An Improved Deep Learning Framework for MR-to-CT Image Synthesis with a New Hybrid Objective Function

There is an emerging interest in radiotherapy treatment planning that uses only magnetic resonance (MR) imaging. Cur-rent clinical workflows rely on computed tomography (CT) images for dose calculation and patient positioning, therefore synthetic CT images need to be derived from MR images. Re-cent efforts for MR-to-CT image synthesis have focused on unsupervised training for ease of data preparation. However, accuracy is more important than convenience. In this paper, we propose a deep learning framework for MR-to-CT image synthesis that is trained in a supervised manner. The pro-posed framework utilizes a new hybrid objective function to enforce visual realism, accurate electron density information, and structural consistency between the MR and CT image domains. Our experiments show that the proposed method (MAE of 68.22, PSNR of 22.28, and FID of 0.73) outperforms the existing unsupervised and supervised techniques in both quantitative and qualitative comparisons.