Automated Data Mining of Lung SBRT Cases for Predicting Dosimetric Indices in Prospective Plans
T Atwood*, L Xing, D Hristov, Stanford University School of Medicine, Stanford, CASU-E-T-217 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall
Purpose: To develop an automated technique for evaluating lung stereotactic body radiotherapy (SBRT) plan quality and predicting achievable lung dose constraints.
Methods: We examined the workflow scripting capabilities of a radiation oncology software package (MIM Software Inc) for automatically generating mean dose gradients and overlap volume histograms (OVHs) for lung SBRT plans. Mean dose gradients were created to assess plan quality by quantifying the dose fall-off from the planning target volume (PTV) into the ipsilateral lung. OVHs were created to evaluate potential predictors of achievable lung dose constraints by measuring the amount of ipsilateral lung volume overlapping the uniformly expanded PTV. Based on the knowledge that OVHs will depend on the size and spatial relationship of the lung and PTV structures, we hypothesize that OVH-derived indices will therefore serve as predictors of achievable lung dose. The workflow scripts were evaluated in a pilot study of 10 lung SBRT plans (50 Gy in 5 fractions) with varying PTV sizes and locations.
Results: The scripted workflows were successful in automatically generating mean dose gradients and OVHs for the PTV and ipsilateral lung. The workflows were able to produce the aforementioned data in a matter of seconds (per patient), which would have otherwise been excessively time consuming and labor intensive. From the pilot study, the average mean dose gradient was -4.1 ± 0.6, indicating that all evaluated patients had similar and acceptable plan quality. Furthermore, the OVH data indicated that a 0 mm PTV expansion and a 10 mm PTV expansion correlated with the mean lung dose (correlation coefficients of 0.71 and 0.56, respectively).
Conclusions: Our current data demonstrate that mean lung dose correlates to multiple OVH indices. These results indicate that automated large-scale mining of retrospective data is a highly promising approach for predicting dosimetric indices in prospective lung SBRT plans.