Predicting Increased Inter-Observer Variation in Pre-Treatment Image Guidance for Lung SBRT Patients

Abstract

To use CT simulation data to identify lung SBRT patients that may be more difficult to register during pre-treatment image guidance. Retrospective image registration was performed for the first and last treatment fraction for 10 lung SBRT patients by 16 individual observers. In the parameter descriptions below, GTV0 is defined as the GTV as contoured using the 0% inhalation phase of a 4DCT reconstruction while surrounding normal lung is defined as the lung volume within the PTV less GTV0. Four metrics were evaluated as possible predictors of higher inter-observer variability, and were named as follows: (1) “target excursion” defined as ITV volume-GTV0 volume, (2) “local target contrast” defined as the area of overlap of the normalized HU distributions for target and surrounding normal lung, (3) “mean HU difference’ defined as local lung mean HU – GTV0 mean HU, and (4) “target density variability” defined as GTV HU standard deviation. Regression analysis was performed to assess significance of the correlation. Three of the 4 quantities evaluated as potential predictive metrics showed statistical correlation with increased inter-observer variation. For “target excursion”, a correlation between ITV-GTV0 volume and mean inter-observer 3D vector difference was observed for both first fraction (r2 = 0.51) and last fraction (r2 = 0.39). For “local target contrast”, one significant outlier was identified having mean local lung HU beyond two standard deviations from the mean of all patients. Upon removal of this outlier, “mean HU difference” (r2 = 0.56 – first fraction, and r2 = 0.56 – last fraction) and “local target contrast” (r2 = 0.48 – first fraction, and r2 = 0.40 – last fraction) were correlated with mean 3D vector difference. Three of 10 patients had inter-observer variability that was considered clinically significant, defined here as having a 95% confidence width greater than 5 mm. All of these patients had higher “target excursion” and/or “local target contrast” scores compared to the remaining patients. Three metrics were shown here to potentially identify patients for which large inter-observer variations would be anticipated and may be used in the future to develop thresholds for additional interventions to mitigate these registration variations.