Yapay sinir ağları ile doku sınıflandırma

Tissue classification using artificial neural networks

Tez No: 39106
Yazar: AYSU SEVEN
Danışmanlar: PROF.DR. ERTUĞRUL YAZGAN
Tez Türü: Yüksek Lisans
Konular: Elektrik ve Elektronik Mühendisliği, Electrical and Electronics Engineering
Anahtar Kelimeler: Belirtilmemiş.
Yıl: 1993
Dil: Türkçe
Üniversite: İstanbul Teknik Üniversitesi
Enstitü: Fen Bilimleri Enstitüsü
Ana Bilim Dalı: Belirtilmemiş.
Bilim Dalı: Belirtilmemiş.
Sayfa Sayısı: 137

Özet

ÖZET Doku sınıflandırma, hastalık teşhisinde etkin bir rol oynamaktadır, örneğin, beyindeki yumuşak doku, kemik ve tümörlü dokunun v.b. tanınması ve yerlerinin belirlenmesinde bilimsel bir yaklaşım kullanmak tıp doktorlarına büyük kolaylık ve güvenilirlik sağlar. Sinir sisteminin modellenmesi için yapılan çalışmalar sonucu oluşturulan Yapay Sinir Ağlan (YSA) modelleri doğal olarak biyolojik sinir sisteminin üstünlüklerine de olabildiğince sahiptir. YSA 1980'lerden sonra paralel işlem yapma, öğrenebilir bir yapıya sahip olma v.b. özelliklerinden dolayı yeniden gündeme gelmiş ve zaman içerisinde değişik problemleri çözmek üzere bir çok farklı yapay sinir ağı modelleri geliştirilmiştir. Bu çalışmada doku sınıflandırma için yapay sinir ağlan kullanılmıştır. Amaçlanan, sınıflandırmayı doğrudan görüntüden alman bilgilerle gerçekleştirmek ve yapay sinir ağlarının bu alandaki kullanımım araştırmaktır. İncelenen YSA modelleri Geriye Yayılma (BP), Sınırlı Coulomb Enerji (RCE), Vektör Kuvantalamayı öğrenme (LVQ), Kendini Düzenleyen özellik Haritası, Büyü ve öğren (GAL) modelleridir. Ağdan önce bir işlem yapmadan doku sınıflamaya gidilmesi, görüntü işleme konusuyla ilgisi olmayanların da YSA eğitiminde etkin olabilmelerim sağlar, örneğin tıp doktorları görüntü üzerinde basitçe seçecekleri uygun örnekler ile ağı sınıflama için eğitebilirler. Simulasyon bazında yapılan sınıflama çalışmalarında SUN IPX 32Mbyte iş istasyonu kullanılmış, sınıflama ilk önce bizzat oluşturulmuş bir görüntü üzerinde gerçekleştirilmiş, sonra bilgisayarlı tomografi ve nükleer manyetik rezonans ile elde edilmiş ve sayısallaştırılmış işleme hazır kafa kesiti görüntüsüne uygulanmıştır. -vı-

Özet (Çeviri)

