Research On Sparse Tissue Segmentation Algorithm Under Low-dose CT

Body Composition Analysis(BCA)is a process of precise quantitative analysis of various tissues in the human body.BCA,as an important research topic in the intersection of medical image processing technology and computer technology,has played an important role in the intelligent identification of lesions,clinical diagnosis,and pathological analysis.The composition of the human body mainly includes visceral fat,subcutaneous fat,muscle fat,muscle,bone and so on.However,at present,there is no unified medical definition and specification for the above components in CT images,and the complex shapes of various human tissues are difficult to label,and there is a lack of relevant semantic segmentation data sets.hinder.First of all,skin tissue(Skn T)and muscle adipose tissue(IMAT)are two important sparse tissues in the human body.Skn T,as the outermost tissue of the human body in direct contact with the environment,plays an important role in resisting germs and protecting internal organs.In the CT image,its main characteristics are linear,continuous and has a certain degree of sparsity,and the number of voxels accounts for a small proportion in the overall CT image.The distribution of IMAT in the human body is more irregular than that of Skn T.In CT images,IMAT is mainly presented as dense blocks and small and sparse shapes.The above characteristics make the quantification of Skn T and IMAT more difficult.Secondly,this paper completes the construction of dense data sets(OAM,IAM,SKN,Dphm,SMR,etc.)and sparse data sets(IMAT,Skn T,VAT,SAT),forming a relatively complete BCA semantic segmentation in the whole body data set.Design an end-to-end semantic segmentation model for sparse class organization.The model includes a convolution module that combines edge detection operators to extract sparse and fine linear features,a feature map dense fusion module that fuses semantic information in the underlying feature map of the model,and an ultra-deep The network architecture is used to fit the complex data distribution of sparse organization,which is significantly improved compared to the existing classic convolutional neural network.Compared with the existing classic segmentation model based on convolutional neural network,the effect on the IMAT and Skn T datasets has been improved by 1.2% and 1.32%,especially for some linear small and sparse tissues,the model in this paper can get better visual effects.Finally,in order to further improve the segmentation accuracy and processing efficiency of sparse class organizations,this paper also proposes a filter semantic segmentation model based on a multi-scale voting mechanism.The model first performs rough positioning of sparse tissues through the sensory check at the center,and then analyzes the data distribution of voxels around the rough positioning results through four voter filters to generate their own segmentation results.Finally,the four filters pass the voting mechanism Determine the final segmentation result and complete the semantic segmentation of sparse class organization.The model surpasses the existing classical convolutional neural network in terms of segmentation effect,and processes CT images with fast speed,high efficiency and accurate results.Compared with the existing classic segmentation models,the filter model proposed in this paper has obtained the best segmentation results of 0.9684 and 0.8383 on the Skn T and IMAT datasets,and can quickly process CT images,and can better process Some thin linear sparse class organization.