Purpose: To explore the most influential factors on the reproducibility of cone-beam computed tomography (CBCT) radiomics features.
Methods: Three sphere materials (acrylic, rubber, and cork: 30 mmφ) were placed at three different locations (upper-central, lower-central, lower-peripheral) in an anthropomorphic chest phantom. The CBCT images were acquired using Halcyon (125 kV, 168 mAs, CTDI(vol): 3.36 mGy) and TrueBeam (125 kV, 225 mAs, CTDI(vol): 3.46 mGy; Half-fan). The images were reconstructed using following three algorithms: (i) Feldkamp-Davis-Kress (FDK) algorithm in Halcyon (FDK(Halcyon)), (ii) FDK algorithm in TrueBeam (FDK(TrueBeam)), and (iii) iterative CBCT (iCBCT) algorithm with medium noise suppression level in TrueBeam (iCBCT(TrueBeam)). The targets were manually delineated in Eclipse treatment planning system version 15.6 by three medical physicists under the lung window setting. A total of 93 radiomics features, including 18 first order and 75 texture features, were extracted from the inside region of interest for each material using PyRadiomics (version 2.2.0). The effects of the inter-observer, target location and reconstruction algorithm variability on radiomics features were evaluated using coefficient of variation (CV). Features with CV value less than 5% were considered highly reproducible.
Results: For inter-observer and inter-target location comparison, the median percentage of reproducible features in FDK(Halcyon), FDK(TrueBeam) and iCBCT(TrueBeam) were 51.6% (range: 38.7-62.4), 51.6% (range: 43.0-76.3) and 55.9% (range: 45.2-69.9), and 35.5% (range: 28.0-48.4), 33.3% (range: 26.9-40.9) and 33.3% (range: 24.7-38.7), respectively. For inter-reconstruction algorithm comparison, the median percentage of reproducible features were 20.4% (range: 11.8-29.0).
Conclusion: Inter-reconstruction algorithm variability may have a greater influence on the reproducibility of CBCT-based radiomic features than the inter-observer and target location variability. Even with the same FDK algorithm, reconstruction device differences have a large effect on the radiomics features reproducibility. It is recommended that the reconstruction algorithm should be unified for CBCT-based radiomics analysis.