Investigation of the optimal data augmentation technique to improve liver disease diagnosis with a deep network architecture from computed tomography images

Recently, deep learning-based methods have been developed for the automatic diagnosis of liver disease. In image classification approaches with deep learning techniques, a large amount of data is needed to train and test network architectures. However, the number of images supplied in the medical field is insufficient. For this reason, various data augmentation methods are applied and network architectures are trained with the increased number of data. However, since the studies in the literature use different types of images, use different network architectures, train and test these architectures with different numbers of training and test data sets, it is not clear which data augmentation technique provides more successful results for which image type. Therefore, in this thesis, after analysis of the data augmentation techniques applied to liver images in the literature, the most effective data augmentation approach has been determined by comparing their effects on the classification of the images. In the studies carried out within the scope of this thesis, ten data augmentation techniques have been applied, training and test data sets have been created with the increased number of images obtained from each technique. A deep neural network architecture with convolutional and residual connections designed has been trained with ten training sets and then tested with test sets. Therefore, classification processes have been carried out as much as the number of data augmentation techniques applied, and the results of each classification process have been evaluated using five different numerical evaluation criteria. It has been determined that the data augmentation technique based on geometric transformations applied integratedly is the most appropriate technique for the classification of liver images taken by computed tomography with high performance. In addition, it has also been determined that providing data augmentation with salt-pepper type noise and shear is the approach that least improves the classification performance of the images