Computational Detection of Latent Quantitative Instability In Daily CBCT Using a Multi‑Material Phantom and Radiomic Variance Analysis
Abstract
Purpose
To deploy a fully automated, Python-based analysis pipeline for evaluating short-term quantitative stability in Elekta CBCT, using a multi-material density phantom to uncover HU fidelity drift that is undetected by conventional QA.
Methods
A prospective longitudinal study will be conducted using a CT density calibration phantom (16 inserts, –1000 to +1000 HU). The phantom will be scanned three times daily (pre-, mid-, and post-clinic) over 30 consecutive treatment days on an Elekta Versa HD (XVI 5.0, standard head protocol). All analysis will be performed using a custom Python pipeline integrating: Automated insert segmentation and ROI placement Computation of HU accuracy, linearity, and short-term precision Extraction of first-order and texture-based radiomic features (entropy, correlation) from uniform regions Variance component analysis and time series decomposition The sensitivity of routine QA metrics to observed quantitative drift will be evaluated via mutual information and multivariate regression, also implemented in Python.
Results
Pilot data suggest routine daily QA will pass tolerances, yet substantial intraday variance (>40% total) will emerge in soft-tissue HU (−100 to +100), undetected by standard methods. Texture radiomics (e.g., GLCM correlation) from uniform regions are projected to correlate strongly (R² > 0.80) with HU drift, outperforming noise metrics, while routine QA shows negligible mutual information (<0.1) with instability.
Conclusion
This work will demonstrate that a computational, Python‑driven approach coupling a multi‑material phantom with radiomic variance analysis can reveal clinically significant quantitative drift in CBCT that is invisible to threshold‑based QA. We intend to propose noise texture as a feasible, Python‑extractable surrogate biomarker for establishing predictive, variance‑aware quality monitoring, providing both a methodological framework and an open‑source toolset for implementation in clinical practice.