Study Of Automatic Detection Algorithm Of Low Concentration Materials In Multi Energy CT Based On Deep Learning

Multi-energy CT(Computed Tomography)can count the number of photons in the separated energy bin under a single X-ray irradiation to realize the recognition of different materials.Due to the narrow energy bin,the material images decomposed by multi-energy CT often have low contrast,so it is very difficult to detect low concentration materials.At the same time,in the CT concentration detection,we need to know the material attenuation coefficient,which is highly restrictive.In order to solve these problems,this paper proposes two pixel level concentration detection algorithms based on multi-energy CT value,and tries to use deep learning method for concentration detection.The two algorithms can detect the concentration or concentration range of each pixel,realize the detection and quantification of concentration,and enhance the image contrast of iodine(I)and gadolinium(Gd).The main contents and contributions are as follows:(1)In order to solve the limitation of CT concentration detection which needs the aid of material attenuation coefficient,this paper proposes a Concentration Detection Based on Regression(CDBR)algorithm.CDBR can detect iodine and gadolinium concentration in multi-energy CT.CDBR uses the idea of regression to fit the correlation between CT value and concentration under different energy bands,and establishes the regression equation between CT value and concentration to realize the detection of concentration.Based on this algorithm,two deep network models are proposed,namely SRCN(Simple Regression CNN)and S-Res Net(Simple Residual Neural Network).SRCN directly learns the relationship between CT value and concentration,and S-Res Net indirectly learns the relationship between CT value and concentration by learning residuals.Finally,after the completion of the CDBR algorithm,iodine and gadolinium concentration prediction image and the concentration binarization image can be obtained.(2)In order to solve the problem that the detection concentration of CDBR<=1mg/cc is not ideal and the sample is not balanced,we propose a Three Stage Concentration Detection(TSCD)algorithm in this paper.The TSCD algorithm applies the idea of classification to realize the phased detection of iodine and gadolinium concentration in multi-energy CT.In the three stages,pixels with concentrations of >=2mg/cc,[1,2)mg/cc,[0.5,1)mg/cc are detected respectively,and finally concentration binarization images with concentrations of >=0.5mg/cc are obtained,and regions with concentrations of 1mg/cc and 0.5mg/cc can be quantified.Based on this algorithm,this paper proposes S-VGG19(Simple VGG19),and proposes the BC-Focal loss function to adjust the imbalance of background and filling material area in data set,and optimize the results of three stages.(3)In order to solve the problem that low concentration <=2mg/cc region of material image is low contrast,this paper proposes a method that use the concentration detection results of CDBR and TSCD to enhance the contrast of material image.By fusing concentration binarization images obtained by CDBR and TSCD with material images obtained by Singular Value Decomposition(SVD)pseudo inverse method,based on the detection results of low concentration,the region of low concentration of material image which is not displayed is displayed,and the contrast of material image is enhanced.The experimental results show that the CDBR algorithm(SRCN and S-Res Net)can detect and quantify the concentration of iodine and gadolinium,and obtain the concentration prediction image.And the contrast of iodine and gadolinium in the region of concentration >=1mg/cc is significantly enhanced,and the contrast in the region of concentration 0.5mg/cc is relatively enhanced.TSCD algorithm(S-VGG19+BC-Focal)can effectively detect the materials with concentration >=0.5mg/cc.The low concentration materials with concentration of 1mg/cc and 0.5mg/cc can be optimized and quantified,and the image contrast of low concentration iodine and gadolinium with concentration of 2mg/cc,1mg/cc and 0.5mg/cc is significantly enhanced.The effectiveness of the two methods is proved.