Research On Low-dose CT Image Enhancement Algorithm Based On Convolutional Neural Networks

Computed Tomography(CT)is an important and widely used method in clinical examination.However,X-ray radiation in CT scans may harm human health,thus,low-dose CT technology is advocated.Low-dose CT technology can effectively reduce the radiation dose that the patients needs to bear,but the noise and artifacts in the reconstructed image will increase significantly,which will affect the doctor’s clinical diagnosis.Therefore,how to improve the quality of low-dose CT images has become a key issue in the field of medical imaging.Conventional low-dose CT image processing methods can be divided into three categories:sinogram projection data domain filtering methods,statistical iterative reconstruction methods and image post-processing methods.Among them,the image post-processing method only processes the image domain data to get the results.It is easy to apply and has high research value.This paper uses deep learning related technologies,based on convolutional neural networks,to build a deep network to more accurately extract the learned image features,so as to more precisely suppress noise and artifacts,and restore clearer image details to achieve the purpose of improving the quality of low-dose CT images.The main research work of this paper is as follows:(1)Improvement on the deep convolutional network based on multi-scale feature perception module.The improved network is composed of multiple core multi-scale feature perception modules.The whole core module is divided into multiple branches,each branch perceives and extracts image features at different scales.The feature information extracted at the large scale will be transferred layer by layer to the smaller scale branches and fused to correct the feature extraction results of other branches,and construct complete and accurate feature information,so as to more accurately distinguish noise,artifacts,organs and tissues and deal with them accordingly.The experimental results on the clinical oral CT dataset show that the improved network has better suppression effects on noise and artifacts in low-dose CT images,and the detailed information in the restored images is more clearer and complete,and better results have been achieved.(2)Improvement on the deep cascaded cycle network based on cycle consistency.The whole improved network is a cyclic structure,with sub-networks of the same structure with different generation directions on both sides,and dual learning is formed between the two subnetworks on both sides to correct the learning of each other’s mapping function.In addition,a cycle consistency loss function was introduced in the network training to constrain the results generated by the cycle network and improve the accuracy of learning.In the experimental results on the clinical hip CT data set,the improved network is more effective in suppressing metal artifacts,the overall structure of the metal implant in the generated image is more complete and accurate,and the detailed information is clearer,indicating that the improved network has better performance.