| Summary: | Об’єктом розгляду є діагностика хвороби Альцгеймера. Предмет роботи –застосування алгоритмів машинного навчання в класифікації зображень, використання штучний нейронних мереж в діагностиці хвороби Альцгеймера. Метою роботи є розробка на основі штучних нейронних мереж системи автоматизованої діагностики хвороби Альцгеймера на основі знімків МРТ. Відсутність ефективних методів діагностики і лікування нейродегенеративних захворювань, поряд з постійним збільшенням їх кількості у світовій популяції, призводить щорічно до величезних економічних втрат. Методи машинного навчання, які демонструють величезні успіхи в різних областях, є одними з головних кандидатів для революції в сфері діагностики. В роботі, послідовно, проведено розгляд основних підходів діагностики, основного провокуючого фактору розвитку деменції - хвороби Альцгеймера, з акцентом на використання знімків МРТ головного мозку людини і методів глибокого навчання для їх аналізу. Зроблено висновок про відсутність знань першопричини цієї хвороби, як головної складності її діагностики. Окремо, викладені основні концепції машинного навчання та глибокого навчання, зокрема. Була розроблена система автоматичної діагностики хвороби Альцгеймера на основі знімків МРТ для виділення трьох класів пацієнтів: здорові, з легкої деменцією через хворобу Альцгеймера, з хворобою Альцгеймера. Система була побудована на основі нейромережевої моделі ResNet50V2. The diagnosis is based on the diagnosis of Alzheimer's disease. The subject of work - the current state of diagnosis of Alzheimer's disease, the use of machine learning algorithms in the classification of images, the use of artificial neural networks in the diagnosis of Alzheimer's disease. The aim of the work is to develop on the basis of artificial neural networks a system of automated diagnosis of Alzheimer's disease based on MRI images. Deep learning reached unprecedented heights in tasks that were considered subject only to man. Not so long ago it was impossible to imagine that a computer could surpass a person in the accuracy of image classification. But this happened in 2014, when the neural network won, then a student at Stanford University, Andrei Karpaty, in a competition to recognize 200 dog breeds. But the triumphal procession of neural networks did not end there. And already in 2017, the neural network defeated the man again, this time she won two games, out of the planned three, the strongest player in Go - Ke Jie. And these are just a few areas of activity where a person cedes his superiority to neural networks. Now the scope of neural networks is unusually wide: from the classification of images to predicting the rate of melting of ice masses in Antarctica, from the generation of automatic responses on the Internet to predicting the evolution of galaxies. Therefore, it is not surprising that now technologies based on neural networks are closely integrated into human everyday life. Sometimes the person himself is not aware of this. But although neural networks have already reached unprecedented heights, their application in the healthcare sector is still extremely insignificant. This happens due to legal restrictions, a biased attitude towards systems whose internal content is not known, and because of inertia, underfunding in some countries of the healthcare system, and in some cases, the lack of convincing results of neural networks in many areas. But even the last circumstance is not critical for many areas, since in principle there are no other effective methods. In particular, such an area is the diagnosis of Alzheimer's disease, where existing methods can determine its presence only in the late stages, when existing treatment strategies turn out to be powerless. Therefore, an attempt to introduce a method that will allow you to gain time in the disease for the patient in the conditions, as mentioned above, of the absence of other alternatives, is, as they say, a win-win option. Therefore, in the thesis, a comprehensive review of the diagnosis of Alzheimer's disease was carried out, with an emphasis on the use of neural networks in the diagnosis of Alzheimer's. The quintessence of work was the creation, on the basis of preliminary studies both in the field of neural networks and the use of neural networks in the diagnosis of Alzheimer's disease, an automated system for diagnosing Alzheimer's disease. For a clear and comprehensive coverage of the state of diagnosis of Alzheimer's disease and the existing capabilities of piece neural networks, the structure of the thesis was chosen, which consists of seven sections. In each of them, all aspects of the current state of diagnosis of Alzheimer's disease, the concept of neural networks, in particular, and machine learning, in general, were consistently presented. In the first section, entitled “Introduction”, the justification of the growing social importance of the problem of Alzheimer's was carried out through the prism of economic litigation, which every year is carried by the global economies and budgets of specific families in which there are people with Alzheimer's disease. Objective indicators of the quality of life faced by people with Alzheimer's disease were given. The forecast was made on the basis of open studies of the growth of the economic burden on the world economies if the current situation in the sphere does not change and Alzheimer's disease is not diagnosed in the early stages or effective drugs are not invented to combat it. In general, in the introductory part, the need is formulated to create an automated system for