SUMMARY TISSUE CLASSIFICATION USING ARTIFICIAL NEURAL NETWORKS The human information processing system consists of the biological brain. The basic building block of the nervous system is the neuron, the cell that communicates information to and from the various parts of the body. The neuron consists of a cell body called a soma, several spine like extensions of the body called dendrites, and a single nerve fiber called the axon that branches out from soma and connects to many other neurons. Inside and around the soma are ions including Sodium (Na+), Calcium (Ca++), Potassium (K+), and Chloride (C1-). The K+ concentrates inside the neuron and the Na+ concentrates outside. When the soma's membrane is electrically stimulated-usually by a voltage drop- its membrane allows the Na+ and other ions such as Ca++ to pass across its membrane and change the soma's internal state. The connections between neurons occur either on the cell body or on the dendrites at junctions called synapses. A helpful analogy is to view the axons and dendrites as insulated conductors at various impedance that transmit electrical signals to the neuron. The nervous system is constructed of billions of neurons with the axon from one neuron branching out and connecting to as many as 10,000 other neurons. All the neurons-interconnected by axons and dendrites that carry signals regulated by synapses-create a neural network. Seemingly the most challenging rival of the biological brain, the computer, processes an element of information as much as a million times faster. If the most advanced computers are able to process information one million times faster than the brain, why is the brain so superior for human information processing problems? The difference between the two can be traced to the processing order. The brain processes information in parallel and the computer processes information serially. -vn-Feldman (1985) has extended a constraint from this distinction called the 100-step program:“If the mind reacts in approximately half a second (500 miliseconds) to a given stimulus (i.e. answering a true-false question or naming a picture) and the cycle time of a neuron averages five milliseconds, then in 100 cycle times of a neuron a decision is reached”. If we use the computer analogy that one step of a program is processed for each time cycle, then the brain runs parallel programs that are only 100 steps long. In contrast to large software programs operating in serial on conventional computers, the brain operates with massively parallel programs that have comparatively few steps, possibly explaining why the brain is superior at human information processing problems despite being as many as 6 orders of magnitude slower. Neural networks, connectionist models, or neuromorphic systems are systems that are deliberately constructed to make use of some of the organizational principles that are claimed to be used in the human brain. An unusual characteristic of neural networking is its interdisciplinary nature. Neurophysiologists, psychologists, optical specialists, even philosophers are heavily interested in neural networks. Artifical neural networks, which have come into being as a result of the work on modelling the nervous systems naturally have the superior characteristics of the biological nervous system, as possible as it can. These advantageous characteristics are mainly as follows: o Parallelism In a neural net model each cell can be its own processor. There are no time dependencies among the connections in the same layer, they may operate completely synchronously. This property provides the artificial neural networks numerous advantages for processing speed and other features. o Fault-tolerance Damage (faults) to individual neurons can occur in the artificial neural network without a severe degradation of its overall performance. -vm-This graceful degradation is called fault-tolerance. In contrast, most conventional computers are not fault-tolerant, instead they are fault- intolerant. Removing any processing component of a conventional computer leads to an ineffective machine, and the corruption of a conventional computer's memory is irretrievable and leads to failure as well. o Sophistication of the model Neural net models usually have literally hundreds of factors at play, some of which may have only a small effect. But the aggregate effect of all of these input factors is a model that is likely to be much more accurate for difficult problems than any statistical model we might formulate. o Simplicity of realization As being distributed and massively parallel systems, artificial neural networks have the advantage of requiring to realize, mainly, simple processes and arithmetic operations instead of complex functions. This evidently, leads to simplicity in realization. o Ability to learn Since artificial neural networks are derived mainly from the work on modelling the nervous system including the biological brain, they naturally intend to solve any problem in the way the brain does, by learning. Due to these superior characteristics of the artificial neural networks, an explosion of interest by a lot of people working on various fields has been observed. One of the most notable of these fields is pattern recognition and classification. Classification has found a wide range of use in various fields such as medical image analysis, remote sensing, oil industry (for locating the reservoirs), etc. ?ix-Classification of tissues plays an important role for the diagnosis of disease in medicine. For instance, to recognize tissues in the brain, such as bone tissue, soft tissue, tumor, and to determine the location of them is a more scientific approach than the doctors determining them by naked eye. The 1980's have shown a renewed interest for artificial neural networks. They have been applied with various methods as the solution of various problems and so the span of interest have been widened. Having the natural ability of learning and the characteristics indicated above, artificial neural networks are used in the fields of classification and pattern recognition Work on classification using artificial neural net models can be viewed in two categories. The first approach is much like the conventional (non-neural network) classification methods. In conventional classification methods, the input data is first processed through a feature extraction block. Here, the classes from various regions of bubbles in the feature space. Then the output of this block is fed into the artificial neural network which produces output data relating the classification of the input space, in accordance to its input. This method is widely used. In the second approach to classification using neural networks, a feature extraction block is not included. Input data are directly taken from the input space, and fed into the artificial neural network. This method also has a wide range of use. The difference between these two approaches is that, if a reliable feature extraction is achieved, the first method requires a sample set (training set) which has smaller number of elements than the one required for the second method. However, it is the advantage of the second method realizing the whole process by the neural network without requiring a preprocessing block. So realizing the model as hardware becomes easier, because in the first method the functions used in the feature extraction block are not always likely to be realized in the hardware base. Even if they are, since they do not process in parallel, significant degradations in the processing rate are observed. -x-The classification of tissues using neural networks without a preprocessing feature extraction block is the objective of this thesis. This also is a kind of research work on the performance of neural net models on these applications, some of them being probably the first work on this type of application. Five different neural net models have been examined and applied to the images. In order to observe whether these various artificial neural networks were able to classify the tissues, first the ability of these networks to recognize the gray levels of simple two dimensional images having three or four gray levels, is examined. Then these models' ability of classifying various types of tissues such as soft tissue, bone tissue, brain liquid or tumor on a real brain image is discussed. The head images used are digitized CT (Computer Tomography), NMR (Nuclear Magnetic Resonance) images. The simulation work on classification have been carried out on SUN IPX 32 Mbyte RAM workstation. Tissue classification without using a preprocessing block provides the opportunity of being effective in the artificial neural network training for people of whom image processing is not a primary interest. For instance, medical doctors can train the neural network for classiffication by simply selecting a sample set from the image. In choosing the neural network models discussed and of which abilities to classify are examined, their suitability for image processing and having at least a few differences from each other have been taken into consideration. The neural net models used are Backpropagation (BP), Learning Vector Quantization (LVQ), Restricted Coulomb Energy (RCE), Grow and Learn (GAL), and Self Organizing Feature Map including its point of origin, Kohonen's Learning Law, and an enhanced mode of it for input probability density functions far from being uniform. Among these only the self organizing feature map is of unsupervised structure and have been used mostly in speech recognition applications till now. The significant distinction between these methods comes forth in the representation of space. The backpropagation method represents the space -xvby hyperplanes, whereas RCE represents the space by closed planes while other methods, used in this work, by dots. In order to put the work which is carried out by this thesis into a more understandable way a general information on artificial neural network concept and classification using artificial neural net models is presented in Chapter 2. In Chapter 3, Backpropagation neural network is discussed. The neural net model examined next, in Chapter 4, is the Restricted Coulomb Energy (RCE). Chapter 5 is the one which Learning Vector Quantization (LVQ) is studied. In Chapter 6, a work on Self Organizing Feature Map is carried out. The last neural network is Grow and Learn (GAL) which is discussed in Chapter 7. Besides, an example image, for didactic purposes, showing the ability of classification of the used neural net models, and the classification results are presented at the end of each chapter from Chapter 3 through 7. This didactic image example seen in Fig 3.7 sends two inputs into the artificial neural network. Hence, this part of the work can be called the classification in two dimensional sample space or input space. In Chapter 8, classification on digitized head images from a computer tomography and nuclear magnetic resonance images is presented and the results of classification is given. Chapter 9 consists of the comparison of the discussed artificial neural net models from the tissue classification viewpoint, and personal comments. -XII'

