MR To CT Synthesis Based On Cycle-Consistent Generative Adversarial Networks

Cancers are common in modern society,and the onset population is becoming younger.Radiation therapy is very effective for the treatment of many malignant tumors,such as head and neck cancer,and radiotherapy requires computerized tomography(CT)of the patient 's lesion images to plan the radiation dose,and also need to take magnetic resonance(MR)images of corresponding spatial structures to segment tumor tissue and healthy organs.Cancer patients need to take MR and CT images separately,which will cause higher social and economic burden and low efficiency.Moreover,when the CT image is taken,the patient is exposed to radioactive ionizing radiation,causing secondary injury to the patient.Therefore,this thesis first made two improvements to the Cycle-consistent Generative Adversarial Networks(CycleGAN)algorithm,and then proposed an algorithm for synthesizing CT images from MR images based on cyclically generated confrontational networks in order to solve these problems.First of all,two improvements were made to the original CycleGAN in this thesis.The first point is for the loss function,and the L1 distance function was added to improve the synthetic ability of the CycleGAN.Then,through data augmentation,a large amount of medical image data that is difficult to obtain in practice was obtained,and the capacity of the training data set was improved.Experimental results show that the mean absolute error(MAE)value of the CT image synthesized by the original CycleGAN method was 105.1 ± 4.4 HU,and the peak signal-to-noise ratio(PSNR)was 44.5 ± 0.8 dB;the mean absolute error value of the improved CycleGAN method was 94.0 ± 3.5 HU,and the peak signal-to-noise ratio was 45.1 ± 0.8 dB.Compared with the original CycleGAN method,the improved CycleGAN method reduced the mean absolute error of the synthesized CT image by 11.1HU and the peak signal-to-noise ratio was increased by 0.6dB,which was statistically significant proved by a paired ttest.Next,a U-Net CycleGAN method is proposedusing an improved U-Net to replace the residual network(ResNet)as the improved CycleGAN generator in this thesis.Experimental results show that the mean absolute error value of the CT image synthesized by the U-Net CycleGAN method was 76.7 ± 3.0 HU,and the peak signalto-noise ratio was 46.1 ± 0.9 dB.Compared with the improved CycleGAN method,the U-Net CycleGAN method reduces the mean absolute error of the synthesized CT image by 17.3HU,and the peak signal-to-noise ratio was increased by 1.0dB,which is statistically significant proved by a paired t-test.In summary,the U-Net CycleGAN method proposed in this thesis has advantages in accuracy and robustness of synthetic images,which proves that it is a more effective medical image synthesis method than the mainstream CycleGAN method.