|
|
Automatic 3D segmentation of brain images
Bafrnec, Matúš ; Dorazil, Jan (oponent) ; Kolařík, Martin (vedoucí práce)
This bachelor thesis describes the design and implementation of the system for automatic 3D segmentation of a brain based on convolutional neural networks. The first part is dedicated to a brief history of neural networks and a theoretical description of the functionality of convolutional neural networks. It represents a fast introduction to the problematics and provides theoretical basics needed for the understanding and creation of the system. Individual layers of the neural network and principles of their functionality and mutual relations are also described in this part. The second part of the thesis is about problem analysis, designing of a solution and a comparison between neural networks and other solutions. The result of a magnetic resonance imaging of the head is a series of black-and-white images representing a 3D scan. The task is to tag a brain and to remove unnecessary information in the form of surrounding tissues. The final image of the brain can be utilized in a volumetry or during a diagnostic of neurodegenerative diseases. The advantage of neural networks in comparison with deterministic systems is their flexibility. They allow an adaptation to other segmentation problems just by changing the training dataset, without a need of changes in the architecture. One of the systems performing fully automatic 3D segmentation is called U-Net – its name comes from the similarity of the architecture with the letter U. Three real solutions, the first implementation of U-Net, extended U-Net and recurrent U-Net were presented. The first version of U-Net has been very memory-demanding, it required a training on a processor instead of a graphic card and has not allowed data processing in full resolution. The extended U-Net has resolved these problems by loading data in overlaying series of three images. In addition to the possibility of a training on a graphic card with related decrease in learning time, the accuracy was increased by adding interconnections to the internal architecture of the network. The last version, recurrent U-Net, aims for the optimization of extended U-Net based on the reusage of existing levels. This brings a decrease in a time and resource difficulty. The number of parameters of the network was lowered to less than 20%, without any increase in case of further level addition. This network is one of first recurrent networks used on the problem of 3D segmentation and provides a foundation to further research. The last part focuses on the evaluation of results and the comparison of accuracy, speed and requirements between particular networks. The accuracy of human and machine segmentation is also compared. The extended and recurrent U-Net have surpassed their human opponent, which in real case could save a lot of doctors time and prevent human mistakes. The result of this work is a theoretical basis providing an introduction to the problematics of convolutional neural networks and segmentation, fully working systems for automatic 3D segmentation and the foundation for further research in the field of recurrent networks.
|
|
|
STATISTICAL LANGUAGE MODELS BASED ON NEURAL NETWORKS
Mikolov, Tomáš ; Zweig, Geoffrey (oponent) ; Hajič,, Jan (oponent) ; Černocký, Jan (vedoucí práce)
Statistical language models are crucial part of many successful applications, such as automatic speech recognition and statistical machine translation (for example well-known Google Translate). Traditional techniques for estimating these models are based on Ngram counts. Despite known weaknesses of N-grams and huge efforts of research communities across many fields (speech recognition, machine translation, neuroscience, artificial intelligence, natural language processing, data compression, psychology etc.), N-grams remained basically the state-of-the-art. The goal of this thesis is to present various architectures of language models that are based on artificial neural networks. Although these models are computationally more expensive than N-gram models, with the presented techniques it is possible to apply them to state-of-the-art systems efficiently. Achieved reductions of word error rate of speech recognition systems are up to 20%, against stateof-the-art N-gram model. The presented recurrent neural network based model achieves the best published performance on well-known Penn Treebank setup.
|
|
|
Automatic 3D segmentation of brain images
Bafrnec, Matúš ; Dorazil, Jan (oponent) ; Kolařík, Martin (vedoucí práce)
This bachelor thesis describes the design and implementation of the system for automatic 3D segmentation of a brain based on convolutional neural networks. The first part is dedicated to a brief history of neural networks and a theoretical description of the functionality of convolutional neural networks. It represents a fast introduction to the problematics and provides theoretical basics needed for the understanding and creation of the system. Individual layers of the neural network and principles of their functionality and mutual relations are also described in this part. The second part of the thesis is about problem analysis, designing of a solution and a comparison between neural networks and other solutions. The result of a magnetic resonance imaging of the head is a series of black-and-white images representing a 3D scan. The task is to tag a brain and to remove unnecessary information in the form of surrounding tissues. The final image of the brain can be utilized in a volumetry or during a diagnostic of neurodegenerative diseases. The advantage of neural networks in comparison with deterministic systems is their flexibility. They allow an adaptation to other segmentation problems just by changing the training dataset, without a need of changes in the architecture. One of the systems performing fully automatic 3D segmentation is called U-Net – its name comes from the similarity of the architecture with the letter U. Three real solutions, the first implementation of U-Net, extended U-Net and recurrent U-Net were presented. The first version of U-Net has been very memory-demanding, it required a training on a processor instead of a graphic card and has not allowed data processing in full resolution. The extended U-Net has resolved these problems by loading data in overlaying series of three images. In addition to the possibility of a training on a graphic card with related decrease in learning time, the accuracy was increased by adding interconnections to the internal architecture of the network. The last version, recurrent U-Net, aims for the optimization of extended U-Net based on the reusage of existing levels. This brings a decrease in a time and resource difficulty. The number of parameters of the network was lowered to less than 20%, without any increase in case of further level addition. This network is one of first recurrent networks used on the problem of 3D segmentation and provides a foundation to further research. The last part focuses on the evaluation of results and the comparison of accuracy, speed and requirements between particular networks. The accuracy of human and machine segmentation is also compared. The extended and recurrent U-Net have surpassed their human opponent, which in real case could save a lot of doctors time and prevent human mistakes. The result of this work is a theoretical basis providing an introduction to the problematics of convolutional neural networks and segmentation, fully working systems for automatic 3D segmentation and the foundation for further research in the field of recurrent networks.
|
|
|
STATISTICAL LANGUAGE MODELS BASED ON NEURAL NETWORKS
Mikolov, Tomáš ; Zweig, Geoffrey (oponent) ; Hajič,, Jan (oponent) ; Černocký, Jan (vedoucí práce)
Statistical language models are crucial part of many successful applications, such as automatic speech recognition and statistical machine translation (for example well-known Google Translate). Traditional techniques for estimating these models are based on Ngram counts. Despite known weaknesses of N-grams and huge efforts of research communities across many fields (speech recognition, machine translation, neuroscience, artificial intelligence, natural language processing, data compression, psychology etc.), N-grams remained basically the state-of-the-art. The goal of this thesis is to present various architectures of language models that are based on artificial neural networks. Although these models are computationally more expensive than N-gram models, with the presented techniques it is possible to apply them to state-of-the-art systems efficiently. Achieved reductions of word error rate of speech recognition systems are up to 20%, against stateof-the-art N-gram model. The presented recurrent neural network based model achieves the best published performance on well-known Penn Treebank setup.
|
host ::
RSS.