Author: Seungtaek Choi, Laurence Edward Court, Eun Young Han, Yusung Kim, Hunter S. Mehrens, Tucker J. Netherton, Shiqin Su 👨🔬
Affiliation: The University of Texas MD Anderson Cancer Center, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center 🌍
Purpose: Automated treatment planning is gaining traction for its enhanced consistency and efficiency. A key challenge, however, lies in the inability of neural network dose predictions directly translating into deliverable VMAT plans, necessitating further processing to compare deliverable plans derived from predictions with the predicted dose itself. In this study, we investigated an automated planning approach for prostate cancer treatment, beginning with dose prediction using a neural network model.
Methods: The predicted dose served as a reference for generating plans through an automated planning protocol (70Gy in 28 fractions). For each patient (N=10), two distinct plans were created: one based on standard clinical goals and the other utilizing DVH parameters derived directly from the predicted dose distribution as planning objectives. Gamma analysis (3%/2mm) was performed to compare dose distributions while plan quality was assessed via DVH metrics and dosiomics features.
Results: All three plans met standard clinical goals for all cases. Gamma analysis comparing the reference plan to the auto-generated plan with derived clinical goal showed >95% pass rate for target volumes suggesting strong agreement. The derived clinical goals plan reproduced DVH metrics from dose prediction, e.g., 70.4±0.2Gy vs 70.5±0.2Gy for PTV(D95%)≥70Gy and 5.4±2.3% vs 5.1±3.1% for rectum(V60Gy)<40%. Both dose prediction and derived clinical goals plans further lowered rectum V60 when comparing to the standard clinical goals plans by 7.0%. Larger kurtosis and skewness in PTV but not in CTV when comparing reference plan to derived clinical goals plan indicating further investigation in improving agreement in the margin areas.
Conclusion: The findings indicate highlight the potential of integrating dose predictions and automated planning to improve plan quality and consistency. Notably, the automated planning approach successfully generated dose distributions that were comparable to those predicted by neural network model.