Unpaired Synthetic Image Generation in Radiology Using GANs

In this work, we investigate approaches to generating synthetic Computed Tomography (CT) images from the real Magnetic Resonance Imaging (MRI) data. Generating the radiological scans has grown in popularity in the recent years due to its promise to enable single-modality radiotherapy planning in clinical oncology, where the co-registration of the radiological modalities is cumbersome. We rely on the Generative Adversarial Network (GAN) models with cycle consistency which permit unpaired image-to-image translation between the modalities. We also introduce the perceptual loss function term and the coordinate convolutional layer to further enhance the quality of translated images. The Unsharp masking and the Super-Resolution GAN (SRGAN) were considered to improve the quality of synthetic images. The proposed architectures were trained on the unpaired MRI-CT data and then evaluated on the paired brain dataset. The resulting CT scans were generated with the mean absolute error (MAE), the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) scores of 60.83 HU, 17.21 dB, and 0.8, respectively. DualGAN with perceptual loss function term and coordinate convolutional layer proved to perform best. The MRI-CT translation approach holds potential to eliminate the need for the patients to undergo both examinations and to be clinically accepted as a new tool for radiotherapy planning.