M Newpower*, S Ahmad, Y Chen, University of Oklahoma Health Sciences Center, Oklahoma City, OK

• 
  M Newpower

PO-GePV-T-75 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: A machine learning model was trained on historic data of spot position errors and MU delivery errors, taken from log files from our Mevion S250i Hyperscan proton therapy system. The output of the model will serve as input into a workflow to predict the spot position errors before treatment and the objective of this work is to assess the dosimetric impact of those errors as part of pre-treatment quality assurance.

Methods: In previous works, a MATLAB script was developed to extract spot position errors and MU delivery errors from a large collection of log files. This script generated data for approximately 4 ×10⁶ spots from 108 patients. Planned spot positions and Monitor Units (MUs), delivered spot positions and MUs, gantry angle and snout extension were used as input for a supervised machine learning algorithm. The MATLAB Statistics and Machine Learning toolbox was used to train a shallow feedforward neural network with 5 neurons. Seventy percent of the data was used for training, 15% for validation and 15% for testing.

Results: The mean squared error of the model fit on the testing data was 0.301 and R=0.99985.

Conclusion: The model was trained well, as evidenced by the R value close to unity. Output from this model will serve as input for the next step in a workflow to input the predicted spot position errors and MU errors back into the RaySearch RayStation Treatment Planning System and recalculate the dose on the patient CT. This recalculated dose, based on the predicted beam delivery, will then be compared to the dose calculated with the optimal spot positions and MUs. These comparisons will be reviewed as part of a pre-treatment quality assurance program.

ePosters

Keywords

Not Applicable / None Entered.

Taxonomy

TH- External Beam- Particle/high LET therapy: Proton therapy – quality assurance

Contact Email