Classification of biomedical signals by machine learning techniques

Early diagnosis of diseases is very important for the survival of living beings and improving their quality of life. Today, machine learning and deep learning algorithms that provide success in every field have become an important part of scientific research studies. It is important to reveal the features that are not visible to the eye in biomedical images and to improve the performance criteria of machine learning algorithms with appropriate parameters. In this thesis study, it is aimed to classify biomedical signals and improve their success in disease diagnosis by using machine learning algorithms. In this context, three different classes of lung X-ray images, four different classes of brain tomography images and ten different classes of heart sound signal data sets taken from the stethoscope have been used. In the studies carried out within the context of this thesis;  Deep learning methods are applied to three-class lung X-ray images for training. In order to improve the performance criteria, iterative neighbor component analysis (INCA), iterative ReliefF and local binary model (LBM) feature selection methods are applied to the trained data set.  Data augmentation and normalization preprocessing are applied to the four-class brain tomography image dataset. Extraction deep feature is performed on the preprocessed dataset using the restricted Boltzmann machine (KBM) method. These deep feature data are given as input data to the generating unit of the generative adversarial network (GAN) method.  A UDA architecture is designed suitable for the ten-class heart sound dataset taken from the stethoscope. The ReliefF feature selection method is applied to this dataset. In addition, the whale optimization method is applied to the UDA parameters. The superiority of these experimental studies is compared with other studies in the literature in terms of performance criteria.