Detection of IMRT delivery errors using a quantitative 2D dosimetric verification system.

We investigated the feasibility of detecting intensity modulated radiotherapy delivery errors automatically using a scalar evaluation of two-dimensional (2D) transverse dose measurement of the complete treatment delivery. Techniques using the gamma index and the normalized agreement test (NAT) index were used to parametrize the agreement between measured and computed dose distributions to seven different scalar metrics. Simulated verifications with delivery errors calculated using a commercially available treatment planning system for 9 prostate and 7 paranasal sinus cases were compared to 433 clinical verifications. The NAT index with 5% and 3 mm criteria that included cold areas outside the planning target volume detected the largest percent of delivery errors. Assuming a false positive rate of 5%, it was able to detect 88% of beam energy changes, 94% of a different patient's plan being delivered, 25% of plans with one beam's collimator rotated by 90 degrees, 81% of rotating one beam's gantry angle by 10 degrees, and 100% of omitting the delivery of one beam. However, no instances of changing one beam's monitor unit setting by 10% or shifting the isocenter by 5 mm were detected. Although the phantom shift could not be detected by the small change it made in the dose distribution, our autopositioning algorithm clearly identified the spatial anomaly. Using tighter 3 %/2 mm criteria or combining dose and distance disagreements in an either/or fashion resulted in poorer delivery error detection. The mean value of the 2D gamma index distribution was less sensitive to delivery errors than the other scalar metrics studied. Although we found that scalar metrics do not have sufficient delivery error detection rates to be used as the sole clinical analysis technique, manually examining 2D dose comparison images would result in a near 100% detection rate while performing an ion chamber measurement alone would only detect 54% of these errors.