connected by a TL
2.2.3 基于Inception网络修改的CNN结构
自GoogLeNet于2014年在ILSVRC中获得第一名以来,Inception模块引起了学者们的广泛关注。Inception模块使用多个小卷积核代替大卷积核,从而提高了参数的利用率并加快了网络的计算速度。如
图4 Incept-BPR,它由三个Inception模块、一个卷积层和一个全连接层组成
Fig.4 Incept-BPR, which consists of three Inception modules, one convolution layer and one FC layer
整体实验设置如
图5 实验设置总流程图
Fig.5 Illustration of experiment setup
实验的评估准则包括准确率(Acc,Accuracy),宏F1(macro-F1),接收者操作特征曲线(Receiver Operating Characteristic Curve,ROC曲线)[
Acc= ∑5i=1QiN, | (5) |
其中N代表测试样本的数量,Qi为测试样本中第i类预测正确的样本个数。宏F1公式如下
macro-F1=2×Pmacro×RmacroPmacro+Rmacro, | (6) |
其中Rmacro代表宏查准率:Rmacro=15∑5i=1TPiTPi+FPi,Pmacro代表宏查全率:Pmacro=15∑5i=1TPiTPi+FNi,TPi为真阳性率,FPi为假阳性率,FNi为假阴性率。i代表类别个数,在本实验中,共有5个类别,分别为头部、颈部、胸部、上腹部以及骨盆。
方法 | 准确率 | 宏F1 | ||||
---|---|---|---|---|---|---|
Res-BPR | Dense-BPR | Incept-BPR | Res-BPR | Dense-BPR | Incept-BPR | |
Img-Net | 98.52±0.01 | 98.34±0.02 | 98.22±0.04 | 98.52±0.02 | 98.35±0.01 | 98.22±0.04 |
Sino-Net | 99.71±0.01 | 99.77±0.03 | 99.06±0.02 | 99.71±0.01 | 99.76±0.03 | 99.07±0.03 |
1)Img-Net是以CT图像作为输入的CNN分类器,Sino-Net是以CT弦图作为输入的CNN分类器。
方法 | LR | LDA | KNN | CART |
---|---|---|---|---|
准确率 | 74.48±0.40 | 77.57±0.14 | 82.50±0.35 | 77.21±0.32 |
方法 | RF | NB | SVM | Sino-Net-Dense-BPR |
准确率 | 85.52±0.42 | 53.68±2.07 | 58.45±0.94 | 99.77±0.03 |
图6 七种传统分类方法和Sino-Net-Dense-BPR的ROC曲线比较
Fig.6 The comparison of ROC curve between seven traditional classification methods and the Sino-Net-Dense-BPR
为了验证本文提出的方法的泛化能力,来自三个不同机构的临床数据集将用于外部验证,以独立测试本文提出方法的性能,结果展示在
本文使用基于CNN的深度学习方法对广泛数据类型的CT弦图数据进行学习并应用于身体部位自动识别,通过改进三种常用CNN结构(ResNet,DenseNet和Inception网络)分别构建三个CNN五分类器,并采用公开数据集DeepLesion和三个临床数据分别进行模型训练、五折交叉验证和外部验证。实验结果表明,基于CT弦图的CNN分类器(Sino-Net)在身体部位识别任务中能达到和基于CT图像方法进行识别类似的效果,甚至优于基于CT图像识别的结果。这表明基于CT弦图的医学图像分析任务值得进一步探究,具有潜力。虽然本实验取得了不错的结果,但是本文只对临床CT扫描中五个最常见的身体部位进行分类识别,可能不足以用于临床实践。在将来的研究中,我们将进一步细分身体部位探究本文方法的性能。此外,由于身体部位过渡区域界限的不明确性[
王芹, 王然冉, 姜述凤, 等. 关于人体部位识别的一种模糊算法[J]. 科学技术与工程, 2004, 4(8):687-690. [百度学术]
WANG Q, WANG R R, JIANG S F, et al. A fuzzy algorithm of recognizing human body parts [J]. Science Technology and Engineering,2004, 4(8):687-690. [百度学术]
COOTES T F, HILL A, TAYLOR C J, et al. Use of active shape models for locating structures in medical images [J]. Image & Vision Computing, 1994, 12(6): 355-365. [百度学术]
LI W, YANG Y, ZHANG K, et al. Dense anatomical annotation of slit-lamp images improves the performance of deep learning for the diagnosis of ophthalmic disorders [J]. Nature Biomedical Engineering, 2020, 4(8): 1-11. [百度学术]
ZHANG K, LIU X, SHEN J, et al. Clinically applicable AI system for accurate diagnosis, quantitative measurements, and prognosis of COVID-19 pneumonia using computed tomography [J]. Cell, 2020, 181(6): 1423-1433. [百度学术]
ZLOCHA M, DOU Q, GLOCKER B. Improving RetinaNet for CT lesion detection with dense masks from weak RECIST labels[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 2019: 402-410. [百度学术]
TANG Y, GAO R, LEE H H, et al. High-resolution 3D abdominal segmentation with random patch network fusion [J]. Medical Image Analysis, 2020, 69: 101894. [百度学术]
TONG N, GOU S, YANG S, et al. Fully automatic multi‐organ segmentation for head and neck cancer radiotherapy using shape representation model constrained fully convolutional neural networks [J]. Medical Physics, 2018, 45(10): 4558-4567. [百度学术]
XU X, ZHOU F, LIU B, et al. Efficient multiple organ localization in CT image using 3D region proposal network [J]. IEEE Transactions on Medical Imaging, 2019, 38(8): 1885-1898. [百度学术]
