MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Machine Learning Based Output Prediction in Uniform Scanning Proton Beam Therapy

Machine Learning Based Output Prediction in Uniform Scanning Proton Beam Therapy

H Grewal1*, T Twyford1 , M Chacko1 , S Ahmad2 , H Jin2 , (1) The Oklahoma Proton Center, Oklahoma City, OK, (2) University of Oklahoma Health Science Center, Oklahoma City, OK

Presentations

(Monday, 7/15/2019) 3:45 PM - 4:15 PM

Room: Exhibit Hall | Forum 2

Purpose: In uniform scanning proton therapy (USPT), an empirical model is used to calculate output factors (OF) based on various treatment conditions. The purpose of this work is to develop machine learning-based models to predict OF based on measured patient quality assurance (QA) data.

Methods: In the current QA approach, the output is predicted by the empirical model and verified for each patient with a water tank measurement. In this alternative approach, three Matlab-based machine learning algorithms (Gaussian process regression-exponential (GPR-E), support vector machines-fine Gaussian (SVM-FG), and neural net fitting (NNF)) were applied to predict the OF in a supervised setting. This study involved 4231 patient QA measurements conducted over the last six years. The four patient QA parameters - range (4 to 31.5 cm), modulation (2 to 25 cm), field size (2 to 26 cm; smallest dimension), and OF - were used to apply these algorithms with training and testing sets containing 90% and 10% of the randomized data, respectively. The model performance during training was accessed using root mean square error (RMSE) and R-squared values. The trained model was used to predict the OF based on range, modulation, and field size from the testing data-set. The percent differences between the predicted and measured OF were calculated.

Results: The prediction accuracy of the machine learning algorithms was higher than the empirical model. GPR-E outperformed the other algorithms. The output predictions with GPR-E, SVM-FG, NNF, and empirical model to be within 2% & 3% differences were 97.16% & 99.76%, 96.93% & 99.29%, 95.04% & 98.11%, and 91.78% & 98.53%, respectively.

Conclusion: The machine learning algorithms predict the OF for USPT with higher accuracy than the empirical model. In clinic, these can be used as primary OF calculation and second check of the OF for USPT treatment plans.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links