Radiomic Features,bridges Between CT And¹⁸F-FDG PET Images

Background:Previous studies have shown that the the density of CT image is related to the SUV maximum（SUVmax）,and the coarse CT image texture analysis shows that some texture features are moderately related to the SUV mean（SUVmean）.In terms of the fact that radiomics can display abundant image information that cannot be discerned by the naked eye,such as establishing a functional relationship between CT image features and PET parameters,this radiomic research may not only reduce the patients’medical expenses and avoid the patients`unnecessary radioactive exposure effectively,but also helps radiologists to diagnose accurately inthefuture.Methods:This study aim to delineate the relevance of the CT image features to the parameters from¹⁸F-FDG PET/CT by making use of PET/CT images of 206 lung cancer patients in our hospital.In total,1198 lymph nodes from both ordinary CT and thin-layer CT images of these206 patients were recruited in our study.The volume of interest（VOI）was drawn for each lymph in CT images.Features were extracted from the two types of CT images,and thereafter the radiographic features and SUVmax or SUVmean were analyzed.The same features in two types of CT images were selected by deep learning method.The prediction formulas with the SUVmax or SUVmean as the independent variables and the features as the dependent variables are established.The relationship between the predicted and measured values was subsequently analyzed.Also,relevant factorsto this relationship were assessed.Results:A total of 141 series（1683 features）features were extracted from each lymph node.Among the top 20 features selected by deep learning method,there were two same features from these two tyes of CT images.The general linear model is used to built the prediction formulas ofSUVmax and SUVmean.The analysis shows that the difference between the measured value and the predicted value of SUVmax is 0.000±2.46（t=0.000,P=1.000）,and the correlation coefficient between them is r=0.573（F=585.475,P=0.000）.The difference between the measured value and the predicted value of SUVmean is 0.000±1.19（t=0.000,P=1.000）,and the correlation coefficient between them is r=0.583.However,the Bland-Altman graph suggests that there are proportional errors in the prediction results of the two formulas.As the measured value increased,the prediction bias increases linearly（linear correlation coefficients are 0.819 and 0.812,respectively）.After analyzing the influencing factors such as pathological type,sex,injection dose,collection time and volume by One-Way ANOVA.It is suggested that the difference between the groups of dSUVmax or dSUVmean is consistent with groups of SUVmax or SUVmean,so ANOVAfailed todetect valuable influencing factors.Conclusion:When the CT image features and SUVmax or SUVmean are regressed by the general linear model,the accuracy of the prediction formula is good but the accuracy is poor.Differences between groups of PET parameters are uncertain.So the influence of image volume,injection dose and acquisition time on the prediction bias cannot be excluded.In summary,the relationship between CT image features and SUVmax or SUVmean is not a simple general linear model.They may be complex functional relationships that deserve further study.