Site-specific alert criteria to detect patient-related errors with 3D EPID transit dosimetry

Abstract

Purpose: To assess the sensitivity of various EPID dosimetry alert indicators to patient-related variations and to determine alert threshold values that ensure excellent error detectability.

Methods: Our virtual dose reconstruction method uses in air EPID measurements to calculate virtual 3D dose distributions within a CT data set. Patient errors are introduced by transforming the plan-CT into an error-CT data set. Virtual patient dose distributions reconstructed using the plan-CT and the error-CT data set are compared to the planned dose distributions by γ(3%/3 mm) and DVH analysis using seven indicators: ΔD_ISOC , γ-mean, near γ-max, γ-pass rate, ΔPTV_{D 2} , ΔPTV_{D 50,} and ΔPTV_{D 98} . Translation and rotation patient setup errors and uniform contour changes are studied for 104 VMAT plans of 4 treatment sites. Lung expansions and contractions to simulate changes in lung density are considered for 26 IMRT lung plans. A ROC curve is generated for each combination of error and indicator. For each ROC curve, the AUC value and the optimal alert threshold value of the indicator are determined.

Results: AUC values for γ-indicators and ΔPTV_{D 2} are consistently higher than for ΔD_ISOC and ΔPTV_{D 98} . For VMAT plans, error detectability to patient position shifts is worse for pelvic treatments and best for head-and-neck and brain plans. Excellent detectability is observed for 5 mm translations in head-and-neck plans (AUC = 0.94) and for 4° rotations in brain plans (AUC = 0.89). All sites but prostate show good-to-excellent detectability (AUC > 0.8) for 10 mm translations and 8° rotations and excellent detectability (AUC > 0.9) for ±6 mm patient contour changes. For head-and-neck, excellent detectability is obtained with γ-mean and γ-pass rate threshold values of around 0.63 and 83%, respectively. For brain and rectum, these threshold values are 0.53 and 90%, respectively. In IMRT lung plans, expansions of 3 mm and contractions of 6 mm are detected (AUC > 0.8).

Conclusions: By combining virtual dose reconstructions with synthetic patient data, we developed a framework to assess the sensitivity of our 3D EPID transit dosimetry method to patient-related variations. The detectability of each introduced error is specific to the treatment site and indicator used. Optimal alert criteria can be determined to ensure excellent detectability for each combination of error type and indicator. The alert threshold values and the magnitude of the error that can be detected are site-specific. In situations where the minimum error that can be detected is larger than the clinically desirable action level, EPID transit dosimetry must be used in combination with IGRT procedures to ensure correct patient positioning and early detection of anatomy variations.

Keywords: ROC analysis; anatomical changes; in vivo EPID dosimetry; patient setup; sensitivity.