Automatic deep learning treatment planning of gradient-optimized match fields for PBS proton therapy
Gunnar Arthur Helgason,
Sweden
PD-0249
Abstract
Automatic deep learning treatment planning of gradient-optimized match fields for PBS proton therapy
Authors: Gunnar Arthur Helgason1, Maja Arvola2, Dominic Maes3, Fredrik Löfman2, Lars Glimelius1
1RaySearch Laboratories, Physics department, Stockholm, Sweden; 2RaySearch Laboratories, Machine learning department, Stockholm, Sweden; 3University of Washington, School of Medicine, Department of Radiation Oncology, Seattle, USA
Show Affiliations
Hide Affiliations
Purpose or Objective
Gradient optimization is a technique used for beam matching in proton PBS and is often used for large targets that extend beyond the maximum field or to drive OAR sparing. In this technique, dosimetric gradients between match fields are designed to be sufficiently gradual to ensure plan robustness with respect to setup uncertainties. This work investigates the use of a deep learning model to infer beam-specific dose distributions to automate robust PBS treatment planning of the prostate and pelvic nodes using gradient-optimized field matching.
Material and Methods
A deep learning model has been trained to predict individual beam dose distributions based on binary volumes of patient geometries, such as target and OAR. The architecture used was a neural network based on the 3D U-Net. The model was trained on a data set consisting of 30 manually planned prostate patients with pelvic nodes, where 25 were used for training and five for validation.
An additional set of five patients was used to test the model by performing automatic planning in a research version of RayStation 11B. It consisted of deep learning prediction of individual beam doses and robust mimic optimization where optimization function penalties are based on quadratic differences to the predicted beam doses. The full optimization problem combines dose level-based optimization functions and mimic optimization functions.
All plans have been evaluated with respect to OAR sparing and robust target coverage (3 mm setup uncertainty along the inferior-superior and anterior-posterior axes and 3% range uncertainty in 16 scenarios).
Results
Figure 1 shows a comparison between the automatic and manual plan of test patient A. Gradient-optimized beam dose distributions are achieved with good target coverage and sparing of OARs.
For automatic plans, the target coverage is robust with respect to setup and range uncertainty (3mm/3%). Table 1 shows that in the high dose region of the prostate (CTV 70) the goal of 95% coverage at 70 Gy(RBE) is fulfilled in the worst scenario for all patients. The same is true for the goal of 95% coverage at 47.88 Gy(RBE) in the low dose target (CTV 50.4).
The sparing of OARs is similar for both automatic and manual plans. Clinical goals for the bladder, rectum and femur heads are fulfilled for all test patients with three exceptions, marked red in Table 1. For the bowel, three of the five automatic plans and two of the five manual plans fail to keep the dose at 1% volume below 49 Gy(RBE). The mimic optimization problem has not been set up to emphasize strict max dose constraints of the OARs.
Conclusion
The deep learning model predicts gradient-optimized beam doses, matched to yield uniform target coverage, despite a limited training set of only 30 patients. Deliverable plans with robust target coverage and OAR sparing on par with manual plans are achieved using mimic optimization. The results show the benefits of a promising method, needed for fully automated PBS proton therapy.