肺结节CT图像中基于多任务特征的语义属性自动评级

发布时间：2018-03-03 17:50

本文选题：肺结节　切入点：语义特征　出处：《深圳大学》2017年硕士论文　论文类型：学位论文

【摘要】：肺癌是发病率、死亡率增长最快的恶性肿瘤之一,早期诊断对肺癌治疗至关重要。肺结节是肺癌的常见表现形式,临床常用计算机断层扫描技术(Computed tomography,CT)筛查肺结节病变。但肺结节CT图像表征多样,人眼筛查具有工作量大、观察者间差异明显、重复性差等缺陷。为解决上述缺陷,肺结节计算机辅助诊断系统(Computer-aided Diagnosis,CAD)得以发展。本文就此提出一种基于多任务学习框架的肺结节自动评级系统,自动给出肺结节CT图像临床语义特征的级别,提高诊断效率与精度。本文提出两种肺结节辅助诊断系统。首先,我们提出基于多任务线性回归(Multi-task Learning Regression,MTLR)模型的评级方法,对肺结节8个语义特征自动作出评级。其次,我们提出多属性辅助诊断肺癌(Multiple Attribute-assisted Diagnosis of Lung Cancer,MAADLC)的模型,在预测8个语义特征级别的同时给出肺结节良恶性诊断结果。计算机特征和临床语义特征的鸿沟是CAD发展的瓶颈。为此本文首先提出使用异构特征的MRLR模型。由描述图像高层次计算机特征的卷积神经网络(Convolutional Neural Network,CNN)特征、堆叠降噪自动编码(Stacked Denoising Autoencoder,SDAE)特征和描述图像灰度低层次计算机特征的随机Haar特征形成异构特征,随后将每个语义特征作为一个任务,使用MRLR模型发掘不同任务间的共享特征和独立特征,自动回归出8个任务的评级结果。其次,我们提出MAADLC模型对肺结节诊断问题进行改进和效果提升。在特征提取上,新增方向梯度直方图(Histogram of Oriented Gradient,HoG)特征,描述肺结节图像的形状与边界信息;在分级问题上,提出多种训练框架的MAADLC模型,将8个语义特征作为辅助信息与良恶性信息联合训练,联合解决分类回归混合问题及高层次诊断问题,达到诊断良恶性的最终目的。我们使用肺部图像数据联盟(The Lung Image Database Consortium,LIDC)公共数据集作为测试数据,在8个语义特征评级上,四组医师组间差异为0.58±0.78(绝对误差均值±标准差,下同),MRLR模型预测结果为0.65±0.56;在联合肺结节良恶性的9个临床语义特征评级上,四组医师组间差异为0.70±0.79,MAADLC混合模型预测结果为0.69±0.59,预测结果接近医师组间差异,说明预测方法达到医师人眼筛查的准确率。本文首次提出肺结节9种属性联合自动评级的方法,该系统使用当前流行的机器学习手段,有效解决临床实际问题。本文提出的评级检索框架是通用的,可扩展使用至医学图像检索引擎中。
[Abstract]:Lung cancer is one of the malignant tumors with the fastest morbidity and mortality. Early diagnosis is very important to the treatment of lung cancer. Pulmonary nodules are a common manifestation of lung cancer. Computed tomographic CT (CT) is commonly used to screen pulmonary nodule lesions in clinic, but the CT images of pulmonary nodules are diverse, the human eye screening has many defects, such as heavy workload, obvious differences among observers, poor repeatability and so on. The computer-aided diagnosis system for pulmonary nodules (CAD) has been developed. In this paper, an automatic classification system of pulmonary nodules based on multi-task learning framework is proposed, which automatically gives the level of clinical semantic features of CT images of pulmonary nodules. In this paper, two kinds of pulmonary nodule assistant diagnosis systems are proposed. First, we propose a multi-task Learning model based on multi-task Learning model to automatically rate eight semantic features of pulmonary nodules. We present a multiple Attribute-assisted Diagnosis of Lung carcinoma MAADLC model. The diagnosis results of benign and malignant pulmonary nodules are given while predicting eight semantic feature levels. The gap between computer and clinical features is the bottleneck of CAD development. In this paper, a MRLR model using heterogeneous features is proposed firstly in this paper. Convolutional Neural Network CNNs that describe high-level computer features of images, Stacked Denoising autocode (SDAE) features and random Haar features that describe the image's low-grayscale computer features form heterogeneous features, and then each semantic feature is treated as a task. The MRLR model is used to explore the shared and independent features among different tasks, and the grading results of 8 tasks are automatically regressed. Secondly, we propose an improved MAADLC model to improve the diagnosis of pulmonary nodules. A new histogram of Oriented histogram was added to describe the shape and boundary information of pulmonary nodule images, and a MAADLC model of various training frames was proposed, in which eight semantic features were used as auxiliary information and benign and malignant information for training. To solve the mixed problem of classification and regression and the problem of high-level diagnosis, we use the common data set of the Lung Image Database consortiums (LIDC) as the test data to achieve the ultimate goal of diagnosing benign and malignant tumors. In the eight semantic feature ratings, we use the common data set of the lung image data alliance (Lung Image Database consortiums) as the test data. The difference among the four groups was 0.58 卤0.78 (mean absolute error 卤standard deviation) and the predicted result of MRLR model was 0.65 卤0.56. The difference among the four groups was 0.70 卤0.79% MAADLC mixed model and 0.69 卤0.59. The predicted results were close to the difference between the physician groups, which indicated that the prediction method reached the accuracy of the doctor's eye screening. The method of combined automatic rating of 9 attributes of pulmonary nodules was put forward for the first time in this paper. The system uses the popular machine learning method to solve the practical clinical problems effectively. The rating retrieval framework proposed in this paper is universal and can be extended to the medical image retrieval engine.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;R734.2

【相似文献】