TANG Y, GAO R, HAN S, et al. Body part regression with self-supervision [J]. IEEE Transactions on Medical Imaging, 2021.DOI:10.1109/TMI.2021.3058281. [百度学术]
YAN K, LU L, SUMMERS R M. Unsupervised body part regression via spatially self-ordering convolutional neural networks [C]//IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE, 2018: 1022-1025. [百度学术]
MEINE H, HERING A. Efficient prealignment of CT scans for registration through a bodypart regressor[J]. arXiv:1909.08898, 2019. [百度学术]
KARCAALTINCABA M, AKTAS A. Dual-energy CT revisited with multidetector CT: review of principles and clinical applications [J]. Diagnostic and Interventional Radiology, 2011,17(3): 181-194. [百度学术]
BASSER P J, JONES D K. Diffusion‐tensor MRI: theory, experimental design and data analysis–a technical review [J]. NMR in Biomedicine: An International Journal Devoted to the Development and Application of Magnetic Resonance In Vivo, 2002, 15(7/8): 456-467. [百度学术]
MILDENBERGER P, EICHELBERG M, MARTIN E. Introduction to the DICOM standard [J]. European Radiology, 2002, 12(4): 920-927. [百度学术]
GUELD M O, KOHNEN M, KEYSERS D, et al. Quality of DICOM header information for image categorization [J]. Proc Spie Medical Imaging, 2002,4685: 280-287. [百度学术]
MUSTRA M, DELAC K, GRGIC M. Overview of the DICOM standard [C]// 2008 50th International Symposium ELMAR, IEEE, 2008,1: 39-44. [百度学术]
张乐锋, 郑逸, 傅超. 用改进的深度差分特征识别人体部位[J]. 微型机与应用, 2015, 34(14): 54-57. [百度学术]
ZHANG L F, ZHENG Y, FU C. Improved depth comparison feature for the recognition of human parts [J]. Microcomputer & Its Applications, 2015, 34(14): 54-57. [百度学术]
田野, 姜娈, 李强. 基于CT 图像的身体部位自动识别方法[J]. 计算机工程与设计, 2017, 38(1): 247-252. [百度学术]
TIAN Y, JIANG L, LI Q. Automated localization of body part in CT images [J]. Computer Engineering and Design, 2017, 38(1): 247-252. [百度学术]
PARK J, KANG G, PAN S, et al. A novel algorithm for identification of body parts in medical images [C]//International Conference on Fuzzy Systems and Knowledge Discovery, 2006: 1148-1158. [百度学术]
HONG L, HONG S. Methods and apparatus for automatic body part identification and localization [J]. U S Patent App, 2008,11.DOI:US20080112605. [百度学术]
CRIMINISI A, ROBERTSON D, KONUKOGLU E, et al. Regression forests for efficient anatomy detection and localization in computed tomography scans [J]. Medical Image Analysis, 2013, 17(8): 1293-1303. [百度学术]
LI S, XU P, LI B, et al. Predicting lung nodule malignancies by combining deep convolutional neural network and handcrafted features [J]. Physics in Medicine & Biology, 2019, 64(17): 175012. [百度学术]
ZHOU Z, LI S, HAO H, et al. A multi-objective based feature selection method for lung nodule malignancy classification [C]//Medical Physics, NJ USA, 2018,45: E678-E678. [百度学术]
林鹏, 张超, 李竹良, 等. 基于深度图像学习的人体部位识别[J]. 计算机工程, 2012, 38(16): 185-188. [百度学术]
LIN P, ZHANG C, LI Z L, et al. Human body part recognition based on depth image learning [J]. Computer Engineering, 2012, 38(16): 185-188. [百度学术]
ROTH H R, LEE C T, SHIN H C, et al. Anatomy-specific classification of medical images using deep convolutional nets [C]//IEEE 12th International Symposium on Biomedical Imaging (ISBI), 2015: 101-104. [百度学术]
YAN Z, ZHAN Y, PENG Z, et al. Multi-instance deep learning: discover discriminative local anatomies for bodypart recognition [J]. IEEE Transactions on Medical Imaging, 2016, 35(5): 1332-1343. [百度学术]
ZHANG P, WANG F, ZHENG Y. Self supervised deep representation learning for fine-grained body part recognition [C]//2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). IEEE, 2017: 578-582. [百度学术]
