A new extreme learning machine auto encoder design for medical datasets

The high classification and prediction performance seen in deep learning architectures recently increased the interest in this field. The widespread use of computer aided systems in the medical field has been important in the diagnosis of the disease, but the problem of the amount of data has come to the fore. The fact that the data sets collected to create disease classes do not have enough data or have unbalanced class distributions can be a problem in terms of the performance of the deep architectures used. However, increasing the desired dataset size is often not possible due to cost, time consuming, geographical conditions, unavailability of the relevant specialist or the equipment to be used, patient privacy and the rarity of some diseases. Data augmenting approaches come into play at this point. In this thesis, an Extreme Learning Machine Auto-Encoder-based data augmentation method with Wavelet function, called ELM-W-AE, is proposed. The proposed method is GAMEEMO dataset consisting of EEG signals for emotion recognition, EEG dataset for autism spectrum disorder diagnosis, EEG dataset for schizophrenia diagnosis, ICBHI breath sound dataset consisting of lung sound files for detecting lung diseases, and lung cancer detection for lung cancer detection. It has been tested within the scope of separate studies on a dataset consisting of computed tomography images. The tested method was compared with the original data set used in the study, the augmented states of these data sets and other recent studies. Comparisons show that the proposed method outperforms the compared methods.