MRI To CT Image Translation Based On Multi-Scale Cycle-Consistent Adversarial Networks

Radiation therapy is a powerful means to cure cancer,and doctors’ accurate diagnosis relies on the comprehensive information provided by medical images of multiple modalities(MRI,CT images,etc.).CT images are necessary for dose formulation,but also radioactive,and in some cases CT images are not of high diagnostic value.Strict registration of different medical images is also complicated and time-consuming.Therefore,accurate and efficient conversion from MRI images to CT images can reduce the economic and health burden of patients and save doctors’ time and energy,which is of great significance.At present,most of the existing methods are based on deep learning methods to achieve MRI-CT image conversion.Supervised methods use regression loss and adversarial training to generate higher-quality images.However,due to their dependence on registered medical data,unsupervised methods have gradually increased in recent years.Under unsupervised conditions,many related works have proposed a series of improvement strategies for the problem of medical image conversion based.However,existing methods have the problem that the quality of the generated image is not fine enough,and there is invalid confrontation training.Therefore,this paper aims at the above problems and proposes the main contribution points as follows:(1)Multi-scale discriminator structure,so that the discriminator can provide different scale discrimination output from pixel level to image level,with multi-scale receptive field,and pass the gradient back to generator for training,so that it can generate images with high quality.(2)Image context loss function based on similar part selection strategy,select images with similar medical structures during training,and use the histogram distance to constrain the context features between real images and generated images,thereby effectively avoiding too invalid training.It enables the generator to learn the medical image conversion of various different structures more effectively,and improves the quality of the generated image.In this paper,a series of evaluation indexes of comparison experiments and ablation experiments,as well as visualization results,show that the proposed method can improve the image quality and realize the purpose of accurate and efficient MRI-CT image translation.