| Radiation therapy planning dominated by MRI needs to obtain CT images for dose calculation,but there is no direct relationship between pixels of MRI and those of CT.Patients have to suffer from high radiation and registration problems are involved in traditional methods,which are complicated.If CT images can be directly generated from MRI images,these problems can be avoided.The magnetic field strength of most MRI image scanners is 0.35 T,so the research on CT image generation is studied based on 0.35 T MRI images.Medical image modality conversion can be regarded as image style transfer.Due to the powerful generation ability and good interpretability of the flow-based model,this model is chosen as the main method.However,the flow-based model cannot be directly used.It needs to be improved.In the improved model,two flow-based models are connected by hidden variables,and each flow-based model can theoretically complete the mapping between pictures and hidden variables.The conversion between hidden variables can be completed by the convolutional neural network,like VGG16,so the model can achieve the CT generation with 0.35 T MRI images.In order to verify the effectiveness of the improved model,2554 pictures were used for model training,1112 for model validation,and 534 for model testing.Also,Unet and Cycle GAN were selected as the comparison models.Qualitative analysis found that the improved model has good image clarity and no defects are generated,which means the overall quality is optimal.Quantitative analysis shows that the MAE loss of the improved model is 37.67 HU,which is very close to 36.93 HU of the Unet model that uses pixel differences as the loss function.It is lower than the 45.92 HU of the Cycle GAN model,indicating that the difference between the picture generated by the improved model and the real picture is smaller at the pixel level.The PSNR value of the improved model is high,which is 24.30 d B,indicating that the picture generated by the improved model has a better reconstruction effect.The improved model has a higher NCC value of 0.967,which is close to 0.970 of the Unet model and higher than 0.953 of the Cycle GAN model,indicating that the pictures generated by the improved model are very similar to the real CT images.From the above results,it can be concluded that the improved model is effective,and the quality of the images it generates is higher than Unet and Cycle GAN.Therefore,the model can be applied to convert MRI images to CT images. |
