(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
Prospective dose prediction is a valuable quality-control tool in the treatment planning process. In the case of proton therapy, it also allows for fast prediction of whether a specific patient is more likely to benefit from advanced planning and delivery methods. We sought to develop a preliminary, three-dimensional dose prediction tool for proton therapy using artificial neural networks (ANN).
Methods:
The initial data set used to develop our preliminary model included 24 pediatric patients with craniopharyngioma treated with pencil beam scanning (PBS) proton therapy. All patients received a total 54 Cobalt Gray Equivalent (CGE) in 30 fractions. Treatment plans consisted of two lateral or oblique beams optimized using the Eclipse treatment planning system (Varian Medical Systems, Inc., Palo Alto, CA). Features from the treatment plan, structure set, and images were extracted for each voxel in the dose grid and pooled to train the ANN. The preliminary dataset was randomly divided with 70% used for training, 15% for validation and 15% for testing. Feed forward neural network models were trained with scaled conjugate gradient back propagation and log-sigmoid activation function. Two separate models were trained corresponding to dose prediction within the clinical target volume (CTV) and outside of the CTV.
Results:
The mean squared differences between ANN predicted dose and the final TPS dose in the plan was 0.59 CGE and 1.96 CGE, respectively, for inside and outside of the CTV. There was a clear trend towards better agreement for voxels close to the CTV.
Conclusion:
We developed preliminary three-dimensional dose prediction algorithm for PBS proton therapy. Encouraged by these results, we intend to expand the validation and testing to include a total of 100 patients treated on the same protocol. ANN modeling can be applied to Linear Energy Transfer prediction.
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email
