Research On Low-dose CT Reconstruction Algorithm Based On Fusion Learning

Although Computed Tomography(CT)has provided doctors with intuitive support for diagnosis,the possible harm caused by high-dose has been highly concerned by researchers.For this reason,the use of low-dose tomography has become the preferred solution.However,the reduction of radiation dose often leads to the introduction of a large amount of noise,which in turn affects the doctor’s diagnosis.Therefore,how to effectively improve the image quality while reducing the radiation dose has become an important challenge for the current CT reconstruction technology research.Generally,CT reconstruction involves two domains: the projection domain(sinogram data)and the image domain.In view of this,the current research has corresponding thinking in both domains.Generally speaking,the sinogram domain aims at solving the reconstructed mapping problem from the perspective of statistics and physics,while the image domain seeks solutions from the perspective of computer vision.Considering that the entire reconstruction process actually involves the signal reprocessing process of the projection and image domains,the previous single domain processing has inherent shortcomings.Although there have been studies dealing with dual domains in the past two years,the mapping process did not take into account the asymmetry of information processing,and tried to use a single method to solve complex problems,which led to uncertain interference in the reconstruction results.In fact,the target of reconstruction technology is the improvement of visual effects.Then,how to construct the inner connection between visual effects and tools has become the core issue of research.This research aims at reconstructing low-dose CT with good visual effects.At present,single-domain reconstruction has problems with residual noise and artifacts,and the iterative algorithm has limitations such as large computational consumption.Aiming at these problems,the following explorations have been specifically made:1、 The single-scale convolution method for feature mapping in the sinogram domain often leads to loss of relevant details.To solve this problem,a multi-receptive field residual convolution method(S-DRN)is proposed to learn a multi-scale nonlinear mapping method for Sinogram data,in order to retain as much contextual information as possible.2、 At present,most existing image domain reconstruction methods use the mean square error loss function in order to achieve the effect of global noise reduction.However,it is often overcorrected(the result is excessively smooth)from the perspective of experimental results.And even if the VGG loss function is used,it is unavoidable to introduce features that have nothing to do with the target CT image in the result.The reason is that the above-mentioned loss functions mainly consider the feasibility of method migration,but lack consideration for the differentiation of CT data.For example,the VGG loss function uses natural texture features as a priori knowledge,which obviously has a gap with the visual diversity of CT images.To this end,this research proposes a feature extraction network(FEN)tightly coupled with CT texture features.And combining the advantages of the above loss function in global feature recovery,construct a fusion feature loss function with trainable weights.3、 Due to the lack of consideration of the global correlation of the extracted features,the existing neural network algorithms are prone to produce different degrees of edge blur in the reconstruction results.To solve this problem,this study proposes a modified residual encoder-decoder network(SRED-Net),which uses the self-attention mechanism to improve the global correlation perception of features,and finally achieve the purpose of restoring texture details.Combining the above innovative thinking of the projection domain and the image domain,this research designs and implements a texture-aware dual-domain mapping network(TADDM-Net),which fully integrates the information of the projection domain and the image domain.Validated by the public AAPM-Mayo-Clinic data set,the experimental results show that S-DRN can effectively retain sinogram information and remove a certain degree of fringe artifacts.SRED-Net can achieve a good balance between noise reduction and texture restoration.Furthermore,compared with the existing dual-domain network,according to the experimental results,the Haralick index of TADDM-Net proposed in this paper is greatly improved.And from intuitive visual observation,the method proposed in this paper can better retain the effective information of the sinogram domain and the image domain,and realize the maximum visual effect restoration of the texture at the same time,so as to better meet the doctor’s diagnosis requirements.