Research On Synthesizing CT Images From MR Images Method Based On Generative Adversarial Network

Magnetic Resonance(MR)image-guided radiation therapy is an emerging radiotherapy technique that relies on the synthesis of Computed Tomography(CT)images from MR images to plan radiation therapy,not only to reduce the impact of cross-modal alignment errors on treatment outcomes but also to avoid the radiation damage caused during CT image acquisition.However,the different imaging mechanisms of MR images and CT images result in significant differences in the appearance of the tissues and organs they display and the medical information they present.Therefore,generating high-quality synthetic CT(SCT)images from MR images is a challenging task.In recent years,methods based on deep learning have shown excellent performance in the field of medical image synthesis,in which the Generative Adversarial Network(GAN)has been widely favored for MR image generation SCT image tasks due to its power generation capability.To generate high-quality SCT images,two GAN-based synthesis algorithms are proposed in this study,and they are applied to the single-sequence MR image generation CT image and doublesequence MR image generation CT image tasks,respectively.The main work of this study is as follows:(1)Currently,the GAN models used to generate SCT images from MR images are all based on convolutional neural networks(CNN).However,the CNN uses convolutional kernels to extract local features of the image,and the dependency between distant voxels cannot be fully utilized for tissues and organs distributed in a wide continuous spatial region.To address the above problems,this study proposes a Residual Transformer Conditional Generative Adversarial Network(RTCGAN)for generating SCT images from single sequence MR images.The model takes advantage of CNN in local texture details and Transformer in global correlation to extract multilevel features,and achieves effective aggregation of local texture features with global features through residual connectivity to obtain more characterization capability.In addition,the model introduces feature reconstruction loss to constrain the potential feature representation to mitigate the over-smoothing or local distortion of the generated images due to local mismatch between MR and CT images.In this study,the validity of the model was verified on pelvic and brain data.The experimental results show that the MAE of SCT images obtained by RTCGAN on the pelvic dataset is 45.05 HU,and the MAE of SCT images obtained on the brain dataset is 59.17 HU(MRT1 sequence)and 62.81 HU(MR-T2 sequence)compared to the real CT images.The synthetic accuracy of the proposed method was demonstrated by comparison with five medical image synthesis methods.The ablation experiments demonstrated the effectiveness of the constructed network structure for improving the quality of SCT images.(2)MR can provide multiple sequences of images,each of which presents its unique anatomical features.In clinical practice,doctors can obtain more comprehensive and accurate diagnostic information by comparing and examining MR images of different sequences for better treatment planning.To effectively utilize MR images of different sequences,this study proposes a Multi-feature Fusion Generative Adversarial Network(MFGAN)for generating SCT images from dual-sequence MR images.Two input channel branches with the same structure are constructed in MFGAN,and each branch corresponds to a different sequence image to obtain the unique features of each sequence image.Also,a reconstruction loss function is constructed on each input channel branch to improve the ability of the model to recognize valid information.In addition,the multifeature attention fusion module in MFGAN enhances the information related to the target task by learning the weights of each sequence image at different positions to efficiently fuse the features of different sequence images.To further improve the network performance,this study uses a multiscale discriminator to enhance the model’s recognition ability for different levels of information.The experimental results showed that the MAE of the SCT images of the brain obtained by MFGAN was 57.82 HU,which was better than the single sequence MR image synthesis CT image model.The ablation experiments demonstrated the effectiveness of the proposed module.