A Style-Based Model for MRI to CT Image Translation

The unsupervised generation of high-quality CT images remains an important and challenging problem. To obtain high-quality CT images, CT scans need to be performed on patients, which may result in radiation exposure and affect human health. In order to prioritize the health of patients, we attempt to estimate one modality of images from another modality, such as estimating CT images from MRI images. Therefore, we propose a novel framework for generating high-resolution CT images. Unlike previous encoder-based methods that can only predict latent codes from real images, our method maps the given medical images to the $\mathcal{W} + $ space and adjusts the generation details of medical images by modifying the variables in the $\mathcal{W} + $ space to ensure the generation of realistic high-resolution medical images. We generate a series of style vectors and then input them into a pre-trained StyleGAN to form an extended $\mathcal{W} + $ space. We assume that the latent vectors can be decomposed into content information and style information. To transform the images of MRI data into CT data, we combine the style information from the MRI images with the style information randomly sampled from the style space of the CT images, and finally generate realistic CT images by combining the content information from the CT images. We demonstrate a method for generating high-resolution medical images by manipulating latent style variables, which effectively yields a CT-like image without exposing the patient to radiation.