AUTHOR=Li Yan , Xu Sisi , Lu Yao , Qi Zhenyu 

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

JOURNAL=Frontiers in Physics

VOLUME=Volume 11 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2023.1088899

DOI=10.3389/fphy.2023.1088899

ISSN=2296-424X

ABSTRACT=Synthesizing CT from MRI has drawn wide research interests aiming to expand the usage of MRI in the radiation therapy and substitute the function of CT, which could save time and financial expenses on scanning CT modality, prevent additional ionizing radiation to patients and avoid MRI-CT registration uncertainties to dose calculation procedure. Deep learning models have become the first choice for MRI-CT synthesis because of its ability to study complex non-linear relations. In order to accurately synthesize the intensity and structural properties of CT, some existing studies have introduced multi-scale learning framework. However, as they applied similar architectures to study the local and global relations and built pure convolution- or transformer- based multi-scale models, they did not utilize the strength of the other method, which could still limit the local or global performance of synthetic CT. To solve this problem, this study proposes a hybrid multi-scale model to explore rich local and global MRI-CT relations, named hybrid multi-scale synthesis network (HMSS-Net). In HMSS-Net, the learning of local and global MRI-CT relations is enhanced by convolution- and transformer- based designs, respectively. In low-resolution module, the method of transformer 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 high-resolution module, residual and dense connections are applied to build the its block for exploring complex local MRI-CT relations under multiple step sizes. Their feature spaces are combined together and utilized to provide synthetic CT. Meanwhile, HMSS-Net also introduces the multi-scale structural similarity index measure (MS-SSIM) loss to provide multi-scale supervision during training. The experimental results on head and neck region of 78 patients showed that HMSS-Net reduced the average of 7.6/3.13 HU on mean absolute error and increase the average of 2.1/1.8% on dice coefficient of bone compared with competing image-to-image synthesis methods. The results implies that HMSS-Net could provide reliable synthetic CT.