Generating Synthesized Computed Tomography (CT) from Magnetic Resonance Imaging Using Cycle-Consistent Generative Adversarial Network for Brain Tumor Radiation Therapy

Abstract

<h3>Purpose/Objective(s)</h3> Brain tumor is the most common malignant tumor of the head and neck in China, postoperative radiotherapy is one of the main methods to improve the survival rate of patients. Magnetic resonance imaging (MRI) has the advantage of its soft tissue contrast, for the sake of getting the tumor boundary accurately, it is necessary to combine CT/MR image fusing in gross tumor volume (GTV) delineation. In order to avoid the uncertainty caused by multimodal images registration and reduce unnecessary radiation dose, replacing CT with MRI has become one of the research hotspots in the field of radiotherapy. The aim of this study was to use unregistered MRI and CT images from brain tumor patients to generate pseudo-CT images based on a cycle-consistent generative adversarial network (CycleGAN) framework. <h3>Materials/Methods</h3> T1, T2-weighted MRI and CT-simulation images of the whole brain were collected from 31 brain tumor patients. In this work, we have used a CycleGAN framework to generate pseudo-CT images from MRI, this model is capable of image-to-image translation using unpaired MRI and CT images in an unsupervised learning method. Due to the influence of head frame in CT-simulation images and the different scanning range, imaging resolution and contrast of each patient, preprocessing of MRI and CT images, such as clipping, background removal, resampling and normalization, was required firstly. Secondly, in order to compensate for the insufficient to separate all major tissue types by the single MR sequence, each patient's T1/T2-weighted MR image pair was served as the input of the CycleGAN framework after registering and fusing. Finally, a simple cross-validation study, randomly selecting 70% samples as the training set and the remaining images as the testing set, was performed to compare the quality of synthetic CT and real CT image. <h3>Results</h3> The CycleGAN method produced the overall average MAE below 0.24 ± 0.02 and the 0.79 ± 0.03 SSIM value for the testing set images, the heterogeneity between the synthetic CT image and the actual CT image was acceptable. <h3>Conclusion</h3> We successfully realized the image-to-image translation using unregistered MRI and CT images based on the CycleGAN framework. The evaluation of pseudo-CT image showed the feasibility and accuracy of this method, which can effectively reduce the error caused by multi-mode image registration in gross tumor volume (GTV) delineation for brain tumor patients.