The Impact Of The Shape And Size Of Roi On The Reproducibility Of CT-based Radiomics Features

Background: It is well known that the various tissues and organs are very different in their anatomical morphology.When lesions occur,there are huge differences in the size and morphology of various diseases.In addition to the different sizes and morphology of different kinds of lesions,there are huge differences in the size and morphology of the same lesions.Especially for digestive tract neoplastic lesions,if the region of interests(ROIs)are set according to the maximum contour of the lesion,similar tumors also vary greatly in the shape and size of the ROIs.During repeated delineation of the same lesion,the shape and size of the ROIs may vary due to posture or breathing and physiological movements,and the shape and size of different levels of the same tumor are also quite different.The shape and size of the ROIs are different when be delineated in the target segmentation and whether it will hamper the process of radiomics features extraction is worth further exploration.Purpose: To investigate the reproducibility of CT image radiomics features(RFs)over different shape and size of ROIs on phantoms and human cohort.Materials and Methods: Four different density liquid phantoms(uniform Water phantom,soy sauce with 52% sodium(Na)chloride,rice vinegar with acetic acid≧3.5 g/100 ml and milk with fat content ≧ 3%)were scanned ten times using abdomen clinical protocols and patients were scanned with the same CT and the same scanning parameters.The same 70 ROIs were delineated repeatedly on the Cross-sectional level of the four phantoms and the normal liver parenchyma,renal parenchyma,spleen,bladder with software,and 489 RFs were extracted.Then the absolute value of the coefficient of variation(|COV|)of these RFs were calculated,and sorting the |COV| of each features in each phantom from little to large in each scan,so,10 ranks of each feature were obtained.Then the mean(X)and standard deviation(SD)of the 10 ranks of each feature in each phantoms and each patients were calculated.Finally,draw a scatter plot of RFs with X as the horizontal axis and SD as the vertical axis.According to the the distribution of the scatter plot,reproducibility of RFs were classified as low,medium and high by reference to 100 <X≦150 and SD≦40,50 <X≦100 and SD≦40 and X≦50 and SD≦40,respectively.Results: Based on the phantoms and clinical experiments to explore the reproducibility of the RFs,the results indicate that there were 28,50,and 30 different degree reproducibility RFs appeared simultaneously in the four different density materials phantoms respectively,and there were 23,44,and 33 different degree reproducibility RFs appeared simultaneously in the clinical patients experiments respectively.In this study,the number of different degree reproducible RFs extracted from phantoms and patients were different,and none of them were statistically significant(P>0.05).There were 23,42,and 28 different degree reproducibility RFs appeared simultaneously in two kinds of experiments respectively.And the RFs of different categories were also affected differently by the shape and size of ROIs.Conclusion: The shapes and sizes of ROI had great impact on the reproducibility of RFs,but the impact was relatively fixed,and they also had different influences on the different categories of RFs.Among the six categories of RFs,the NID features were most affected by the shapes and sizes of ROI.