In recent years, neural networks have been widely used as the method of choice for classification tasks in computer vision and image processing. The networks that currently provide the best results are very deep residual Convolutional Neural Networks. New network structures, based on the previously used topologies are developed each year, using different approaches. In this thesis the application of a canonical genetic algorithm to optimize the topology of neural networks is investigated. Current state of the art networks are used as a baseline to start the optimization form. The generated architectures achieve an improved classification accuracy, compared to reference networks. The developed network structures differ from commonly used architectures in their higher level architecture. In contrast to the reference networks they are not homogeneously from a sequence of different blocks. The improved classification accuracy suggests that the use of heterogeneous structures may be advantageous. Furthermore it is shown that the generated neural networks converge in fewer iterations, producing better results.