Research On Denoising Algorithm Of Low-dose CT Image Based On Deep Convolutional Sparse Coding

Computed Tomography（CT）,based on the attenuation of X-rays penetrating objects,can non-invasively reconstruct tomographic images.It is widely used in medical imaging and industrial inspection,and is one of the greatest scientific and technological achievements at the end of the 20th century.However,some studies have shown that excessive X-rays will increase the risk of cancer.In order to reduce the"secondary damage"to the subject,hospitals seek to obtain clear CT images with a lower radiation dose.In the field of image post-processing,traditional CT denoising methods are prone to remain noise and artifacts.Deep learning methods,based on convolutional neural networks,have excellent denoising and detail retention ability,but it is difficult to understand the internal logic of the denoising process.Therefore,under the premise of effectively improving the quality of low-dose CT,how maintaining good interpretability of the denoising algorithm is a hot spot in medical image processing research.In this paper,with the combination of convolutional neural network and convolutional sparse coding,two studies on low-dose CT denoising are carried out by means of deep convolutional sparse coding.The specific research contents are as follows:（1）It proposed a deep convolutional sparse coding algorithm based on multi-scale edge extraction and weighted L₁ norm.The proposed learnable weighted convolutional sparse coding model adds a learnable weight to the penalty term of the convolutional sparse feature map on the objective function.This is because the image results of different individuals and different parts of the same individual show different noise levels.Theoretically,the convolutional sparse feature maps with low noise are sparser than those with heavy noise,and different degrees of punishment can better suppress noise.The objective function is calculated by the learnable weighted iterative soft threshold algorithm,and the unfolding process is clearly explainable.In the implementation,the additional weights on the sparse feature map are parameterized by the channel attention module,and the texture details lost in the denoising process are supplemented by the designed multi-scale edge extraction module.The experiments show that the number of parameters of the proposed edge extraction-based deep convolutional sparse coding network,compared with the networks in the comparison methods,is lightweight.Furthermore,the proposed denoising model plays a good role in noise suppression and detail protection for low-dose CT images,which is helpful for clinicians to accurately judge the CT images.（2）It proposed a deep convolutional sparse coding（also known as deep convolutional dictionary learning）algorithm based on adaptive window function.Convolutional dictionary atoms and convolution kernels in convolutional neural networks can be regarded as special filters with rectangular windows,which are easily affected by spectrum leakage in the process of truncated filtering.For this reason,the convolutional dictionary learning model based on adaptive window function is proposed.Through the designed two-dim adaptive window function,the image blur caused by spectrum leakage is effectively alleviated.Furthermore,the penalty term in the objective function no longer adopts fixed manual priori,but takes the form of generalization,and uses the proximal gradient descent algorithm to iteratively optimize the objective function.In the implementation,the proposed edge extraction-based convolutional dictionary network with adaptive window expresses the generalization priori through the convolutional neural network,and applies the hybrid dilated convolution kernel instead of conventional convolution kernel to further alleviate spectrum leakage.The overall architecture of the network can one-to-one correspond to the iterative process of the objective function.In addition to adding the edge extraction module to improve details,the loss function combines the global pixel-level loss and the local energy loss based on the image blocks.The experimental results show that the number of parameters and the inference times of the proposed interpretable denoising network is very competitive,and its denoising effect on low-dose CT images is better than the comparison methods qualitatively and quantitatively.