论文摘要 ( 英文 ) Nowadays, doctors need to identify multimodal medical images combined with clinical characterization to diagnose the patient's condition a

Size: px

Start display at page:

Download "论文摘要 ( 英文 ) Nowadays, doctors need to identify multimodal medical images combined with clinical characterization to diagnose the patient's condition a"

便皇苗
5 years ago
Views:

1 博士学位论文公示材料学生姓名王洋学号二级学科计算机应用导师姓名刘积仁论文题目人工智能与多模态影像组学在若干疾病分析中的研究论文研究方向医学图像处理深度学习人工智能论文关键词医学图像 ; 多模态影像组学 ; 人工智能 ; 深度学习 ; 异构数据融合论文摘要 ( 中文 ) 如今医生需要通过识别多个模态的医学图像并结合临床表征来综合诊断病人的病情, 确定病灶, 这需要大量的时间和丰富的经验而现有计算机辅助诊断多针对于某一模态图像, 忽略了其他诊断数据基于充分利用临床多模态异构数据, 通过合理的设计与训练, 人工智能可以辅助完成非常多的临床任务在人工智能领域, 机器学习与深度学习已经提高了目标检测和识别的准确率, 还在医学领域取得了不小的成绩本文选取多模态影像数据与临床数据, 并结合改进的卷积神经网络对乳腺癌帕金森疾病阿尔兹海默病进行辅助诊断具体工作如下 : 1 本文通过应用机器学习方法对乳腺钼靶影像的良恶性肿块进行区分, 建立六种机器学习模型, 分别对 BCDR-F03 数据集进行良恶性分类, 并且对六种有监督机器学习分类器的实验结果进行详细评估通过实验发现, 对训练样本量特征维数的优化组合能够降低特征维数并且明显提升分类器分类结果的准确率通过可重复性研究, 在测试过程中随机森林的分类准确率最高人工神经网络和支持向量机具有较好的泛化能力稳定性 2 本文对帕金森病人和未患帕森病者的多模态像, 将 T2-MRI 单模态图像集 PET 单模态图像集和多模态融合图像集分别训练测试卷积神经网络, 采用交叉验证法得出平均诊断准确率对设计的卷积神经网络进行改进, 通过改变卷积核的大小卷积核的个数神经网络的层数替换可以自主学习的激励函数增加 Dropout 层和引入卷积核来改进得到性能更好泛化能力更强的卷积神经网络通过两种卷积神经网络的六组实验结果表明, 卷积神经网络对多模态融合后的图像集诊断正确率高于单模态图像集两种卷积神经网络的诊断结果同样可以说明, 改进后的卷积神经网络诊断效果远优于改进前的卷积神经网络, 更适合辅助诊断帕金森病最后对卷积神经网络的特征图进行提取, 并将特征图进行可视化, 还原包含患者脑部的疾病特征的图片, 辅助医生进行诊断 3 本文选取脑肿瘤与阿尔茨海默病和未患病者进行区分诊断, 将临床表征和量表处理为矩阵形式存储以便网络进行下一步的学习和计算机辅助诊断本文基于 AlexNet 模型进行网络搭建, 并在每个卷积核上放置一个先验分布, 并将每个 kernel-patch 对与变分分布进行近似整合, 以适应本文中较小的数据样本, 防止过拟合在搭建的改进后的 AlexNet CNN 上运行小数据量的 CIFAR-10 数据集结果与改进前相比, 测试精度有了显著提高在进行诊断训练前先测试网络性能并证明网络具有良好的泛化能力, 同时测试不同的激活函数对网络的影响进行网络测试后, 文中采用十折交叉验证法进行了对计算机辅助诊断结果的对比与分析, 将单独通过 MRI-CNN PET-CNN 单独诊断的结果与将 MRI-CNN PET-CNN 提取到的特征和临床数据输入 SVM 进行分类的结果加以对比, 改进后网络的平均准确率达 92.35% 最后提取特征图进行灰度映射后还原显示特征, 用于后期诊断充分利用临床异构数据, 发挥人工智能在医疗领域的潜力, 辅助医生的临床决策, 为患者提供更准确的治疗方案, 可提升病人的生存质量和生存期, 并能最大限度的利用医疗资源, 形成和谐的医患关系, 促进我国卫生健康事业和谐发展 1

