Low-Dose CT Image Quality Improvement With Fusion Of Perception Loss And Attention Mechanism

In the problem of improving the quality of low-dose CT images,there is a mismatch between the quantitative index and the qualitative index of the denoising algorithm.In medical imaging,the amount of X-ray radiation in a CT scan is much higher than that of conventional X-ray imaging.One of the most commonly used methods to reduce the amount of X-ray radiation is to reduce the operating voltage or current of the X-ray tube,but this method will cause a lot of quantum noise,degrades the image quality,and affects the diagnosis of the patient's condition.Therefore,eliminating or suppressing the noise of low-dose CT images and improving image quality have become a research hotspot.In recent years,researchers have proposed many classic machine learning algorithms and deep learning methods to try to improve the quality of reconstructed CT images.Compared with machine learning algorithms,deep learning methods have achieved better experimental results in both quantitative and qualitative indicators,but there is a problem of mismatch between qualitative and quantitative indicators.For this reason,we proppose a low-dose CT collaborative denoising network with perceptual loss and attention mechanism.The collaborative algorithm can significantly improve the problem of low quantitative indicators of existing methods on the basis of ensuring visual effects.The model also introduces an 8-direction edge detection layer at the network input,which optimizes the original 4-direction edge detection layer,which can extract richer texture and structure information,and further enhance the network effect.The experimental comparison results based on the phantom data set and the real clinical data set show that compared with the mainstream work,the method in this paper has better performance in subjective visual perception and objective indicators PSNR and SSIM indicators;compared with the general neural network based on perceptual loss The network further solves the problem of mismatch between qualitative and quantitative indicators in the low-dose CT denoising problem.The main contributions of this thesis include the following:(1)In order to illustrate the effectiveness of general deep networks for improving the low-dose CT image quality,the denoising performance of traditional machine learning methods and the common convolutional neural networks are analyzed and compared.Analysis and comparison of machine learning algorithms TV,KSVD,BM3 D,and a low-dose CT denoising network based on the SRCNN network.Experiments were performed on the above-mentioned networks using the phantom data set,and the PSNR values of the CT images after the experiment were obtained.The experimental results show that the common deep network is effective in improving the experiment of quantitative indicators.(2)Optimize the edge detection layer to improve the effect of the deep network for low-dose denoising work.The 8-direction Sobel gradient operator is applied to the edge detection layer of the low-dose CT denoising network based on the perceptual loss and edge detection layer,and the low-dose CT denoising network based on the perceptual loss and the8-direction Sobel gradient operator edge detection is obtained.The network can extract more detailed features of the gradient.The network was experimented based on the phantom data set and clinical data set.The quantitative and qualitative evaluation results of image quality improvement experiments based on phantom data sets and clinical data sets show that our work to optimize the edge detection layer is effective for improving the quantitative and qualitative experimental effects of the network.(3)A network of cooperative perception loss and attention mechanism is proposed to improve the quality of low-dose CT images,which effectively solves the problem of mismatch between qualitative and quantitative indicators.The work is mainly to introduce 4multi-channel attention mechanism modules into the low-dose CT denoising network based on the perceptual loss and edge detection layer,so as to obtain a low-dose CT denoising network that synergizes the perceptual loss and attention mechanism.Then,based on the phantom data set and clinical data set,the same multi-channel attention mechanism was introduced on the 4-direction and 8-direction perceptual loss network to conduct related experiments.The qualitative and quantitative results of the final experiment show that the low-dose CT denoising network that cooperates with the multi-channel attention mechanism and the loss of perception is effective for solving the problem of qualitative and quantitative index mismatch in low-dose CT denoising.