Efficient knowledge-based planning model construction in institutions with limited cases using plan quality metrics.

Abstract

Prostate cancer volumetric modulated arc therapy (VMAT) planning often faces challenges in the construction of high-quality RapidPlan models when the number of cases is limited. In the present study, we retrospectively scored 90 VMAT plans using Plan Quality Metrics (PQM) and Adjusted PQM (APQM) and constructed 12 RapidPlan models from various combinations of cases with high and low PQM or APQM scores, each trained on 30 cases. Six representative models were selected for a detailed evaluation, including the P_H model based on the top 30 PQM cases and the AP_H model based on the top 30 APQM cases. All models were tested on ten independent cases that exhibited varying planning difficulties. The overall plan quality was assessed using PQM scores and dose-volume histogram (DVH) metrics for targets and organs at risk (OARs). The P_H model demonstrated significantly higher PQM scores than all other models (p < 0.05), with superior consistency and improved OAR sparing. Although the AP_H model performed well, the results were inconsistent. In challenging cases, the P_H model maintained a stable quality and outperformed both manual plans and APQM-based models. These findings indicated that case selection based on the actual clinical plan quality (PQM) is more effective than selection based on theoretical dose distributions (APQM) for building robust RapidPlan models, particularly when data are limited. This method is practical for small institutions and could be further improved by standardizing the PQM-based selection criteria and optimizing priority settings to enhance the generalizability and clinical utility of knowledge-based planning.