Predictive performance of an OVH-based treatment planning quality assurance model for prostate VMAT: Assessing dependence on training cohort size and composition

Radiotherapy treatment planning quality assurance models are used to assess overall plan quality in terms of dose-volume characteristics, by predicting an optimal dosimetry based on a dataset of prior cases (the training cohort). In this study, a treatment planning quality assurance model for prostate cancer patients treated with volumetric modulated arc therapy was developed using the concept of the overlap volume histogram for geometric comparison to the training cohort. The model was developed on the publically available Erasmus iCycle dataset in order to remove the effect of plan quality/inter-planner variability on the model's predictive capabilities. The model was used to predict anus, rectum, and bladder dose volume histograms. Two versions were developed: the n = 114 case (leave-one-out method) which made predictions using the complete Erasmus dataset, and the similarity index (SI)-based model which used a smaller training cohort allocated in order of geometric similarity determined using an overlap volume histogram-derived SI. The difference in mean dose (predicted-achieved) of the SI model at cohort sizes of 10, 20, 30, 40, 50, 75, and 100 was compared to the leave-one-out method for 5 patients, in an attempt to determine the “optimum” cohort size for the SI-based model in this dataset. Performance of the optimized SI model was compared to the leave-one-out method for all patients using the following metrics: difference in mean and median dose, difference in V65Gy and V75Gy (rectum only), similarity of predicted and achieved mean dose, and mean dose volume histograms residual. The “optimum” cohort size for the SI-based model was determined to be 45. The SI-based model implementing this cohort size yielded slightly better outcomes in all performance metrics for the rectum and anus, but worse for the bladder. SI-based training cohort allocation can lead to better predictive efficacy, but the cohort size should be optimized for each individual organ.