Trust but Verify: Evaluating Human Modifications to Auto-Generated Rapidplan Prostate SABR Radiotherapy Plans
Abstract
Purpose
Knowledge-based planning tools like RapidPlan (RP) estimate dose-volume histograms (DVH) for organs-at-risk based on anatomical features and historical data. DVH estimates can then be used to guide treatment planning. However, planners may modify the optimization objectives generated by RP. This study quantifies the frequency and types of modifications to RP objectives and evaluates whether these manual interventions improve plans.
Methods
We retrospectively analyzed 253 prostate SABR plans created since clinical RP implementation. RP-generated optimization objectives were compared against the final clinical plan objectives using automated extraction and matching algorithms. Modifications to optimization parameters were categorized as changes to priority, dose, volume, or combinations thereof, as well as added or removed objectives. For the subset of patients with identified modifications, we generated unaltered RP-guided plans and compared dosimetric endpoints including target coverage (CTV V40Gy, PTV V36.25Gy) and bladder and rectum V37Gy, V36Gy, V33Gy, V29Gy, V18Gy. Protocol compliance rates were assessed for both clinical and automated plans.
Results
At least one optimization objective was modified in 60.9% of plans. Modifications to objectives priority (312 occurrences) were the most common change type, followed by dose modifications (181) and combined priority-dose changes (117). Among plans with optimization parameter modifications, 52.6% of clinical plans met all primary goals versus 44.8% of corresponding automated plans generated with unmodified RP objectives. Most minor protocol deviations were due to unfavorable bladder anatomy.
Conclusion
RP enabled fully automated optimization in 39.1% of prostate SABR plans. When parameter edits were required, dosimetric differences between clinical and automated plans were generally small but increased the rate at which plans met all primary protocol objectives. These findings support a tiered workflow where straightforward cases proceed with automated planning while planner expertise remains essential in case of unfavorable anatomies.