CT synthesis from MRI with an improved multi-scale learning network

Introduction: Using MRI to synthesize CT and substitute its function in radiation therapy has drawn wide research interests. Currently, deep learning models have become the first choice for MRI-CT synthesis because of their ability to study complex non-linear relations. However, existing studies still lack the ability to learn complex local and global MRI-CT relations in the same time , which influences the intensity and structural performance of synthetic images.Methods: This study proposes a hybrid multi-scale model to explore rich local and global MRI-CT relations, relations, namely, the hybrid multi-scale synthesis network (HMSS-Net). It includes two modules modelling different resolution inputs. In the low-resolution module, the Transformer method is applied to build its bottleneck part to expand the receptive field and explore long-range MRI-CT relations to estimate the coarse distribution of widely spread tissues and large organs. In the high-resolution module, residual and dense connections are applied to explore complex local MRI-CT relations under multiple step sizes. Then, the feature spaces of two modules are combined together and utilized to provide synthetic CT. HMSS-Net also introduces the multi-scale structural similarity index measure loss to provide multi-scale supervision during training.Results: The experimental results on head and neck regions of 78 patients showed that HMSS-Net reduced the average of 7.6/3.13 HU on the mean absolute error and increased the average of 2.1/1.8% on the dice coefficient of bone compared with competing image-to-image synthesis methods.Conclusion: The results imply that HMSS-Net could effectively improve the intensity and structural performances of synthetic CT.