the early diagnosis of Alzheimer's disease. The second part of the thesis was an introduction to the basic concepts of neurodegenerative diseases and Alzheimer's disease, in particular. The difference between dementia and Alzheimer's disease has been carefully described, due to the fundamental difference between these concepts and the fact that in an environment remote from the medical, these concepts can often be identified or confused with each other, which in the end result can lead to incorrect interpretation of some research results. and possibly misinterpretation of scientific results. What can cause unintentional falsification of scientific results. An existing algorithm for diagnosing Alzheimer's disease was also described, where it was noted that there are no reliable methods that would determine exactly Alzheimer's disease. But there are only a number of methods that, with varying degrees of reliability, should, as they say, give indirect evidence, in this case, for other diseases, to establish Alzheimer's disease. The emphasis is placed on the technological backwardness of the methods for diagnosing Alzheimer's disease, in comparison with modern means available to medicine. The fact of insufficient efficiency of using modern technical means, such as magnetic resonance imaging and positron emission tomography, was also noted. The stages of development of Alzheimer's disease are described. At the end of the section, a conclusion was drawn, in which, based on the section, the main difficulties associated with the diagnosis of Alzheimer's were summarized. Thus, the introduction makes a complete introduction to the basic concepts and formulates the necessary terminological basis for the diagnosis of Alzheimer's disease. A section called "Machine Learning" is intended to give a soft introduction to machine learning, to introduce its basic concepts. It describes the history of the creation of machine learning algorithms from first ideas to algorithms running on supercomputers. In parallel with this, the main tasks are introduced, which should be solved by machine learning algorithms. A formal statement of the problem of machine learning. A soft introduction is made that machine learning is nothing more than the task of approximating the objective function, with the problems inherent in the problem of approximating the objective function. Accordingly, the analysis of methods for approximating the objective functions is carried out. In the course of which the problem of retraining machine learning systems and methods of dealing with this phenomenon are described. The section "Piece neural networks", as well as the section "Machine learning", describes the history of the creation of the science of piece neural networks, at the same time introduces and describes the basic concepts of piece neural networks, such as: piece neuron, activation function, and others. A comparative analysis of a biological neuron and a piece neuron is carried out. The current state of the field of deep learning is described in detail, various kinds of information are provided on the scope of the use of piece neural networks, on the competition of piece neural networks with humans, on the market for the use of piece neural networks and its growth. A full description of the first algorithms built on piece neural networks is made, in particular, a detailed description of one of the first neural network algorithms, the Rosenblatt perceptron, is carried out. The two most used architectures for image analysis are introduced, namely, a fully connected piecewise neural network and a convolutional piecewise neural network. After this, a complete derivation of the most commonly used method in the training of neural networks is done - the method of back propagation of errors. In the section "Use of neural networks in the diagnosis of Alzheimer's disease" a detailed review of the currently existing methods is carried out, with a description of the procedure for working with them, the use of neural networks in the diagnosis of Alzheimer's disease. The description of all the methods of using piece neural networks in the diagnosis of Alzheimer's disease provides links to scientific articles that reveal the essence of each of them at the level of the experiment. The emphasis is on the advantages and disadvantages of using convolutional architectures such as: VGGNet, ResNet, ResNet50V2 and so on. The main obstacles for creating an effective diagnostic system based on piece neural networks are formulated. In the next section, entitled “Description of the developed scheme”, in the first part, the rationale and a detailed description of the selected model for creating an automated diagnosis of Alzheimer's disease are made. The second part describes in detail, based on the experiments, the functionality of the developed system for the automatic diagnosis of Alzheimer's disease, namely, to evaluate the accuracy and stability of the system for the automatic diagnosis of Alzheimer's disease. The final part of the thesis entitled “Conclusions” summarizes the work done, gives the main functional characteristics of the developed system for the automatic diagnosis of Alzheimer's disease, compares the developed system for the automatic diagnosis of Alzheimer's disease with other automatic systems for the diagnosis of Alzheimer's disease, describes the main ways to improve the developed system for the automatic diagnosis of Alzheimer's disease and a conclusion is drawn on the practical value of the developed system for diagnosing Alzheimer's disease.
|
|---|