2 论文摘要 ( 英文 ) Nowadays, doctors need to identify multimodal medical images combined with clinical characterization to diagnose the patient's condition and determine the lesion, which requires a lot of time and experience. The existing computer-aided diagnosis system ignores other diagnostic data for multi-modal images. Based on the full use of clinical multi-modal heterogeneous data, through reasonable design and training, artificial intelligence can assist to complete a lot of clinical tasks. In the field of artificial intelligence, machine learning and deep learning have improved the accuracy of target detection and recognition, and have made great achievements in the field of medicine. In this thesis, multimodal imaging data and clinical data are selected, and the improved convolution neural network is used to diagnose breast cancer, Parkinson's disease and Alzheimer's disease. The specific work is as follows: 1. This thesis used machine learning method to distinguish benign and malignant breast tumors in molybdenum target images, established six machine learning models, classifies BCDR-F03 data sets for benign and malignant tumors, and evaluated the experimental results of six supervised machine learning classifiers in detail. The experiments showed that the optimal combination of training sample size and feature dimension can reduce the feature dimension and significantly improve the accuracy of classification results. Through repeatability study, the accuracy of random forest classification was the highest during the test. Artificial neural network and support vector machine have better generalization ability and stability. 2. For the multimodal images of Parkinson's disease patients and non-parkinson's disease patients, the T2-MRI single-modal image set, PET single-modal image set and multimodal fusion image set were trained to test convolution neural network, and the average diagnostic accuracy was obtained by cross-validation. The convolution neural network was improved by changing the size of convolution kernel, the number of convolution kernels, the layers of neural network, replacing the excitation function which can be learnt independently, adding Dropout layer and introducing convolution kernel to improve the convolution neural network with better performance and stronger generalization ability. Six experimental results of two kinds of convolutional neural networks showed that the diagnostic accuracy of convolutional neural networks was higher than that of single-mode image sets. The diagnostic results of the two kinds of convolution neural networks also showed that the improved convolution neural network was much better than the improved convolution neural network in diagnosing Parkinson's disease. Finally, the feature map of convolutional neural network was extracted, and the feature map was visualized to restore the image containing the patient's brain disease features to assist doctors in diagnosis. 3. In this paper, brain tumors and Alzheimer's disease and undiagnosed patients were selected for differential diagnosis. The clinical manifestations and scales were processed into matrix form and stored for network learning and computer-aided diagnosis. In this paper, we built the network based on AlexNet model, place a prior distribution on each convolution kernel, and integrated each kernel-patch with the variational distribution approximately to adapt to the smaller data samples in this paper, to prevent over-fitting. The CIFAR-10 data set with small amount of data running on the improved AlexNet CNN was more accurate than before. Before diagnostic training, the network performance was tested and proved to have good generalization ability. Meanwhile, the influence of different activation functions on the network was tested. After the network test, ten fold cross validation method was used to compare and analyze the results of computer-aided diagnosis. The results of MRI-CNN and PET-CNN were compared with those of MRI-CNN and PET-CNN respectively, and the features extracted from MRI-CNN and PET-CNN and clinical data were input into SVM for classification. The average standard of the improved network was also compared. The accuracy was 92.35%. Finally, the feature map was extracted to restore the display features after gray mapping, and was used for later diagnosis. 2

3 Making full use of clinical heterogeneous data, giving full play to the potential of artificial intelligence in the medical field, assisting doctors in clinical decision-making, providing patients with more accurate treatment programs, can improve the quality of life and survival of patients, and can maximize the use of medical resources, form a harmonious doctor-patient relationship, and promote China's health cause Harmonious development. 3

4 论文主要创新点 1 利用多种机器学习方法利用纹理分析法对乳腺癌钼靶影像进行特征提取建立四种机器学习模型对乳腺钼靶影像进行分类, 并且评估不同分类器的最优样本量和特征量 2 设计并实现适合多模态帕金森疾病的医学图像融合算法 3 设计并搭建适合帕金森病的卷积神经网络, 通过网络的搭建对帕金森病多模态图像进行特征提取 4 基于 AlexNet 设计并搭建网络模型, 对患者 MRI PET 图像和临床数据进行训练和测试评估, 并对网络进行了改进将采用新型 dropout 机制以适应较小数据量的卷积神经网络提取的特征和临床数据输入支持向量机进行分类 5 利用改进后的神经网络对阿尔茨海默病脑肿瘤和未患病者的异构数据进行计算机辅助诊断, 对患者是否患病及其疾病类型进行判断, 并探索研究还原疾病特征攻读博士学位期间取得的学术成果注 : 按有关规范填写本人为第一作者或导师是第一作者本人为第二作者, 且第一署名单位为东北大学的研究成果 1 Wang Yang, Yiqi Zhao, Jiren Liu, Dazhe Zhao, Zuitian Tao, Auxiliary diagnosis of heterogeneous data of Parkinson's disease based on improved convolution neural network[j], Journal Medical Imaging and Health Information, 2018,SCI 检索, 已录用. 2 王洋; 刘积仁 ; 赵大哲 ; 邱道云, 卷积神经网络在 MRI 图像诊断中的应用 [J], 东北大学学报自然科学版,EI 检索, 已录用. 3 Wang Yang, Jin-Ren Liu, Research and development of medical image fusion, ICMIPE Conference2013,EI 检索, 已见刊. 4

5 5

报告 1: 郑斌教授, 美国俄克拉荷马大学医学图像特征分析与癌症风险评估方法摘要 : 准确的评估癌症近期发病风险和预后或者治疗效果是发展和建立精准医学的一个重要前

报告 1: 郑斌教授, 美国俄克拉荷马大学医学图像特征分析与癌症风险评估方法摘要 : 准确的评估癌症近期发病风险和预后或者治疗效果是发展和建立精准医学的一个重要前东北大学中荷生物医学与信息工程学院 2016 年度生物医学与信息工程论坛会议时间 2016 年 6 月 8 日, 星期三,9:30 至 16:00 会议地址会议网址主办单位东北大学浑南校区沈阳市浑南区创新路 195 号生命科学大楼 B 座 619 报告厅 http://www.bmie.neu.edu.cn