| Abstract [eng] |
This paper focuses on electroencephalograms (EEG) - the main tools in diagnosis and treatment of specific neurological disorders characterized by epileptic seizures. Definition of EEG signals, significance in medicine field, theory of their peaks and possibilities to improve processing using computers were analysed during this work and described in this paper. First part of this work is dedicated to implementation of algorithm based on series of mathematical morphological operations and filters. This algorithm was selected because it filters out background activity, distinguishes centrotemporal spikes and has high reliability according scientific literature. Eficient parameters of this algorithm were analysed and applied. Testing and verification of implemented algorithm were done with real patients' electroencephalograms already fully analysed by doctors. Second part of this work focuses on theory and implementation of artificial neural networks (ANN). Two types of ANN are detailed in this paper -- feedforward neural network and convolutional neural network. Practical tests using various parameters were investigated in order to achieve better results and precision of ANN. Also, both architectures of neural networks were compared to identify which one is more suitable for analysis of EEG. Moreover, an idea to combine algorithms mentioned in both parts - morphological filters and neural networks - was presented and analysed. On this purpose, a hybrid model for EEG peaks detection was implemented. Python programming language was used for implementation. Conclusions and observations are presented at the end of this paper. In general, experiments show that compared to algorithm based on morphological operations and filters ANN has achieved better results - approximately 90% of accuracy - in EEG peaks detection. However, results of hybrid model do not meet the expectations and no significant improvement was achieved. |
