Impact of database quality in knowledge-based treatment planning for prostate cancer

Purpose

This article investigates dose-volume prediction improvements in a common knowledge-based planning (KBP) method using a Pareto plan database compared with using a conventional, clinical plan database.

Methods and materials

Two plan databases were created using retrospective, anonymized data of 124 volumetric modulated arc therapy (VMAT) prostate cancer patients. The clinical plan database (CPD) contained planning data from each patient’s clinically treated VMAT plan, which were manually optimized by various planners. The multicriteria optimization database (MCOD) contained Pareto-optimal plan data from VMAT plans created using a standardized multicriteria optimization protocol. Overlap volume histograms, incorporating fractional organ at risk volumes only within the treatment fields, were computed for each patient and used to match new patient anatomy to similar database patients. For each database patient, CPD and MCOD KBP predictions were generated for D₁₀, D₃₀, D₅₀, D₆₅, and D₈₀ of the bladder and rectum in a leave-one-out manner. Prediction achievability was evaluated through a replanning study on a subset of 31 randomly selected database patients using the best KBP predictions, regardless of plan database origin, as planning goals.

Results

MCOD predictions were significantly lower than CPD predictions for all 5 bladder dose-volumes and rectum D₅₀ (P = .004) and D₆₅ (P < .001), whereas CPD predictions for rectum D₁₀ (P = .005) and D₃₀ (P < .001) were significantly less than MCOD predictions. KBP predictions were statistically achievable in the replans for all predicted dose-volumes, excluding D₁₀ of bladder (P = .03) and rectum (P = .04). Compared with clinical plans, replans showed significant average reductions in D_mean for bladder (7.8 Gy; P < .001) and rectum (9.4 Gy; P < .001), while maintaining statistically similar planning target volume, femoral head, and penile bulb dose.

Conclusion

KBP dose-volume predictions derived from Pareto plans were more optimal overall than those resulting from manually optimized clinical plans, which significantly improved KBP-assisted plan quality.

Summary

This work investigates how the plan quality of knowledge databases affects the performance and achievability of dose-volume predictions from a common knowledge-based planning approach for prostate cancer. Bladder and rectum dose-volume predictions derived from a database of standardized Pareto-optimal plans were compared with those derived from clinical plans manually designed by various planners. Dose-volume predictions from the Pareto plan database were significantly lower overall than those from the clinical plan database, without compromising achievability.