Use of generative adversarial networks in medical image synthesis and segmentation

Generative Adversarial Networks (GANs) are one of the popular deep learning architectures in recent years. While performing image synthesis, images very close to real images were obtained. GANs have attracted the attention of many researchers in areas such as signal-to-image conversion, image synthesis and segmentation, resolution enhancement and completion of missing parts. It has been observed that it surpasses most methods in the field of computer vision. In this thesis, the most up-to-date approaches in areas such as creating medical datasets and image synthesis, segmentation and resolution enhancement on medical data were examined, and new GAN-based architectures were developed and contributions were made to the literature. In the first experimental study, the effect of classical GAN architectures CycleGAN and Pix2Pix methods on PAPSMEAR cell nuclei production was investigated. While comparing the performances of the architectures, Jaccard and Dice similarities were used. Image synthesis performances were examined separately and the results were presented in charts and visual forms. The second experimental study within the scope of the thesis is the use of mathematical modeling-based method to create synthetic PAPSMEAR dataset. The created method is named MathModel. In the first part of the third experimental study conducted within the scope of the thesis, new images were synthesized on the PAPSMEAR dataset by using classical MPA architectures (CycleGAN, Pix2Pix, DiscoGAN and AttentionGAN). The created images were compared according to MSE, SSIM and PSNR. Considering these outputs, a new GAN architecture called Pix2PixSSIM has been proposed. Obtained results are presented with charts and visuals. In the second part, the effect of classical and current comparison metrics on the PAPSMEAR dataset on the cost of GAN architectures is examined. The aim of this study is to discover a cost function suitable for general use in future GAN architectures. Results are presented visually and tabularly. In the fourth experimental study conducted within the scope of the thesis, classical GAN architectures (CycleGAN, CUT, FastCUT, DCLGAN, SimDCL) were used while performing image segmentation on the PAPSMEAR dataset. While comparing the method results, both visual and tabular results are included. LPIPS, PSNR, FID and KID similarity metrics were used when comparing the outputs. In the fifth experimental study within the scope of the thesis, classical GAN architectures (CycleGAN, CUT, FastCUT, DCLGAN, SimDCL) were used while performing image segmentation on the 2D brain MRI dataset, and a new GAN architecture called SSimDCL was proposed. It has been observed that the proposed method gives successful results both in increasing image resolution and in segmentation. LPIPS, PSNR, FID and KID methods were used when comparing the created images. In the last stage of the study, a visual and metric comparison was made with VolBrain results, which can now make good segmentation in brain MRI texturally. The sixth experimental study conducted within the scope of the thesis is the continuation of the fifth experimental study. Classical GAN architectures (CycleGAN, CUT, FastCUT, DCLGAN, SimDCL) and SSimDCL method were used while performing image segmentation on 2D Brain MRI, PAPSMEAR, CHASEDB and XRAY datasets. The aim is to examine the efficiency of the proposed method when segmentation on other medical images. Outputs are presented in a visual and tabular form. In the seventh experimental study conducted within the scope of the thesis, SSimDCL (recommended) from GAN architectures and nnU-Net method from deep learning architectures were compared in brain tumor segmentation. BraTs was used as the brain tumor dataset. In this study, datasets were improved and results were obtained over these datasets. Outputs are presented both visually and schematically. It is a continuation of the fourth experimental study. LPIPS and PSNR methods were used as similarity metrics. Outputs are presented visually and in tabular form. As a result of experiments and investigations, it has been observed that GAN-based approaches make positive contributions to many problem solutions.