Benzer Tezler

Tez No
112235
Uydu görüntü verisinin yapay sinir ağları ile sınıflandırılması
Classification of satellite imagery data with artificial neural networks
COŞKUN ÖZKAN
Doktora
Türkçe
2001
Jeodezi ve Fotogrametri İstanbul Teknik Üniversitesi
Jeodezi ve Fotogrametri Mühendisliği Ana Bilim Dalı
DOÇ.DR. FİLİZ SUNAR ERBEK
Tez No
100840
Doku analizi sınıflandırma ve bölütleme için yöntemler
Techniques for texture analysis, segmentation and classification
OSMAN NURİ USLU
Doktora
Türkçe
2000
Uçak Mühendisliği İstanbul Teknik Üniversitesi
DOÇ. DR. B. GÜLTEKİN ÇETİNER
Tez No
335740
Yapay sinir ağları ve destek vektör makineleri ile kemik erimesinin teşhisi
Diagnosis of osteoporosis using artificial neural networks and support vector machines
MUSTAFA İSTANBULLU
Yüksek Lisans
Türkçe
2013
Biyomühendislik İstanbul Teknik Üniversitesi
Elektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
PROF. DR. SEDEF KENT
Tez No
250926
Sanat eserlerinin yapay sinir ağları ile özgünlük tespiti ve ayrıştırılması
Authentication and separation of art works with using artificial neural network
BAYBORA TEMEL
Doktora
Türkçe
2009
Güzel Sanatlar Marmara Üniversitesi
Güzel Sanatlar Eğitimi Ana Bilim Dalı
PROF. DR. BÜNYAMİN ÖZGÜLTEKİN
Tez No
213079
MS segmentation by using MRI images
MR görüntüleri kullanarak MS bölütlemesi
ŞULE ÜŞÜMEZOĞLU
Yüksek Lisans
İngilizce
2007
Elektrik ve Elektronik Mühendisliği Dokuz Eylül Üniversitesi
Elektrik-Elektronik Mühendisliği Ana Bilim Dalı
Y.DOÇ.DR. AHMET ÖZKURT

Geri Dön