Research On Low-dose CT Image Noise Reduction Method Based On Improved Generative Adversarial Network

Computed tomography（CT）is an important medical diagnostic technique.High-dose Xrays will cause certain radiation damage to the human body during CT scanning.In order to reduce the damage of X-rays to the human body,low-dose CT（Low-dose CT,LDCT）is widely used in clinical diagnosis.Dosing at the same time introduces quantum noise into the projection data,resulting in reduced quality of the reconstructed CT images.Aiming at the image noise problem of low-dose CT,the traditional image noise reduction method cannot meet the actual needs due to the complicated noise reduction process or the defects of the algorithm itself.Based on the neural network method proposed in recent years,this paper uses a semi-supervised learning method to study the noise of LDCT images.The main work is as follows:（1）This paper proposes a low-dose CT image denoising method based on improved generative adversarial network,extracts image details from low-dose images,fuses features of different scales,and adopts batch normalization to improve network performance.The experimental results show that the proposed model effectively reduces the noise and artifacts of low-dose CT images,while retaining more detailed information and improving the image quality of low-dose CT images.（2）A network model based on residual module is established,three identical residual blocks are added to the original network,and the residual function after adding residual blocks is optimized.In the generation network part,deconvolution is used instead of the fully connected layer to reshape the denoised image,and the improved network is more stable in performance.The test data verifies the effectiveness of the network model.