Encrypted login | home

Program Information

A Method for Statistical Comparison of Multiple Dose-Volume Histograms in Radiation Therapy: ROC Hypercurves

C Cavedon

M Giri , C Cavedon*, Azienda Ospedaliera Universitaria Integrata - Verona, Italy


SU-I-GPD-T-441 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall

Purpose: The purpose of this study was to compare multiple dose-volume-histograms (DVH) through a new statistical index (dn). The proposed applications concern DVH comparison in cohorts of patients receiving the same nominal treatment, and the intra-patient evaluation of the constancy of dose distributions delivered day-by-day.

Methods: The method is based on the extension in more than 2D of the ROC curve, defined in 2D as ROC={Sᵢ(c), i=1,2} where Sᵢ(c) denotes the survivor function for a continuous variable Y, c=[0, Ymax]. A ROC curve describes the separation between distributions. A multidimensional ROC curve (mROC) can be defined preserving the above formalism. The statistical index dn here proposed is the maximum distance between mROC and the n-dimensional diagonal line. The dn null distribution was obtained by numerical simulation and fitted to theoretical functions (Generalised Extreme Value distributions). A relationship between fitting parameters and experimental simulation conditions (number n of distributions, number n_obs of observations) was searched for. The method was applied to 5 prostate patients treated by IMRT with the same nominal prescription, to assess whether their DVHs were equivalent. The method was also used to evaluate the reproducibility, through daily CBCT, of a prostate VMAT treatment (30 fractions).

Results: Three non-linear relationships between fitting parameters and n and n_obs were found and used to generate the dn null distribution. No statistical difference was found in PTV DVHs of the IMRT patients (dn=0.048, p-value=0.157). The analysis of the 30 DVHs calculated after registration of the planning CT on CBCT showed that daily PTV irradiations were not statistically equivalent (dn=0.19, p-value=0.004). However, truncating the DVH at 98% of the prescription dose, we obtained dn=0.074 and p-value=0.997, showing that PTV coverage was statistically met.

Conclusion: The proposed dn index could be useful as a tool to improve the statistical comparison of DVHs in radiotherapy.

Contact Email: