基于深度学习的肝硬化识别研究

发布时间：2018-01-19 19:48

本文关键词： 改进LBP 卷积神经网络深度信念网支持向量机超限学习机　出处：《青岛大学》2017年硕士论文　论文类型：学位论文

【摘要】：B超图像中肝硬化的识别在临床上对肝病的诊断具有重要意义,但是临床采集的图像中有噪声、尺度不一、病变边缘模糊、回声不均匀等因素。传统的图像识别技术在对肝硬化进行识别时,准确率较低,主要是由于传统的特征提取算法不能很好地描述肝硬化病变的特点。针对上述问题,本文提出了一种改进的局部二值模式的特征提取算法,很好地解决了病变区域边缘模糊的问题;利用深度学习理论,设计了两种卷积神经网络模型和一种深度信念网络模型,得到了表现力更强的特征。本文主要工作有以下几个方面:(1)将局部二值模式(Local Binary Pattern,LBP)算法进行改进,并通过信息熵来度量通过该改进LBP算法得到的样本的信息量。选出包含纹理信息最多的样本,作为支持向量机(Support Vector Machine,SVM)和超限学习机(Extreme Learning Machine,ELM)的输入。实验结果表明,所提出的改进算法对于提高肝硬化的识别率很有帮助。(2)利用深度学习思想,设计了两种基于卷积神经网络(Convolutional Neural Networks)的模型,分别是Liver-CNN1和Liver-CNN2,并将其与SVM和ELM结合进行分类,替换了普通卷积神经网络最后的全连接神经网络,不仅取得了更高的分类准确率,而且在分类时间上也有明显的减少,具有实时性。卷积神经网络中感受野概念和权值共享技术的使用,大大减少了网络参数的数量,提高了学习的速度。经实验对比得知,卷积神经网络自学习得到的特征相对于传统的特征有更高的识别率,说明本文的两种卷积神经网络学习到的特征对肝硬化病变有更准确的描述。(3)利用深度信念网(Deep Belief Nets,DBNs)与SVM和ELM结合的分类方法,将一般的深度信念网的最顶层的逻辑回归,用SVM和ELM代替,省去了回归值与类别之间的度量过程。而且利用DBNs的预训练和分层学习,得到了更接近全局最优的解,网络收敛更快。
[Abstract]:The recognition of liver cirrhosis in B-ultrasound images is of great significance in the diagnosis of liver diseases. However, the images collected in the clinic have noise, different scales and blurred lesion margins. Traditional image recognition technology has low accuracy in the recognition of liver cirrhosis. The main reason is that the traditional feature extraction algorithm can not describe the characteristics of liver cirrhosis well. In view of the above problems, this paper proposes an improved feature extraction algorithm of local binary pattern. The fuzzy edge of the lesion area is well solved. Two kinds of convolution neural network models and a deep belief network model are designed by using the theory of deep learning. The main work of this paper is as follows: 1) improve the local binary mode (Local Binary pattern) algorithm. The information entropy is used to measure the information of the samples obtained by the improved LBP algorithm, and the samples with the most texture information are selected. As support Vector Machine (SVM) and extreme Learning Machine. The experimental results show that the improved algorithm is helpful to improve the recognition rate of liver cirrhosis. Two models based on Convolutional Neural Networks are designed. Liver-CNN1 and Liver-CNN2 were combined with SVM and ELM to replace the final fully connected neural network of the general convolutional neural network. It not only achieves higher classification accuracy, but also has a significant reduction in classification time, real-time. The concept of perception field and weight sharing technology are used in convolution neural network. It greatly reduces the number of network parameters and improves the speed of learning. The experimental results show that the characteristics obtained by convolution neural networks have a higher recognition rate than the traditional features. It is suggested that the features of the two convolutional neural networks in this paper are more accurate in describing the liver cirrhosis lesions.) Deep Belief Nets is used in this paper. DBNs) combined with SVM and ELM, the logical regression of the top layer of general deep belief net is replaced by SVM and ELM. The measurement process between regression value and category is omitted, and by using the pretraining and hierarchical learning of DBNs, the global optimal solution is obtained, and the convergence of the network is faster.
【学位授予单位】：青岛大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R575.2;TP18;TP391.41

【参考文献】