Room: Exhibit Hall | Forum 8
Purpose: Radiomics, the extraction and analysis of quantifiable features from medical images, is expected to play important roles in radiation oncology. There are several radiomics software packages developed to extract the image features and our study is to investigate the uncertainty and consistency of the features calculated using two open-source radiomics software packages.
Methods: Six brain MRI T1-weighted images with single brain metastasis lesion contoured were used. Two publicly available and widely used and accepted packages were included for comparison in this study. Three feature classes are extracted without image pre-processing: first order, shape and gray level co-occurrence matrix (GLCM). Among all the features extracted, 37 features (16 first order, 5 shape and 16 GLCM) were paired according to the feature definition in the two packages. Pearson correlation coefficient was calculated for each matched feature pair to evaluate the correlation of the features obtained from the two packages.
Results: For first order feature, the uniformity, skewness, and entropy feature percentage differences were: 44.67% Â± 1.66%, 18.01% Â± 27.91% and 11.89% Â± 2.81%. For shape feature, the surface area, surface volume ratio, and sphericity feature percentage differences were: 18.51% Â± 2.85%, 18.81% Â± 2.74%, 18.71% Â± 2.76%. For 26 out of 37 feature pairs, the Pearson correlation coefficient was higher than 0.9. The reasons for the feature discrepancy between the two packages include the RT structure ROI interpreted differently and the different GLCM orientation definition.
Conclusion: Algorithmic implementation of features in each package could vary and there is still inconsistency between radiomics software packages. Some efforts have been made, including the image biomarker standardization initiative (IBSI), to standardize feature definitions. When selecting an radiomics package, the user should verify that it meets their task-specific need and provides expected results. A benchmark imaging dataset with known texture values would be helpful.