Biyomedikal Görüntülerin Sınıflandırılması için Yeni Bir Evrişimli Sinir Ağı Mimarisi

Abstract

Convolutional neural networks (CNN) are a specialized version of deep learning and widely used among artificial intelligence algorithms recently. It can easily distinguish complex data such as biomedical images. In this study, a new CNN architecture which has 34 layers was proposed. This proposed architecture was named as OzNet. OzNet was applied for the classification of various biomedical data and obtained quite high performances. First, one-dimensional electrocardiogram (ECG) signal dataset was used in this study. The dataset was converted into two-dimensional images with continuous wavelet transform (CWT) and classified with OzNet. In addition, OzNet was compared to AlexNet and SqueezeNet and classification results displayed that OzNet was more effective than these architectures. In this study, it was combined with support vector machines (SVM) to increase the efficiency of performance. As a result, Oznet-SVM architecture obtained an accuracy of 99.21% for the classification of the ECG dataset. Next, coronavirus (COVID-19) computerized tomography (CT) images were used in this study. These images were classified with OzNet and pre-trained architectures using various preprocessing methods. According to the findings, discrete wavelet transforms (DWT) increased the classification performance of architectures. Eventually, DWT-OzNet architecture achieved an accuracy of 99.5% in the classification of COVID-19 CT images. Finally, CT images which were including brain hemorrhage types were applied in this study. OzNet also achieved high classification success on these images. However, it was combined with neighborhood component analysis (NCA) and machine learning algorithms to obtain more successful results and increase the reliability of the study. In this part, OzNet was employed as an automatic feature extractor and 4096 features were acquired from the fully connected layer for each image. These features were reduced by using NCA to get significant and informative features with minimal loss. Then, reduced features were classified with classifiers. According to the experimental results, OzNet-NCA-ANN structure obtained through artificial neural networks (ANN) was the best classifier model with 100% accuracy.