MO-DE-207B-09: A Consistent Test for Radiomics Softwares




The purpose of this study is to investigate the consistency of the features extracted by different radiomics software.


CT sets of 212 patients with rectum tumor and three existing radiomics feature extraction tools (IBEX from MD_Anderson, RADIOMICS from Maastro, and a MATLAB based code-set from our institution) were enrolled in this study. Among thousands of features that were extracted by three softwares, 273 (between our codes and RADIOMICS), 51 (between IBEX and RADIOMICS) and 34 (between our codes and IBEX) feature pairs were proven to be conjugated according to feature definition. As for each matched feature pair, a Spearman's rank correlation test was conducted to measure the correlation of the feature pair yielded by different feature extraction tools. Furthermore, each of three datasets were standardized by z-score method and undergone a clustering analysis with NMF, respectively. Finally, the consistent of clustering was verified by a chi-square test.


The consistent between IBEX and RADIOMICS (34 out of 54 features' coefficient of correlation were above 0.9) were better than consistent between these two tools and our codes, from FDSCC (9 out of 33 and 9 out of 34 features' coefficient of correlation were above 0.9, respectively). One of the causes we expected to be reasonable is the different GLCM definition among the algorithms in three softwares. We used 2D GLCM features, while IBEX used 2.5D and RADIOMICS used 3D, respectively. On the other hand, the most consistent clustering was between datasets yielded by IBEX and RADIOMICS, with an accuracy of 0.87 and a p-value (Chi-sqr test) of 4.5e-27. Meanwhile, the consistence accuracy of clustering between our codesets and other software was about 0.7.


This work indicated the potential inconsistency between different software has little impact on the clustering outcome. Additional attention should be paid when using different software in radiomics research.