MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Residual Learning by An Artificial Neural Network for a Radiotherapy Beam Monitoring System

Y Cho1*, (1) Cleveland Clinic, Cleveland, OH

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Interpretability of artificial intelligent systems are highly desirable, especially when quality and safety of radiotherapy is involved. In our previous study, the performance of an Artificial Neural Network (ANN) has been shown to be equivalent to a complex analytic algorithm for predicting the radiotherapy treatment beam signals of the Integral Quality Monitoring (IQM) system. To improve accuracy and interpretability of the prediction, we investigated a hybrid approach: a commercial version of the analytic model and an ANNRes trained for residual errors. The performance of the hybrid approach is evaluated and compared to the conventional ANN approach.
Methods: Training dataset included 270 rectangular fields, 801 clinical IMRT fields delivered using 6MV beams on Varian TrueBeam and Elekta Infinity and measured by the IQM system. Field shapes and corresponding monitor units of the segments were converted to 10 image features as in the previous modeling study. First, the IQM signal was calculated using the analytic model and the corresponding errors were determined. Twelve different configurations of ANN(Res) were simulated to train the errors of analytic solution. For training, 80% of the dataset was randomly selected and remaining 20% was used for validation. The IQM output was computed by combining of analytic model results and the output of ANN(Res).
Results: The modeling error was -0.66%±2.98% (average ± standard deviation), -0.148%±2.24% and 0.02% ±1.44% for analytic model, ANN, and residual learning ANN(Res), respectively. The magnitude of residual learning is 0.66%±2.09% compared to the analytic model and greatly improve the interpretability and creditability due to the much reduced correctional action. The training is accelerated greatly and smaller network than conventional ANN may be suffice.
Conclusion: This hybrid method to learn residual learning greatly improve the interpretability of AI in radiotherapy quality assurance program and improve the performance of modeling as well.

Keywords

Modeling, Dosimetry

Taxonomy

TH- Radiation Dose Measurement Devices: Development (new technology and techniques)

Contact Email