节点文献

基于深度学习的医学CT影像分析研究

Research on Medical CT Image Analysis Based on Deep Learning

【作者】 高鹏

【导师】 程晓荣;

【作者基本信息】 华北电力大学 , 计算机科学与技术, 2021, 硕士

【摘要】 肺癌在所有癌症中发病率最高,死亡率逐年增长,且我国肺癌新增率和死亡率高于世界平均水平,肺癌诊断与治疗刻不容缓。肺癌早期表现形式为孤立性肺结节,针对肺结节检测最有效的手段之一是计算机断层扫描(CT)。CT影像检阅和诊断对于医生是一种耗时耗力的工作,且不同医生对于同一份CT影像的诊断存在主观性差异。随着基于深度学习的计算机辅助诊断(CAD)系统的出现,CAD系统可以帮助医生诊断CT影像,降低主观性因素,提高诊断准确率,实现早期诊断和治疗,提高肺癌患者的生存机会。本文重点研究基于深度学习的肺结节分割方法,主要研究内容如下:(1)本文对肺部CT影像和肺结节分布研究发现,对肺部CT影像需要进行预处理操作和肺实质分割操作。首先对肺部CT影像进行预处理操作,对三维肺部CT影像进行切片处理,提取距离肺结节最近的三层切片图像,对切片图像进行图像去噪和图像增强,增强CT图像的特征。然后采取大津法(OTSU)算法对肺实质进行粗分割,得到初始肺实质图像。最后开运算、抽取肺实质、孔洞填充和滚球法等一系列操作,对初始肺实质图像进行细分割,得到完整的肺实质区域图像。(2)本文对深度学习中的U-Net网络模型进行研究,分析其优缺点,针对其不足,设计了一种改进的U-Net网络分割模型。针对图像输入与输出大小不一样和特征信息裁剪丢失等问题,本文使用VGG-16网络作为U-Net网络的主干特征提取网络,加强特征提取网络部分保持不变,可以有效解决输入输出图像大小不一致问题和特征信息裁剪丢失问题;针对U-Net网络模型在网络训练时出现的网络性能退化问题,本文使用Res Net网络中的残差学习单元,对每一层卷积层进行残差学习,可以有效解决网络性能退化问题;最后本文使用一种混合损失函数,可以更进一步的提升改进的U-Net网络模型的肺结节分割性能。(3)本文使用部分LUNA16数据集验证了改进的U-Net网络模型分割肺结节的性能更好。对混合损失函数系数进行设计,使用subset0数据集对不同系数下的混合损失函数进行实验,选取改进的U-Net网络分割性能最佳的损失函数系数;对选取好混合损失函数系数的改进的U-Net网络模型,使用subset0-2数据集进行实验,验证了改进的U-Net网络模型比原始的U-Net网络模型的肺结节分割性能更好;使用消融实验分析,验证了经过预处理和肺实质分割的CT图像的肺结节分割性能更好。

【Abstract】 Lung cancer has the highest incidence rate of all cancers and the mortality rate is increasing year by year,and the rate of new lung cancer cases and mortality in China is higher than the world average,diagnosis and treatment of lung cancer is urgent.The early manifestation of lung cancer is a solitary lung nodule,one of the most effective means of detecting lung nodules is computed tomography(CT),CT image review and diagnosis is a time-consuming and laborious task for doctors,and different doctors have subjective differences in the diagnosis of the same CT image.With the emergence of deep learning-based computer-aided diagnosis(CAD)systems,CAD systems can help doctors diagnose CT images,reduce subjectivity,improve diagnostic accuracy,achieve early diagnosis and treatment,and improve the survival chances of lung cancer patients.This thesis focuses on the study of lung nodule segmentation based on deep learning.The main research contents are as follows:(1)In this thesis,the study of lung CT images and lung nodule distribution reveals that pre-processing operation and lung parenchyma segmentation operation are required for lung CT images.Firstly,pre-processing operation is performed on the lung CT images,slicing the 3D lung CT images,extracting the three layers of sliced images closest to the lung nodules,and image denoising and image enhancement are performed on the sliced images to enhance the features of the CT images.Then the Otsu method(OTSU)algorithm was adopted to perform coarse segmentation of the lung parenchyma to obtain the initial lung parenchyma image.Finally,a series of operations such as open operation,lung parenchyma extraction,hole filling and rolling ball method are performed to finely segment the initial lung parenchyma image to obtain the complete lung parenchyma region image.(2)In this thesis,the U-Net network model in deep learning is studied,its advantages and disadvantages are analysed,and an improved U-Net network segmentation model is designed to address its shortcomings.To address the problems of inconsistent image input and output sizes and the loss of feature information cropping,this thesis uses the VGG-16 network as the backbone feature extraction network of the U-Net network,and strengthens the feature extraction network partly to keep the same,which can effectively solve the problems of inconsistent input and output image sizes and the loss of feature information cropping;Aiming at the network performance degradation problem of the U-Net network model during network training,this thesis uses the residual learning unit in the Res Net network to perform residual learning on each convolutional layer,which can effectively solve the network performance degradation problem;finally this thesis Using a hybrid loss funct ion can further improve the lung nodule segmentation performance of the improved U-Net network model.(3)In this thesis,I used part of the LUNA16 dataset to verify that the improved U-Net network model has better performance in segmenting lung nodules.T he mixed loss function coefficients were designed and experiments were conducted using the subset0 dataset for the mixed loss functions with different coefficients to select the best loss function coefficients for the improved U-Net network segmentation performance;the improved U-Net network model with the selected mixed loss function coefficients was experimented using the subset0-2 dataset to verify that the improved U-Net network model better lung nodule segmentation performance than the original U-Net network model;using ablation experimental analysis,the better lung nodule segmentation performance of CT images with pre-processing and lung parenchyma segmentation was verified.

节点文献中: 

本文链接的文献网络图示:

本文的引文网络