基于受限玻爾茲曼機的醫(yī)學圖像分類方法研究
發(fā)布時間：2018-09-11 10:05
【摘要】：隨著計算機技術(shù)的發(fā)展,越來越多的醫(yī)學圖像分析技術(shù)也應運而生。利用數(shù)據(jù)挖掘方法對醫(yī)學圖像做分析是目前研究的熱點之一,目前很多數(shù)據(jù)挖掘方法已經(jīng)成功應用在了醫(yī)學圖像的分類中,但這些方法中的很大一部分都是先提取醫(yī)學圖像中的統(tǒng)計學特征,然后再在此特征數(shù)據(jù)集基礎上對圖像進行分析,從而實現(xiàn)對醫(yī)學圖像的診斷。目前這種通過特征提取來對醫(yī)學圖像進行分析的方法主要有關(guān)聯(lián)規(guī)則、決策樹、遺傳算法、人工神經(jīng)網(wǎng)絡、貝葉斯網(wǎng)絡、粗糙集、支持向量機等方法。但是基于統(tǒng)計特征提取的醫(yī)學圖像分析方法中特征提取的好壞直接影響著圖像的分析結(jié)果,而且特征的選擇會受到經(jīng)驗等主觀因素的影響。目前比較流行的特征提取方法是深度學習方法,此方法利用深度信念網(wǎng)絡(Deep Belief Network,DBN)模型對輸入數(shù)據(jù)進行特征學習。DBN實際上是一個有向圖模型,它的基礎模型是無向圖模型受限玻爾茲曼機(Restricted Boltzmann Machine,RBM),由于在多層有向圖中,推斷隱層單元的后驗分布是相當困難的,于是DBN特征的學習采用每次學習一個RBM的方式。這是因為RBM二部圖結(jié)構(gòu)可以讓對隱層單元的狀態(tài)的推斷變得很簡單。每個RBM學習到的特征即其隱層單元的狀態(tài)將會作為下一層RBM的輸入數(shù)據(jù),以此類推,完成DBN的訓練。而且Hinton等人已經(jīng)證明每增加一層RBM,DBN的訓練數(shù)據(jù)的對數(shù)概率可變邊界就會降低,即DBN對訓練數(shù)據(jù)的表達能力就會更強。DBN因為其學習輸入數(shù)據(jù)中復雜的高層次的特征結(jié)構(gòu)的能力,已經(jīng)得到了廣泛的研究和應用。本文主要研究了DBN的基礎模型RBM,并從特征學習的角度,利用RBM針對現(xiàn)有的應用在醫(yī)學圖像分類上的方法對醫(yī)學圖像的診斷做了兩個方面的改進:1.提出利用RBM模型對醫(yī)學圖像進行特征學習的特征提取方法RBM是DBN模型的基礎模型,它有一個二部圖結(jié)構(gòu),是一個無向圖,RBM本身也是一種有效的特征提取器。本文利用機器學習模型受限玻爾茲曼機的特征學習能力對圖像的特征提取階段進行改進,然后再利用組合分類器:基于Bagging的概率神經(jīng)網(wǎng)絡對圖像進行分類。在乳腺X光圖像的標準數(shù)據(jù)集(MIAS)上的實驗結(jié)果表明:利用RBM學習到的特征進行的分類精度比起人工選擇的分類精度更高。2.提出利用DRBM模型對醫(yī)學圖像分類的圖像分類方法由于前面的改進存在一定程度的局限性,比如受限玻爾茲曼機學習到的特征可能并不適應于所有的分類器,所以本文又采用一種新的醫(yī)學圖像分析方法:判別式受限玻爾茲曼機(Discriminative Restricted Boltzmann Machine,DRBM)對醫(yī)學圖像進行分類分析。DRBM可以直接利用學習到的特征對圖像進行分類,避免了特征形式與要求的數(shù)據(jù)形式不匹配的問題。DRBM是一種無向判別模型,它可以自動的從圖像中學習特征,并利用學習到的特征直接對圖像進行分類。在乳腺X光圖像標準數(shù)據(jù)集上的實驗結(jié)果表明,DRBM對醫(yī)學圖像的分類準確率要好于基于Bagging的概率神經(jīng)網(wǎng)絡利用學習到的特征對圖像進行分類的效果。本文最后列出了目前醫(yī)學圖像分類研究中存在的一些問題以及今后需要進一步開展的研究工作。
[Abstract]:With the development of computer technology, more and more medical image analysis techniques have emerged. Data mining is one of the hotspots in medical image analysis. At present, many data mining methods have been successfully applied to medical image classification, but a large part of these methods are extracted first. At present, the main methods of analyzing medical images by feature extraction are association rules, decision trees, genetic algorithms, artificial neural networks, Bayesian networks, rough sets and support directions. However, the quality of feature extraction in medical image analysis method based on statistical feature extraction directly affects the result of image analysis, and the selection of features will be affected by subjective factors such as experience. In fact, DBN is a directed graph model. Its basic model is Restricted Boltzmann Machine (RBM). Because it is very difficult to infer the posterior distribution of hidden layer units in multi-layer directed graphs, the learning of DBN features is based on every one of them. This is because the RBM bipartite graph structure makes it easy to infer the state of hidden layer units. Each RBM learns that the state of its hidden layer units will be used as input data to the next layer of RBM, and by analogy, complete the training of DBN. Hinton et al. have demonstrated that each additional layer of RBM, DBN. DBN has been widely studied and applied because of its ability to learn complex high-level feature structures in input data. This paper mainly studies the basic model of DBN, RBM, and uses RBM from the point of view of feature learning. Two improvements have been made to the existing medical image classification methods: 1. RBM is the basic model of DBN model, which has a bipartite graph structure and is an undirected graph. RBM itself is an effective feature. 2. In this paper, the feature extraction stage of an image is improved by using the feature learning ability of a Boltzmann machine with machine learning model constraints, and then the image is classified by using a combined classifier: Bagging-based probabilistic neural network. The classification accuracy of the feature is higher than that of the manual selection. 