Research On The Algorithm Of Head And Neck MRI Image Synthesis CT Image Based On CycleGAN

Cancer mortality has surpassed cardiovascular mortality in some regions,posing a serious threat to human health.Intensity-Modulated Radiotherapy(IMRT)is an effective way to treat cancer.In conventional IMRT planning,a patient’sComputed Tomography(CT)and Magnetic Resonance(MR)images are taken.However,taking a CT can expose the patient to additional radiation,resulting in possible damage to the diseased area.It also increases the financial burden on the patient.Therefore,this paper investigates the synthetic CT(sCT)images from MR images based on deep learning to solve these problems.In MR-only radiotherapy planning,an improved Cycle-consistent Generative Adversarial Network(CycleGAN)is designed for the problems of reducing clinical data requirements and improving the accuracy of MR-based synthetic CT generation.Firstly,the MR data and CT data are pre-processed and a data cropping strategy is proposed to prepare the available data for subsequent network training.Secondly,a cycle-consistent structural control domain is proposed in terms of the loss function to constrain the structural consistency of the generator input and output.Meanwhile,the generator and discriminator of the CycleGAN are improved to achieve the purpose of balanced adversarial.And the ablation experiments are done to verify the effectiveness of the improvement by different evaluation metrics.Taking into full consideration the different focus of different MR sequence expressions,this paper verifies the accuracy of different individual MR sequences to generate sCT.The experimental results show that the sCT generated based on T1-weighted MR sequences is more accurate.To take full advantage of the complementary nature of MR sequences.In this paper,the improved CycleGAN is used to validate the accuracy of sCT generation based on hybrid sequence MR.The mixed MR sequences are used in the training phase of the model.During the testing phase of the model,it is further verified which MR sequences are more suitable for generating sCT.Five sCT generation strategies are proposed in this paper.The experimental results show that the accuracy of the sCT generated using mixed MR sequences during the training of the model and T1-weighted MR sequences based on T1-weighted MR sequences during the testing of the model is higher.