Fifty-seventh annual meeting of the American association of physicists in medicine
TH-CD-BRA-04: Assessing How Stochastic CT Noise Can Lead to Systematic Proton Range Errors
To demonstrate that the discontinuous nature of the CT number to stopping power ratio (SPR) calibration curve, combined with the presence of uncorrelated zero-mean Gaussian CT noise, leads to non-negligible and tissue-dependent systematic errors in SPRs and proton range, typically not taken into account in usual safety margins for proton therapy.
Increased systematic errors with noise standard deviation have first been observed in proton range Monte-Carlo simulations with stoichiometric calibrations, whereas only zero-mean random errors were expected. Their existence has then been proved analytically for arbitrary calibration curves and material distributions along the proton path and validated through continuous slowing down approximation (CSDA) simulations. Their importance relative to the other sources of uncertainty has then been estimated in head-and-neck, lung, and pelvis patient data for multiple beam orientations. CT noise has first been reduced using a double-pass median filtering approach and a Gaussian noise has then been added to obtain total standard deviations between 10 to 40 HU.
This study provides close form equations for the systematic error and uncertainty on SPR and proton range due to uncorrelated noise. They have shown to accurately match CSDA simulation results with realistic calibration curves and material distributions. Depending on the tissue distribution and the position of the discontinuities along the curve the resulting effect on range varies but has shown never to cancel out completely as opposed to common beliefs. The analysis performed on patient data with clinical calibration curves has confirmed that fact with estimated systematic range errors of 0.2–0.5% and uncertainties (4 σ) between 0.5 and 1% with typical CT noise levels.
A new source of SPR and range systematic errors has been highlighted and proved not to be negligible compared to the 3.5% uncertainty reference value used for safety margin design
This study is linked to a public partnership between UCL and IBA funded by the Walloon region under convention number 1017266 and 1217662