| Abstract [eng] |
Intracerebral aneurysm is one of the most common causes of subarachnoid hemorrhage, also known as a stroke, when the blood spills into subarachnoid space. Whereas this disease often ends up in death while the majority of surviving patients stay disabled, it is important to regularly perform analysis of patients in order to notice developing brain aneurysms on time and to monitor their change. In order to lighten and automate diagnostics of intracerebral aneurysms, deep learning methods of convolutional neural networks were used in this scientific work. The aim of this master's thesis is to create and analyze models based on deep neural networks. These models are intended for classification and segmentation of head volume images which are taken during magnetic resonance angiography procedure. For image classification whether an aneurysm is visible or not, classic convolutional neural networks and ResNet-34 network architecture were used. Three different U-Net architectures (U-Net-15, U-Net-27, U-Net-36) were used for segmentation of intracerebral aneurysms. After examining the results, it was noticed that classic convolutional neural networks classify images, which are taken during magnetic resonance angiography procedure, better than ResNet-34. The largest U-Net architecture (U-Net-36) that was used in this research is best suited for segmentation of intracerebral aneurysms. However, even though networks that were used in this research were able to detect aneurysms most of the time, they were segmented larger than they actually are. The project consists of three parts: literature review, data and research methods, the results and summary of the research. In the part of literature review, detection, classification and segmentation problem of intracerebral aneurysms is reviewed. In the part of data and research methods, images of magnetic resonance angiography procedure and their processing are being reviewed, as well as network architectures that were used in the research. In the part of results and summary, results which were obtained by using convoluted neural networks aneurysms in order to solve the problem of classification and segmentation are being described. |