2. The DRBM model is proposed to classify medical images because of the limitations of the previous improvements. For example, the features learned by the restricted Boltzmann machine may not be suitable for all classifiers, so this paper uses another one. A new method of medical image analysis: Discriminative Restricted Boltzmann Machine (DRBM) is used to classify and analyze medical images. DRBM can classify medical images directly by using the learned features, avoiding the problem that the feature form does not match the required data form. DRBM is an undirected discriminant. The experimental results on the mammogram standard dataset show that the classification accuracy of DRBM is better than that of Bagging-based probabilistic neural network. Finally, some problems in current medical image classification research and further research work in the future are listed.
【學位授予單位】：西北師范大學
【學位級別】：碩士
【學位授予年份】：2015
【分類號】：TP391.41
本文編號：2236378
[Abstract]:With the development of computer technology, more and more medical image analysis techniques have emerged. Data mining is one of the hotspots in medical image analysis. At present, many data mining methods have been successfully applied to medical image classification, but a large part of these methods are extracted first. At present, the main methods of analyzing medical images by feature extraction are association rules, decision trees, genetic algorithms, artificial neural networks, Bayesian networks, rough sets and support directions. However, the quality of feature extraction in medical image analysis method based on statistical feature extraction directly affects the result of image analysis, and the selection of features will be affected by subjective factors such as experience. In fact, DBN is a directed graph model. Its basic model is Restricted Boltzmann Machine (RBM). Because it is very difficult to infer the posterior distribution of hidden layer units in multi-layer directed graphs, the learning of DBN features is based on every one of them. This is because the RBM bipartite graph structure makes it easy to infer the state of hidden layer units. Each RBM learns that the state of its hidden layer units will be used as input data to the next layer of RBM, and by analogy, complete the training of DBN. Hinton et al. have demonstrated that each additional layer of RBM, DBN. DBN has been widely studied and applied because of its ability to learn complex high-level feature structures in input data. This paper mainly studies the basic model of DBN, RBM, and uses RBM from the point of view of feature learning. Two improvements have been made to the existing medical image classification methods: 1. RBM is the basic model of DBN model, which has a bipartite graph structure and is an undirected graph. RBM itself is an effective feature. 2. In this paper, the feature extraction stage of an image is improved by using the feature learning ability of a Boltzmann machine with machine learning model constraints, and then the image is classified by using a combined classifier: Bagging-based probabilistic neural network. The classification accuracy of the feature is higher than that of the manual selection. 2. The DRBM model is proposed to classify medical images because of the limitations of the previous improvements. For example, the features learned by the restricted Boltzmann machine may not be suitable for all classifiers, so this paper uses another one. A new method of medical image analysis: Discriminative Restricted Boltzmann Machine (DRBM) is used to classify and analyze medical images. DRBM can classify medical images directly by using the learned features, avoiding the problem that the feature form does not match the required data form. DRBM is an undirected discriminant. The experimental results on the mammogram standard dataset show that the classification accuracy of DRBM is better than that of Bagging-based probabilistic neural network. Finally, some problems in current medical image classification research and further research work in the future are listed.
【學位授予單位】：西北師范大學
【學位級別】：碩士
【學位授予年份】：2015
【分類號】：TP391.41
【參考文獻】
相關(guān)期刊論文 前4條
1 楊秋霞;;中值濾波與均值濾波的應用研究[J];內(nèi)江科技;2010年11期
2 張超;蔣宏傳;;舒怡乳腺診斷儀在乳腺癌診斷中的應用[J];中華腫瘤防治雜志;2010年19期
3 畢華;梁洪力;王玨;;重采樣方法與機器學習[J];計算機學報;2009年05期
4 周志華,陳世福;神經(jīng)網(wǎng)絡集成[J];計算機學報;2002年01期
，本文編號：2236378
本文鏈接：http://www.lk138.cn/shoufeilunwen/xixikjs/2236378.html
最近更新
教材專著
