26. Stepping validation of “knowledge-based” models for semi-automatic plan optimization: The case of concomitant treatment of pelvic nodes and prostatic bed in post-prostatectomy radiotherapy

Purpose To develop and apply a stepping validation approach for Knowledge-based (KB) models for planning optimization, starting from the complex and clinically relevant case of the treatment of pelvic nodes and prostate + seminal-vesicles bed in post-prostatectomy patients (pts). Methods Fifty-two VMAT (Varian, RapidArc) plans optimized by two planners (reference of the protocol) were selected to generate KB-model using the RapidPlan tool implemented in the Eclipse system (v.13.5). The DVH prediction model was set on the lower DVH population quartile for each OARs (rectum, bowel, bladder, femurs and penile bulb). A stepping-validation was performed: 20 randomly chosen pts used to generate the model (Closed-loop); 10 new pts, planned by the same two operators (Open-loop); 20 new pts, planned by 4 additional planners (Wide-loop). KB-plans were generated following two modalities: RP + planner intervention ( RP ) and fully-automatic RP ( only-RP ). All plans were compared against clinical plans ( RA ), in terms of dose-volume parameters and generalized equivalent uniform dose (gEUD). Two tails paired t-test was used to assess statistically significant differences. Results KB-based plans were generally better than or equivalent to clinical plans; similar trends could be appreciated without major differences between the three steps. PTVs coverage and homogeneity were comparable (Fig. 1). Concerning OARs, RP was always better with most of the improvements statistically significant; for only-RP, a better average OARs sparing was obtained (few significant differences). Concerning gEUD, RP was significantly better than RA : >1 Gy gEUD reduction was found in 80% of pts for rectum, bladder and bowel. For only-RP vs RA, the average differences were significant for few parameters: however, >1 Gy gEUD reduction was found in 50% of pts for rectum/bladder and in 70% for bowel. Conclusions A comprehensive validation of KB-model showed the possibility of automatically generating high-quality plans; planner interaction further improved performances. The model is going to be introduced in the clinical routine.