【作者】 林涛；

【导师】 鲍旭东； 苏娟；

【作者基本信息】 东南大学 ， 计算机技术（专业学位）， 2017， 硕士

【摘要】 胸片因其廉价、安全等特性,被广泛运用于肺部疾病的常规检查。然而肺部病灶的判读常常受到诸如肋骨,锁骨等骨组织的干扰,造成医务人员对病人病情的误判。通过技术手段去除胸片中的骨组织,无论对医务人员还是计算机辅助诊断系统都有好处。本文先通过双能减影数据训练卷积神经网络,学习普通胸片与骨组织胸片的数值映射关系,再将所学到的网络模型运用于去除胸片中的骨组织,属于一种虚拟双能减影技术。本文首先介绍了基本骨组织去除卷积神经网络,网络以常规胸片的梯度图作为输入,以常规胸片减去软组织胸片的梯度图作为理想图片,通过Adam算法进行训练。针对卷积网络训练时间过长的问题,本文通过将大卷积核换成小卷积核和降低卷积层输出特征尺寸两种方法,在保证网络精度的同时,缩减了网络的训练时间和运行时间。基于基本卷积网络,本文讨论了不同网络参数和不同网络结构对网络精度和运行时间的影响。实验结果显示骨组织去除卷积神经网络无论在客观数值指标还是在主观视觉感受上都较好的去除了胸片中的骨组织。本文最后针对基本网络使用固定小块进行训练的不足,通过训练多个不同分辨率的卷积网络,并将网络输出特征进行融合,规避了固定小块提取信息有限的问题。相比基本卷积网络,多分辨率卷积网络在骨组织去除数值指标上有较大幅度的提高。更多还原

【Abstract】 Chest radiography(CXR)is widely used in routine examination of lung diseases for its cheapness and safety.However,the interpretation of lung lesions is often interfered by the bone structures,such as ribs and clavicle.Suppression of bones in chest radiography would be potentially useful for radiologists as well as computer-aided system.This article introduces a method which uses dual-energy subtraction data to train the convolution neural network,learning the mapping between chest radiography and bone structures.Then the statistic model is used to suppress the bone structure in chest radiography.This method belongs to a virtual dual-energy subtraction technique.In this article,we first introduce the basic neural network.The gradient maps of chest radiograph are set as the network’s input,while the gradient maps of chest radiograph minus tissue radiograph are set as the network’s label data,the network is trained by Adam algorithm.Aiming at the problem of long training time of convolutional network,by replacing the large convolution kernel into smaller one and decreasing the size of output features,we have successfully reduce the network’s training time and the running time.Based on the basic convolutional network,we have discussed the influence of different network parameters and network structures on the network’s accuracy and running time.The experimental results manifest that convolution network works well both on quantitatively metrics and visual perception.According to the shortage of basic fixed patch size network,we have trained a multi-resolution convolutional network,which integrates the features at different resolution,to avoid the problem of limited-information of fixed patch size.Compared to the basic convolution network,the multi-resolution convolution network has a significant improvement in quantitative metrics.更多还原