SU-E-I-61: The Role of Interpolation Method in Computing Dose-Volume Histograms




To study the effects of dose interpolation method on DVH computation, in order to make recommendations for quality assurance and consistency of treatment.


We defined two IMRT plans (maximum doses: 84/79Gy) on a physical phantom using a commercial TPS, and exported four dose volumes with different spatial resolutions (2–5mm in-plane; 2.5–5mm thickness). Contours for fifteen structures were defined on the CT. Using two different voxel counting schemes (partial volume weighting or centre inside) and four different interpolation methods (structures mapped to dose resolution, or nearest neighbour, linear or cubic interpolation of dose to the CT grid), we computed 960 DVHs, each with 1024 bins. Ground truth was taken to be the partial-volume-weighted DVHs on the highest-resolution dose. These were compared to the suboptimal DVHs using the percentage of failing gammas (PFG) distance (doi:10.1088/0031-9155/55/11/N04) at nine different comparison points (dose differences 1,2,3Gy, volume changes 1,2,3%), giving 8640 PFG measurements. The effects were analysed via multi-way ANOVA.


DVH accuracy depended significantly on dose resolution and interpolation type (p<1e-15 in both cases). As expected, the plan, contour and PFG parameterization were also significant sources of variation. The least significant source was voxel counting method (p=0.0030). Using the etasquared measure to examine effect size, the interpolation type (eta^2=0.0209) was marginally more important than the dose resolution (eta^2=0.0312). These importances rose to eta^2=0.130 and eta^2=0.0903 respectively if comparisons used the PVW-DVHs from fine-grained doses interpolated onto the CT as ground truth instead.


For applications where DVH comparability between clinical sites with different TPS is imperative (e.g. multi-centre studies), the dose resolution and interpolation method should be strictly identical in order for DVHs to be truly comparable. For doses sampled at a lower spatial resolution, using linear — or even better, cubic — interpolation can reduce errors over nearest neighbour.