2405 - Automated Planning Dose Optimization Methods for ProstateSBRT to Generate Synthetic Dose Data for Training a Plan Optimization Deep Learning Neural Network

J. J. Sackett^1,2, L. Rosa¹, R. Clark¹, and A. Magliari¹; ¹Varian Medical Systems Inc, Palo Alto, CA, ²Department of Radiation Oncology, University of Cincinnati, Cincinnati, OH

Purpose/Objective(s): Quantify the plan quality of a Knowledge Based Planning (KBP) model with a dosimetric scorecard to generate large amounts of quantified high quality dosimetric data for training a deep learning dose optimization tool. Materials/

Methods: A preliminary KBP model was created with previously treated cases on clinical trial. A dosimetric scorecard was manually created with input from multiple physicians and dosimetrists, including the clinical trial authors. The dosimetric scorecard is a comprehensive physician’s clinical intent expressed with DVH based Dose at Volume and Volume at Dose metrics with relative points assigned on piecewise linear scoring functions for each metric, which, when summed, provide a single number quantitative score. This scorecard was used to score and iteratively guide retrospective improvements to cases before using them to train the KBP model, no direct clinical doses were used. Then, the scorecard was used to also tune the automatically generated optimization objective priorities for the KBP model. Five separate validation cases were created by the KBP model with VMAT and IMRT beam arrangements.
Results: KBP plans are generated with high quality on the validation set cases. Previous average score of clinical cases before they were manually improved and used as the training set for the KBP model was lower, 162.3/215 points.
Conclusion: This preliminary KBP model, trained with only 24 cases, is capable of generating higher quality plans than have been clinically observed. Combining this KBP model with scripting automation could be useful in generating large sums of dosimetric data required for training a Deep Learning Neural Network. Future work will explore using the dosimetric scorecard as a direct optimization method and utilizing the KBP model or direct optimization to generate data for training a neural network, with the goal of creating an autoplanning dose optimization engine which understands dosimetric tradeoffs and offers flexibility without requiring a specific scorecard for each protocol or intent and simultaneously maintaining plan quality.

	KBP optimization results on validation cases
	VMAT			IMRT
CTV [cc]	Score [X/215]	PTV-U V@36.25Gy [%]	Rectum D@50% [Gy]	Score [X/215]	PTV-U V@36.25Gy [%]	Rectum D@50% [Gy]
40	185.62	99.65	10.33	184.24	99.41	11.92
55	186.06	99.66	12.09	181.14	99.56	11.59
48	168.93	99.51	12.87	168.82	99.27	13.22
67	167.83	99.11	12.84	172.25	99.11	12.37
74	189.52	99.63	11.47	189.34	99.40	11.77
Average	179.59	99.51	11.92	179.16	99.35	12.17