COMPARISON OF GENERATIVE ADVERSARIAL NETWORKS IN MEDICAL IMAGING APPLICATIONS - MR to CT image synthesis

Abstract: Cancer is one of the leading causes of death worldwide with about half of all cancer patients undergoing radiation therapy, either as a standalone treatmentor in combination with other methods such as chemotherapy. ‘The dose planning of radiation therapy is based on medical images such as CT and MR images. A recent trend is to move towards “MR-only” workƒows, and previous work have shown good results when synthesizing CT images from MR with machine learning methods. Eliminating the need for CT scans in the dose planning procedure removes a potentially carcinogenic part of the procedure, saves clinical resources, and could shorten the time until treatment can begin. Th‘is thesis builds upon earlier work, where a Cycle-Consistent Adversarial Network (CycleGAN) was used successfully to synthesize CT images from MR Iimages. We compare the CycleGAN architecture to the original Generative Adversarial Network (GAN), and a variant called Wasserstein GAN (WGAN). Th‘e UNetwas used as a generative sub-network for all models.Th‘e CycleGAN utilized a PatchGAN discriminative sub-network while the other models used a custom Convolutional Neural Network (CNN). Th‘e GAN architecture was tested both with paired and unpaired data. Th‘e best results were obtained by the GAN using unpaired data, with an MAE of 30:4 HU and PSNR of 26:7 dB. ‘The CycleGAN model failed to produce images what could pass as suitable synthetic CT images. We could not reproduce the results of earlier work in this regard while either using the hyperp arameters of previous studies or other con€gurations. But we conclude that it is possible to produce synthesized CT images using GANs with both paired and unpaired data.