Bridging the Gap Between Paired and Unpaired Medical Image Translation

Medical image translation has the potential to reduce imaging workload, by removing need capture some sequences, and annotation burden for developing machine learning methods. GANs have been used successfully translate images from one domain another, such as MR CT. At present, paired data (registered CT images) or extra supervision (e.g. segmentation masks) is needed learn good models. Registering multiple modalities annotating structures within each of them a tedious laborious task. Thus, there develop improved methods unpaired data. Here, we introduce modified pix2pix models tasks CT$\rightarrow$MR MR$\rightarrow$CT, trained with data, MRCAT pairs generated scans. The proposed modifications utilize ensure alignment between input translated images, MR$\rightarrow$CT model produces realistic-looking works well real input. variants outperform baseline pix2pix, pix2pixHD CycleGAN in terms FID KID, generate more realistic looking translations.