Research On Noise Reduction In Low-Dose CT Based On Deep Convolutional Neural Networks

With the continuous development of computed tomography(CT)technology and its important role in clinically assisted diagnosis,since the invention of the first CT scanner by Godfrey Hounsfield in 1973,CT has become more and more widely used.Since the generation of CT images is related to the absorption and transmission of different amounts of X-rays in different tissues of the human body,this poses a potential risk to genetic damage,cancer and other tumors.In order to solve this problem,it is a highly feasible means to reduce the radiation dose of the CT scans by reducing the tube current and other methods.However,the reduction in radiation dose increases the noise and artifacts of the image,which is unacceptable for clinical diagnosis.In recent years,in order to reduce the radiation dose of CT scans while ensuring the diagnostic utility of low-dose CT images,researchers have done a lot of related work.The traditional image processing methods include sinogram filtration before reconstruction,iterative reconstruction and image post-processing after reconstruction.Many breakthrough results have been achieved.However,the processing time of low-dose CT images is too long,and there may be problems such as edge blurring.In terms of CT scanning needs,further improvement is needed.With the significant improvement of computer computing power,allowing a large amount of data to be processed in deep networks,the deep learning-based method has an explosive development in the field of medical image processing.These methods have achieved good results in low-dose CT image noise reduction and show great potential for development.And trained models spend much less time processing images in the application phase than traditional methods.In the study of noise reduction of low-dose CT images,in order to reduce the noise and artifacts present in low-dose CT images,and at the same time make the processed images have high-quality processing results in subjective visual effects of the noise reduction,soft tissue texture preservation and small structure fidelity,and obtain high scores in quantitative analysis of objective indicators,a low-dose CT image noise reduction method based on perceptual convolutional network is designed in this paper.The method uses the residual learning to learn the mapping between low-dose CT images and image noise,which reduces the training difficulty.And combining perceptual loss and L1 loss,the expression capabilities of deep learning and the characteristics of medical image data are integrated.Through experimental verification,the proposed method can better balance the noise reduction degree of the image and the complete preservation of the texture details of the tissue,especially for the processing effect of the difficult soft tissue window in the field.A multi-dose CT image denoising method based on deep convolutional network was designed for different doses of CT scan.Based on the CT image denoising model,the model structure is further optimized to improve the noise reduction performance.The blind noise reduction method is introduced.For different noise reduction images,a good subjective visual effect and objective index scores are obtained by using the model processing on different datasets.