CT synthesis from MRI images based on deep learning methods for MRI-only radiotherapy

Abstract

MRI-only radiotherapy is expected to be safer and more precise compared with conventional CT-based radiotherapy. But unlike CT, MR images is not related with electron density for radiotherapy planning. To use MR images for treatment planning, we proposed a residual learning based u-shaped deep neural network (RUN) to synthesize CT images from MR images. The RUN network is an encoding-decoding network consisted of 16 residual learning based units which are constructed with parameter-free shortcuts to alleviate the degradation problem of deep network and strengthen feature propagation and reuse. We input MR and corresponding CT images to the RUN network to learn a voxel-by-voxel mapping from MR to CT images. Synthetic CT images are generated for newly input MR images through the learned mapping model. Our datasets contain 35 patients with more than 5000 T1/T2-weighted MR and CT images pairs. We compare Hounsfield Unit (HU) discrepancies between synthetic CT and original CT images. The mean absolute error (MAE) was $66.12 \pm 5.95$ HU, peak signal noise ratio (PSNR) was $28.52 \pm 0.64$ dB, and structural similarity index (SSIM) was $0.97 \pm 0.005$ for T1-weighted MR. The comparison results of the same metrics for T2-weighted MR were $63.79 \pm 4.18$ HU, $28.8 \pm 0.55$ dB and $0.973 \pm 0.004$ respectively. Converting CT images from MR images for one patient takes about 3 seconds on average. Experimental results show that the proposed RUN network is accurate, robust, and efficient for predicting synthetic CT from MR images for MRI-only radiotherapy.