HE J, WANG Y, MA J. Radon Inversion via Deep Learning [J]. IEEE Transactions on Medical Imaging, 2020, 39(6): 2076-2087. [百度学术]
HE J, YANG Y, WANG Y, et al. Optimizing a parameterized plug-and-play ADMM for iterative low-dose CT reconstruction [J]. IEEE Transactions on Medical Imaging, 2019, 38(2): 371-382. [百度学术]
LEE H, HUANG C, YUNE S, et al. Machine friendly machine learning: interpretation of computed tomography without image reconstruction [J]. arXiv:1812.01068, 2018.[https://arxiv.org/abs/1812.01068v1] [百度学术]
HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778. [百度学术]
HUANG G, LIU Z, MAATEN L V D, et al. Densely connected convolutional networks [C]//IEEE Conference on Computer Vision & Pattern Recognition, 2017: 2261-2269. [百度学术]
SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 2818-2826. [百度学术]
KE Y, WANG X, LE L, et al. DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning [J]. Journal of Medical Imaging, 2018, 5(3): 036501.DOI:10.1117/1.JMI.5.3.036501. [百度学术]
YU L, SHIUNG M, JONDAL D, et al. Development and validation of a practical lower-dose-simulation tool for optimizing computed tomography scan protocols [J]. Journal of Computer Assisted Tomography, 2012, 36(4): 477-487. [百度学术]
ELBAKRI I A, FESSLER J A. Statistical image reconstruction for polyenergetic X-ray computed tomography [J]. IEEE Transactions on Medical Imaging, 2002, 21(2): 89-99. [百度学术]
KEATING K A. Use and interpretation of logistic regression in habitat-selection studies [J]. Journal of Wildlife Management, 2011, 68: 774-789. [百度学术]
YE J, JANARDAN R, QI L, et al. Feature extraction via generalized uncorrelated linear discriminant analysis [C]//Machine Learning, Proceedings of the Twenty-first International Conference (ICML), Banff, Alberta, Canada, 2004. [百度学术]
GUO G, WANG H, BELL D, et al. KNN model-based approach in classification [C]// OTM Confederated International Conferences "On the Move to Meaningful Internet Systems", 2003: 986-996. [百度学术]
SABARIAH M M K, HANIFA S A, SA'ADAH M S. Early detection of type II Diabetes Mellitus with random forest and classification and regression tree (CART) [C]// International Conference of Advanced Informatics: Concept, Theory and Application (ICAICTA), IEEE, 2014: 238-242. [百度学术]
BREIMAN L. Random forests [J]. Machine Learning, 2001, 45(1): 5-32. [百度学术]
MOŽINA M, DEMŠAR J, KATTAN M, et al. Nomograms for visualization of naive bayesian classifier [C]//European Conference on Principles of Data Mining and Knowledge Discovery, Berlin, Heidelberg, 2004: 337-348. [百度学术]
SELVARAJ H, SELVI S T, SELVATHI D, et al. Brain MRI slices classification using least squares support vector machine [J]. International Journal of Intelligent Computing in Medical Sciences & Image Processing, 2007, 1(1): 21-33. [百度学术]
DAVIS J, GOADRICH M. The relationship between precision-recall and ROC curves [C]//the 23rd International Conference on Machine Learning, 2006: 233-240. [百度学术]
LOBO J M, JIMÉNEZ-VALVERDE A, REAL R. AUC: a misleading measure of the performance of predictive distribution models [J].Global Ecology and Biogeography, 2008,17(2): 145-151. [百度学术]
YAN K, WANG X, LU L, et al. Deep lesion graphs in the wild: relationship learning and organization of significant radiology image findings in a diverse large-scale lesion database [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 9261-9270. [百度学术]
85
浏览量
131
下载量
0
CSCD
相关文章
相关作者
相关机构