Direct Comparison between an Automated Deep Learning KBP Pipeline and Rapidplan In Lung SBRT
Abstract
Purpose
Stereotactic body radiotherapy (SBRT) plays a central role in the management of early-stage and oligometastatic lung cancer. Over recent years, there has been significant interest in automating treatment planning through the development of knowledge-based planning (KBP) techniques. Earlier methods utilized DVH estimators, while more modern methods use a deep learning (DL) model to predict voxel-level dose distributions. In this work we present a direct comparison of a commercial implementation of the former (RapidPlan) to an in-house version of the latter for lung SBRT.
Methods
We created an automated treatment planning pipeline that uses a DL model (HDUnet) to predict a dose distribution and then performs inverse planning to generate a treatment plan. We compared this pipeline to a RapidPlan model created at our institution. The HDUnet and RapidPlan models were trained on 195 and 154 lung SBRT plans respectively. They were evaluated on the same 21 non-training lung SBRT plans. The evaluation data and RapidPlan training data were drawn from the same distribution, with prescriptions ranging from 40Gy/5 to 60Gy/8. The HDUnet training data was more diverse and included prescriptions ranging from 20Gy/1 to 60Gy/8.
Results
The plans created using the dose prediction pipeline were comparable to the clinical plans. Global clinical constraint pass rates were 81% (n=17), 76% (n=16), and 57% (n=12) for Ground Truth, DL pipeline, and RapidPlan, respectively. The maximum chest wall constraint was most challenging for the RapidPlan cases, with 33% of plans failing to meet it.
Conclusion
We demonstrate that voxel-based KBP approaches may outperform current commercial solutions in lung SBRT, supporting further clinical translation and prospective evaluation. The DL method outperformed RapidPlan despite increased heterogeneity in the training data, suggesting better generalizability. Future work includes plan quality characterization beyond constraint pass rate, including complexity, quality assurance and delivery times.