111 - DeepTuning: A Novel Deep Learning Approach for Interactive Plan Tuning and Trade-Off Exploration

L. Ma¹, H. Liu², M. Chen³, and W. Lu³; ¹Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, ²Mayo Clinic, Rochester, MN, ³Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX,, Dallas, TX

Purpose/Objective(s): Back-and-forth plan revisions between physicians and dosimetrists occur in the planning process. They collaborate to tune plans and explore desired trade-off. To reduce back-and-forth, we proposed a new deep learning framework called DeepTuning. It can predict dose distributions with varying trade-offs from contours, so physicians can complete contours and subsequently explore trade-offs before dosimetry planning. Materials/

Methods: DeepTuning can predict doses with different trade-offs by manipulating deepest layer Z (a 1x1x1024 vector). It leverages two encoders for prior and posterior inference respectively. Prior encoder takes contours as input and extracts geometric information for conventional dose prediction, which predicts average dose with no trade-off. Posterior encoder takes both contour and dose as input and extracts ?Z encoding the trade-off of input dose. Given a template plan from a treated patient with desired trade-off, posterior inference can extract trade-off information ?Z. When predicting dose for a new patient with just contours, prior inference route predicts not only an average dose, but also doses with desired tradeoffs when we apply the ?Zs extracted
Results: We validated DeepTuning with a prostate dataset of 99 cases. We retrospectively optimized two VMAT plans for each case, prioritizing PTV coverage (pro-ptv) and rectum sparing (pro-oar). DeepTuning was trained / tested by 70 / 29 cases. The baseline is prior inference route that predicts fixed “average” dose distributions. Mean rectum doses are 49.5±9.0Gy, -3.1±5.3% cooler than ground truth (GT) pro-ptv doses (52.0±10.6Gy) and 6.6±7.9% hotter than GT pro-oar doses (44.2±12.2Gy). Then we extracted ?Zs for pro-ptv trade-off and pro-oar trade-off from the two plans of a training case. With ?Zs applied, DeepTuning can predict doses with two different trade-offs. The mean rectum doses of the pro-ptv predictions are 51.8±8.5Gy, -0.27±5.1% different from GT, while pro-oar predictions are 43.8±10.1Gy, -0.6±7.5% away from GT. The deviations from GT are much lower than “average” dose prediction
Conclusion: We introduced a novel deep learning framework, DeepTuning, capable of encoding trade-offs from treated plans and predicting doses with varying trade-offs for new cases. DeepTuning empowers physicians to tune doses and explore trade-offs immediately after contouring. As the trade-off selection occurs before dosimetry planning, back-and-forth can be minimized and treatment planning workflow can be revolutionized. Furthermore, it holds promise to clinical application of auto-planning. Physicians can generate the doses distributions they desired, which serve as optimization objectives for auto-planning better than templated objectives.

Table 1

Mean rectum dose		Pro-ptv	Pro-oar
Ground truth (GT) doses		52.0±10.6Gy	44.2±12.2Gy
Prior model	Prediction	49.5±9.0Gy
	Difference	-3.1±5.3%	6.6±7.9%
Prior inference with ?Z	Prediction	51.8±8.5Gy	43.8±10.1Gy
	Difference	-0.27±5.1%	-0.6±7.5%