Room: Stars at Night Ballroom 2-3
Purpose: To evaluate the stability of radiomics features using 4D-CT as an alternative to test-retest CT scans.
Methods: 4D-CT images with 10 breathing phases were acquired with Varian RPM for 10 patients with lung tumors. For each patient, two contours (GTV and GTV-1mm) were delineated on each breathing phase. GTV-1mm was generated by subtracting a 1mm inner margin from GTV to reduce the low-density air enclosed. A total of 64 first-order histogram-based and second-order texture-based features (Gray Level Co-occurrence Matrix as GLCM, and Gray Level Run Length Matrix as GLRLM) were extracted using an open-source CERR radiomics platform. The mean, standard deviation and coefficient of variation (COV) of each feature were calculated from all breathing phases for each patient. Features with mean of COV under 5% were deemed stable features. The variation was also calculated among all features for each patient to identify source of large variation in unstable features.
Results: Regardless of difference in contours, 2 out of 22 1st order histogram-based features (entropy and rms) and 9 out of 42 2nd order texture-based features (jointEntropy, sumEntropy, invDiffMomNorm, invDiff, invDiffNorm, secondInfCorr, sre, re and rlnNorm) are identified as stable. While some other features as 5 histogram-based (mean, median, skewness, kurtosis and coeffVariation) and 3 texture-based features (clustShade, clustPromin and haralickCorr) are found to be unstable, with COVs beyond 100%. In addition, all features were sensitive to patientâ€™s motion. Larger COVs were observed for patients with larger breathing motions despite the fact that the same tumor image captured within 3-5 minutesâ€™ scan time.
Conclusion: The results showed that the stability of radiomics features could be evaluated using 4D-CT as an alternative to test-retest CT scans. We were able to identify stable radiomics features which may have the potential to be used for radiomics-guided clinical studies.