The purpose of this work is to develop a reliable deep-learning-based method that is capable of synthesizing needed CT from MRI for radiotherapy treatment planning. Simultaneously, we try to enhance the resolution of synthetic CT. We adopted pix2pix with a 3D framework, which is a conditional generative adversarial network, to map the MRI data domain into the CT data domain of our dataset. The original dataset contains paired MRI and CT images of 31 subjects; 26 pairs were used for model training and 5 were used for model validation. To identify the correctness of the synthetic CT of models, all of the synthetic CTs were calculated by the quantized image similarity formulas: cosine angle distance, Euclidean distance, mean square error, peak signal-to-noise ratio, and mean structural similarity. Two radiologists independently evaluated the satisfaction score, including spatial, detail, contrast, noise, and artifacts, for each imaging attribute. The mean (±standard deviation) of the structural similarity indices (CAD, L2 norm, MSE, PSNR, and MSSIM) between five real CT scans and the synthetic CT scans were 0.96 ± 0.015, 76.83 ± 12.06, 0.00118 ± 0.00037, 29.47 ± 1.35, and 0.84 ± 0.036, respectively. For synthetic CT, radiologists rated the results as evincing excellent satisfaction in spatial geometry and noise level, good satisfaction in contrast and artifacts, and fair imaging details. The similarity index and clinical evaluation results between synthetic CT and original CT guarantee the usability of the proposed method.
Keywords: MR-only simulation; attenuation correction; brain tumor; deep learning; generative adversarial net (GAN); radiotherapy planning.