Research On Low-dose CT Image Noise/Artifact Suppression Algorithm Based On Deep Learning

Computed Tomography(CT)is a medical imaging technology widely used in clinical medicine.Compared with other imaging technologies,CT technology has the advantages of short inspection time,high accuracy and reasonable price,and can realize the visualization of human fine structure in a short time.However,CT scans can cause ionizing radiation hazards and increase the risk of cancer in patients.Therefore,in order to reduce the radiation dose,lowdose CT(LDCT)scanning is usually performed by minimizing the X-ray flux.However,this will lead to an increase in CT image noise and artifacts,reducing image quality and accuracy.Aiming at the needs of LDCT image denoising and artifact removal,this paper uses the advantages of data-driven,high performance and fast execution of deep learning to construct a variety of low-dose CT image denoising algorithms,and pays special attention to the interpretability of the model.The proposed algorithm effectively improves the performance of LDCT image denoising,provides higher quality and more accurate image reconstruction results for the field of medical imaging,and provides more reliable image information for doctors.The main research work and innovative contributions of this paper are as follows:1.Aiming at the over-smoothing problem that may occur when Convolutional Neural Network(CNN)processes low-dose medical CT images,a dual-path three-stage network and a cross-connected multi-scale feature fusion network are designed.By extracting features of different levels and scales,denoising and feature fusion,the model aims to play the role of different types of features,and the effectiveness of the model is verified by experiments.2.In order to explore the effect of frequency domain information on low-dose medical CT image denoising,this paper proposes a low-dose CT image denoising method based on dual domain.This method uses the characteristics of both the Discrete Cosine Transform(DCT)domain and the spatial domain to remove image noise and improve image clarity.In the discrete cosine transform domain,a new convolutional attention residual network is designed to enhance the internal and external relationships of different channels.In the spatial domain,a top-down multi-scale codec structure is proposed to obtain multi-scale information.In order to better optimize the results,a composite loss function composed of L1 loss,Charbonnier loss and edge loss is used.The experimental results show that the proposed method performs well in removing noise and preserving texture details.It not only shows excellent performance in objective evaluation indicators,but also shows significant improvement in visual effects.3.In order to solve the problem that most of the current low-dose CT image denoising algorithms based on neural network are more like black box,this study combines the physical model and proposes a decomposition iteration strategy suitable for low-dose CT image denoising.First,the low-dose CT image is decomposed into a high-frequency part and a lowfrequency part using the Hessian matrix.Then,the deep learning network is used to process the high frequency part and the low frequency part.Finally,the high-frequency information after noise reduction and the low-frequency information after enhancement are re-fused to obtain an iterative denoised image.By subtracting the results of the previous iteration from the results of this iteration,and using the difference as the input of the next iteration,the process is repeated continuously.Through experimental verification,the denoising method shows a good noise artifact suppression effect on low-dose CT images.4.In order to deal with the challenge of noise artifact removal in low-dose CT images,this paper proposes a method combining traditional denoising methods and deep learning networks,and designs a more accurate and more interpretable denoising network framework.The framework integrates the noise optimization model of Adaptive-weighted Total Variation(AWTV)expansion,the Gaussian curvature guided edge detection model and the image restoration model into a CNN framework.Firstly,the noise optimization model is constructed by learning the parameters in the adaptive weighted total variation regularization model to approximate the noise level in the normal dose CT image.Then,an edge detection network is constructed using Gaussian curvature to directly predict clear edges from noisy images.Finally,under the guidance of the noise optimization model and the edge detection model,the finer details are restored and the denoised image is reconstructed.The experimental results show that the network recovers the structure of accurate low-dose CT images with limited image details.