Biomedical Image Analysis using Deep Neural Networks

Abstract

The use of machine learning is very prevalent now-days and more new applications are continuously discovered. One of the prominent paradigms in computer vision are Convolutional Neural Networks (CNN). The purpose of this thesis is to introduce the topics of machine learning, convolutional neural networks and to test and evaluate experimental deep neural network architectures on functional Magnetic Resonance Images (fMRI). A series of various multi-layer CNN architectures with alternating hyper-parameters was tested against two well-known benchmark problems in the area of image classification: MNIST, CIFAR-10. The models' capacity was also evaluated against a real-world dataset of fMRI images. The network's model was rebuilt with each test run, rotating between the possible configurations. The proposed models, while performing relatively well on benchmark problems, were not able to surpass the current state of the art in brain image classification. To achieve possibly better results, they would need to be expanded to allow a broader set of features to be absorbed and classified. Also the limitations of the used hardware and the resulting impact were established. Based on the empirical results, it can be concluded that CNN are a viable tool for image pattern recognition.