Research And Application Of Medical Image Processing Based On Deep Learning And Auto-encoding

In recent years,significantprogressof artificial intelligence hasbeenmadeinmanyfields,such as biomedicine,intelligent transportation,and smart home etc.,and medical image processing is one of the hot research directions in the field of biomedicine.There are new solutions to this problem with the rapid development of artificial intelligence technology,especially the rise of deep learning.Although with certain progress in applying deep learning technology to solve common medical image processing problems,there are still a number of challenges:1)Being time-consuming and expensive,the acquisition of certain images may even cause harm to human body,so how to avoid these limitations and obtain sufficient medical images to assist doctors in diagnosis? 2)In consideration of the indistinctness of current medical images generated by deep learning model,how to improve the current network model and bring forth high-quality medical images? 3)Demanding experimental data sets are necessary to some methods,such as paired matching data,well-labeled data,etc.,so how to improve the method to reduce the strong dependency of the network model on the data?4)"Stripe shadows"(artifacts)have frequently emerged in medical images generated by existing models,resulting in inferior quality of generated images.How to solve the problem of artifacts? In response to the above challenges,thefollowing research work has been carried out:First of all,for the generation of high-precision medical images,this paper proposes a deep model Res VNet&Patch GAN based on multi-level residuals and the theoretical conception of a deep learning network model.The generator of this model receives a random noise,it generates a complete image through this random noise,and the discriminator is used to judge whether the images are "true".Meanwhile,a multi-level residual and patch-GAN module is added to the model to strengthen information transmission efficiency of the Res VNet&Patch GAN model,for the sake that the quality of generated image is improved.In addition,the model further improves the accuracy of the generated image and the robustness of the model by random cropping and data enhancement.Finally,the detection time of the patient and the damage during the detection can be reduced with the medical images generated by a high-precision network model.Secondly,this paper proposes a network model based on transfer learning and deep learning(Dual3D&Patch GAN)in response to the harsh requirements of the data sets and the artifacts in generated images.Detailedly,this network model mainly uses the GAN,3D convolution,Patch-GAN and other modules,and the original two-dimensional network model is modified to three-dimensional one.The generator generates asmall three-dimensional blocks,as a resultthat the problem of striped "shadows" in the images generated by the two-dimensional network model can be solved.Since the Dual3D&Patch GAN model is a network model established on transfer learning,extremely high requirement on the data sets and image pairing are not indispensable,and selecting high-quality images in two domains can be more practical in clinical medicine.In terms of experimental results,Dual3D&Patch GAN precedes the results of other models.