Research On Brain CT To MRI Image Generation Based On Pix2Pix Residual GAN

Computed Tomography(CT)and Magnetic Resonance Imaging(MRI)are two common techniques that capture images of human tissue in a noninvasive manner.CT has the advantages of fast and clear imaging,while MRI has no ionizing radiation damage to human body and the soft tissue imaging is clear.However,MRI scans take a long time,and the presence of metal implants in the patient’s body limit the use of MRI.Therefore,for some brain diseases,such as brain tumors and stroke,if effective information that is difficult to detect in CT images can be fully detected,and those information can be presented in a way that doctors are familiar with for auxiliary diagnosis,it will be of great significance to reduce the risk of missing the best treatment time.With deep learning technology,great breakthroughs have been made in cross-modal generation of medical images in recent years.However,cross-modal generation at present mostly use high information density image to generate low information density image.By comparing the information contained in CT and MRI,CT contains less effective information than MRI,so it is more difficult to achieve cross-modal image generation from CT to MRI.At present,Generative Adversarial Network has made good progress in image generation,this thesis mainly completes the following work:(1)This thesis designs a targeted method to preprocess the data to reduce interference because the data in the original dataset has some interference problems.In order to prevent the algorithm from over fitting,the preprocessed data is augmented and the amount of training data is enlarged.(2)For solving the problem that CT contains less effective information than MRI,which makes it difficult to generate cross-mode from CT to MRI,this thesis proposes an algorithm based on Pix2 Pix Residual Generative Adversarial Network(PRGAN).This algorithm can generate cross-modal images from brain CT to MRI,which can accurately generate images of brain skull,brain soft tissue and brain lesions.PRGAN combines Pix2 Pix network with residual block and is applied to cross-modal image generation after improvement.The generator can further extract high-level semantic features of CT images after introducing residual block,and PRGAN can also avoid gradient disappearance through residual block,so that the generated MRI has more details.Moreover,loss function is used to improve the generator’s ability of generating objects,which can enhance the stability of generative adversarial model.Compared with other algorithms,the algorithm in this thesis has been tested on open source data sets,which has achieved satisfactory results in quantitative analysis,as well as qualitative analysis by radiologists.(3)Based on the algorithm proposed in this thesis,a Web visualization system for brain CT to MRI cross-modal image generation is design.ed.It encapsulates the relevant algorithms so that users can achieve fast cross-modal image generation without understanding how the underlying algorithms are constructed.The PRGAN model proposed in this thesis,combining the advantages of Pix2 Pix network and residual network structure,verifies its wide applicability in Pix2 Pix network and its ease of use in medical image cross-modal.Not only that,It achieves cross-modal image generation from CT to MRI,and also achieves certain effects in quantitative and qualitative analysis.What’s more,a Web visualization system of brain CT to MRI cross-modal image generation is designed based on this model,which is helpful for auxiliary diagnosis and has practical value.