Browsing by Author "Şen, Sena Yağmur"

Now showing 1 - 1 of 1

Kalp aritmilerinin evrişimsel sinir ağları ve spektrogram tabanlı yöntemle sınıflandırılması
(2021) Şen, Sena Yağmur; Özkurt, Nalan
In this thesis, the main objective is to classify heart rhythms which were acquired from MIT-BIH Arrhythmia Database using Convolutional Neural Networks (CNN) architecture. The classified heartbeats are normal sinus rhythm, premature ventricular contraction (PVC), left bundle branch block (LBBB), and right bundle branch block (RBBB). Two main studies are described as follows; CNN Classification of Time-Series vs. Spectrogram Study and Spectrogram-CNN-Hyperparameter Tuning with Adam Optimization Algorithm Study. In CNN Classification of Time-Series vs. Spectrogram study, the classification of arrhythmias was carried out via using both raw signals and Short-Time Fourier Transform (STFT) in order to analyze the characteristic of the heartbeat signals. In time-domain classification, normal sinus rhythm, PVC and RBBB heart signals were used as a 1-D vector. The large number of electrocardiogram (ECG) time-series signals were classified with CNN by using their raw form. With the aid of STFT, Hamming window was applied in order to obtain the information from the signal. The normal sinus rhythm, PVC, and RBBB heart signals which are in the time domain were transformed into the time-frequency domain with the help of STFT. The STFT provided the acquisition of spectrograms from heart signals, and these spectrograms were classified with CNN through using as their RGB image form. The proposed CNN Classification of Time-Series vs. Spectrogram study demonstrated the high accomplishment rates in the deep learning approach, and according to accuracy, sensitivity and specificity terms, and also showed to better than traditional feature extraction methods. In Spectrogram-CNN-Hyperparameter Tuning with Adam Optimization Algorithm study, heartbeat signals were examined by tuning CNN hyperparameters. The ECG heart signals which are normal sinus rhythm, LBBB, and RBBB were transformed into their corresponding spectrograms in order to acquire characteristics of the heart signals. These spectrograms were restricted with a particular time/frequency resolution rate, and the suitable time/frequency resolution rate was identified heuristically. Adam was selected as an optimization algorithm of the deep learning network in order to train the ECG spectrograms. The tuned hyperparameters were the learning rate, gradient decay factor and squared gradient decay factor of the Adam algorithm. The identification of hyperparameters was performed by using the grid search method in order to compare the results. The effect of the tuning process according to learning rate and the moment estimation coefficients were represented as their validation loss graphs. The proposed Spectrogram-CNN-Hyperparameter Tuning with Adam Optimization Algorithm study yielded great achievement rates with respect to accuracy, sensitivity and specificity terms.