Research On Low-dose Liver Dualsource CT Denoising Algorithm

With the development of deep learning,the field of medical imaging has also begun to use deep learning for research.The main research problem in this article is to use deep learning methods to solve the low-dose CT denoising problem.First of all,because the examination of the upper abdomen requires multiple enhanced scans,the radiation that the patient needs to receive will increase exponentially compared to other parts.In order to reduce the radiation caused by CT to patients,low-dose CT came into being.Although low-dose CT reduces the radiation to patients,it also brings some problems,reduced radiation leads to more noise,these noises are generally composed of Gaussian noise and Poisson noise.These noises will interfere with the doctor’s judgment,it affects the diagnosis efficiency of doctors,and affects the diagnosis results of doctors,cause very serious consequences.To solve this problem,this paper uses deep learning methods to denoise low-dose CT and assist the doctor’s diagnosis.The first is the collection of data sets,I prepared two data sets.One of them is the Mayo dataset of the “Low-Dose CT Challenge” provided by the American Medical Association in 2016,The other is a data set provided by the Second Affiliated Hospital of Mudanjiang.This paper uses the same method for these two different data sets.Using the Mayo dataset to verify the feasibility and performance of the method,at the same time use the clinical data of the hospital to test the real effect of the algorithm.(The Mayo dataset is not real data).Secondly,this paper proposes an unsupervised denoising method refers to the architecture of CYCLE-GAN.At the same time,the U-net structure is used to extract multi-scale features from medical images.The Attention mechanism is used for feature fusion,and the residual network is used to transform the features.It also considers the comparison of GAN network models and introduces perceptual loss and other technologies to improve the network in a targeted manner.At the same time,in view of the three-dimensional characteristics of CT images,this paper attempts to use transfer learning to use three-dimensional information.After training the 2D network,keep the 2D network parameters unchanged,change the network to a 3D structure,and then move the parameters.Input multiple slices at the same time during training to use 3D information to improve the experimental effect.Finally,a large number of comparative experiments are designed to verify the method in this paper,and the evaluation standards in the image field and the medical field are also used.Compared with the classic method,this paper uses unsupervised learning to solve the problem that real data cannot use supervised learning for effective learning.Because the network structure considers the particularity of the application,compared with the classical method,the experimental results of this unsupervised method still have better results.After the professional evaluation of the imaging doctors in the hospital,this algorithm basically meets the standards of complementary medicine.