Research On Low-dose CT Noise Reduction Algorithm Based On Image Frequency Separation

Computed Tomography(CT)technology plays an important role in image diagnosis.In recent years,the radiation dose problem of CT examination has received more and more attention,which is due to the ionizing radiation generated by X-rays during CT scanning can cause certain damage to the human body,which can lead to an increased risk of radiation-related diseases such as cancer.However,lowering the radiation dose increases noise/artifacts in CT images,resulting in lower image quality and interfering with the doctor’s diagnosis of the disease.To solve the noise/artifact problem in low-dose CT heavy images,this paper is committed to studying efficient image processing algorithms to obtain high-quality low-dose CT images,so as to assist doctors to identify and analyze lesions more accurately,and provide strong support for clinical decision-making.In this paper,from the perspective of image frequency separation,combining the advantages of traditional methods and deep learning,the characteristics of low-dose CT images are deeply explored,and four low-dose CT image noise reduction algorithms are proposed.The main work is as follows:(1)We putforward a denoising algorithm for low-dose CT image based on optimal wavelet basis and morphological component analysis(MCA),which aims to solve the problem of severe noise and artifacts in low-dose CT imaging.First,the high-frequency(HF)component coefficients in the horizontal,vertical,and diagonal directions of low-dose CT after the stationary wavelet transform(SWT)are weighted to obtain the wavelet basis selection coefficients,and the wavelet basis with the smallest wavelet select coefficient is selected as the optimal wavelet basis.Second,the artifacts are processed using the MCA algorithm based on online dictionary learning(ODL)for the HF component.Third,the improved low-dose CT images are obtained using the inverse stationary wavelet transform(ISWT),which uses the low-frequency(LF)component and the denoised HF component.The extensive experiments on simulated and real data demonstrated the images denoised using the optimal wavelet basis algorithm showed the highest objective evaluation index,followed by the other wavelet-based algorithms.In addition,the proposed method is superior to NLM,KSVD,BM3 D,WNNM and Wav Res Net in both quantitative and qualitative evaluation,which verifies the effectiveness of wavelet selection and the feasibility of the proposed algorithm.(2)To solve the problem of large amount of noise/artifacts,edge blurring and detail loss in low-dose CT images,we propose an edge protection and global attention mechanism(GAM)densely connected convolutional network for low-dose CT denoising.First,edge information was extracted using the improved eight-directional Prewitt operator,and then passed to each convolutional block through skip connections.Subsequently,a multiscale feature extractor and global attention mechanism were used for feature extraction.The final predicted images were then obtained by the noise reduction module.Further,a compound loss function based on mean squared error and perceptual loss was used to enhance the texture detail and improve the visual quality of the image.Extensive experiments on Mayo and Piglet datasets showed the effectiveness of the proposed method in reducing noise/artifacts while preserving edges.Compared with RIDNet,MWCNN,CNCL,EDCNN,RED-CNN,CTformer,ADNet and Q-AE methods,it has better performance in both objective index and subjective effect.(3)To more precisely extract the high-frequency information in low-dose CT images,preserve the edge and texture information of the images,and make up for the insufficiency of the traditional operators to extract the edge information,we propose a gradient extraction based multi-scale dense cross network for low-dose CT image denoising.The method firstly uses a shallow feature extraction block to extract features from the low-dose CT image;at the same time,the low-dose CT image is input into a gradient extraction network based on the U-Net model to extract the gradient(high-frequency information)of the image.Secondly,the shallow features are summed up with the high-frequency images to obtain rich feature maps.Finally,the feature maps are fed into Backbone Net to obtain better quality prediction images.In addition,we propose a compound loss function based on Charbonnier loss and gradient loss to enhance the texture details and improve the visual quality of the images.Extensive experiments on Mayo and piglet datasets show the model can effectively remove noise/artifacts from low-dose CT images while preserving the image structure and edge information.(4)To address the problem that the traditional convolutional neural network relies heavily on cascading convolutional layers to extract high-level features,which can only perceive the features in the local area,and the previous methods focus on the high-frequency noise individually while ignoring the low-frequency part,which often fails to achieve the optimal image quality.Therefore,we propose a combined frequency separation network and Transformer for low-dose CT denoising.Firstly,the model decomposes the low-dose CT images into low-frequency images and multi-layer high-frequency images by frequency separation blocks.Then,the low-frequency components are fused with the high-frequency components of different layers to remove the noise in the high-frequency components with the help of the potential texture of low-frequency parts.Next,the estimated noise images can be obtained by using Transformer stage in the frequency aggregation denoising block.Finally,they are fed into the reconstruction prediction block to obtain improved quality images.In addition,a compound loss function with frequency loss and Charbonnier loss is used to guide the training of the network.The performance of the model was validated and evaluated based on the Mayo dataset and the Piglet dataset.The model achieves optimal metrics compared to previous representative models in different architectures.The experimental results show that the model is a state-of-the-art model with noise/artifact suppression and texture/organization preservation.