Preliminary Study of Radiomics Feature Stability with HyperSightTM CBCT Imaging for SBRT Treatments
Abstract
Purpose
To evaluate the stability of radiomic features extracted from HyperSightTM CBCT imaging in stereotactic body radiation therapy (SBRT) treatments.
Methods
A total of 10 cancer patients including prostate, lung, spine and kidney treatments undergoing SBRT radiotherapy with a TrueBeam Linac were retrospectively collected for this study. To assess the repeatability of CBCT-based radiomics features, a second CBCT was acquired after beam delivery when the patients were still in the treatment position for every fraction. The interval between these two CBCTs was less than 10 minutes. Agreement between the features calculated from the pre- and post- CBCT images using pyradiomics were evaluated by the statistical measure concordance correlation coefficient (CCC). For the feature reproducibility and detection of meaningful tissue biological changes, we examined CCC between the radiomic features using pre-CBCT images from the first and subsequential fractions.
Results
We collected a total of 96 CBCT images. A comprehensive set of 91 texture features were initially selected, including first-order conventional (18) features, second-order grey-level co-occurrence matrix (GLCM, 22), grey-level dependence matrix (GLDM, 14) grey-level run length matrix (GLRLM, 16), neighboring grey-level dependence matrix (NGLDM, 5), and grey-level size zone matrix (GLSZM, 16) features. However, 26 features were considered reproducible if the criterion of the CCCs between pre- and post- CBCTs is set to 0.9. With these selected features, feature reproducibility was conducted to examine the feature progression across subsequential fractions. We observed that the lung cases yielded fewer acceptable features, especially the first-order ones due to the evident effect of motion.
Conclusion
Feature stability has been a known issue in radiomics analysis to reliably predict clinical endpoints. Our preliminary study showed that radiomics features that were pre-selected as repeatable could become not reproducible due to various factors such as different imaging protocols and motion effects.