Predicting liver SBRT eligibility and plan quality for VMAT and 4π plans

Abstract

Background: It is useful to predict planned dosimetry and determine the eligibility of a liver cancer patient for SBRT treatment using knowledge based planning (KBP). We compare the predictive accuracy using the overlap volume histogram (OVH) and statistical voxel dose learning (SVDL) KBP prediction models for coplanar VMAT to non-coplanar 4π radiotherapy plans.

Methods: In this study, 21 liver SBRT cases were selected, which were initially treated using coplanar VMAT plans. They were then re-planned using 4π IMRT plans with 20 inversely optimized non-coplanar beams. OVH was calculated by expanding the planning target volume (PTV) and then plotting the percent overlap volume v with the liver vs. r _v , the expansion distance. SVDL calculated the distance to the PTV for all liver voxels and bins the voxels of the same distance. Their dose information is approximated by either taking the median or using a skew-normal or non-parametric fit, which was then applied to voxels of unknown dose for each patient in a leave-one-out test. The liver volume receiving less than 15 Gy (V_<15Gy), DVHs, and 3D dose distributions were predicted and compared between the prediction models and planning methods.

Results: On average, V_<15Gy was predicted within 5%. SVDL was more accurate than OVH and able to predict DVH and 3D dose distributions. Median SVDL yielded predictive errors similar or lower than the fitting methods and is more computationally efficient. Prediction of the 4π dose was more accurate compared to VMAT for all prediction methods, with significant (p < 0.05) results except for OVH predicting liver V_<15Gy (p = 0.063).

Conclusions: In addition to evaluating plan quality, KBP is useful to automatically determine the patient eligibility for liver SBRT and quantify the dosimetric gains from non-coplanar 4π plans. The two here analyzed dose prediction methods performed more accurately for the 4π plans than VMAT.

Fig. 1

a OVH for all 20 patients. b Expansion distance at 10% OVH vs. percentage of liver receiving less than 15 Gy

Fig. 2

Distance to the target vs. dose received at every liver voxel for a 4π and b VMAT plans

Fig. 3

Boxplot results for the leave-one-out cross validation tests for the 21 patient cohort, where median, non-parametric, and skew-normal are the SVDL methods. For each box, the central mark represents the median and the edges the 25th and 75th percentiles. Whiskers cover the remaining data not deemed outliers, *p < 0.05. a Percent error in predicting the liver V < 15Gy using OVH and SVDL for 4π and VMAT plans. b Residual sum of squares analysis using SVDL to predict 4π and VMAT DVHs. c Residual sum of squares analysis for predicting the 3D dose wash using median SVDL prediction for 4π and VMAT

Fig. 4

a, c, e DVHs and b, d, f residuals of the various SVDL prediction methods for three example 4π cases

Fig. 5

a, c, e DVHs and b, d, f residuals of the various SVDL prediction methods for three example VMAT cases

Fig. 6

Actual and predicted dose for a sagittal slice of an example patient using the median SVDL prediction method for 4π and VMAT. While the actual dose includes dose for the whole body, SVDL only predicts dose within the liver