Evaluation of Synthetic CT Image Generated Using a Neural Network for MR-Only Radiotherapy

Abstract

The MR-only radiotherapy workflow has been an active topic because of the increasing use of Magnetic Resonance (MR) image in the identification and delineation of tumors, while a fast generation of synthetic Computer Tomography (sCT) image from MR image is still one of the key challenges. This study aims to develop a neural network to generate the sCT and evaluate its performance in terms of both CT number difference and dosimetry accuracy. A generative adversarial network (GAN) is designed to translate MRI to sCT. For the generation of sCT, we use a "U-net” shaped encoder-decoder network with some image translation-specific modifications. Then a discriminator network is designed to distinguish between synthetic and real CT images. We enrolled 30 patients with brain tumors who have undergone external beam radiotherapy acquiring both CT and MRI for their simulation. 24 pairs of 2D T1-weighted turbo dark fluid MR images and rigidly registered CT are used to train the GAN, and the remaining 6 pairs are used to access the translation through the metric of mean absolute error. Furthermore, the dose distribution of the clinical VMAT treatment plan on sCT is calculated and compared with those on real CT using gamma analysis. On average, 5 and 10 seconds are needed to generate one sCT on GPU and CPU respectively. The mean absolute error between synthetic and real CT is 52±15 HU over six testing patients. For dose distribution on sCT and CT, the average gamma pass rates using the 3%,3mm and 2%,2mm criteria are above 97 and 92%, respectively. The GAN model is able to generate synthetic CT based on a single MRI sequence within seconds, and no significant radiotherapy dose difference is found between sCT-based and real CT-based plans.