An artificial neural network to model response of a radiotherapy beam monitoring system.

PURPOSE

The integral quality monitor (IQM) is a real-time radiotherapy beam monitoring system, which consists of a spatially sensitive large-area ion chamber, mounted at the collimator of the linear accelerator (linac), and a calculation algorithm to predict the detector signal for each beam segment. By comparing the measured and predicted signals the system validates the beam delivery. The current commercial version of IQM uses an analytic method to predict the signal, which requires a semi-empirical approach to determine and optimize various calculation parameters. The process of developing the calculation model is complex and time consuming, and moreover, the model cannot be easily generalized across various beam delivery platforms with different combinations of beam energy, beam flattening, beam shaping elements, and Linac models. Therefore, as an alternative solution, we investigated the feasibility of developing a machine learning (ML) method, using an artificial neural network (ANN), to predict the ion chamber signal. In developing an ANN, it is not necessary to explicitly account for each of the elements of beam interactions with various structures in the beam path to the ion chamber.

METHODS

The ANN was designed with multilayer perceptron (MLP). The input layer consisted of multiple features, derived from the geometrical characteristics of beam segments. Gradient descent error backpropagation technique was used to train the ANN. The combined training dataset included 270 rectangular fields, and 801 clinical IMRT fields delivered using 6 MV beams on Varian TrueBeam and Elekta Infinity . Each of 12 different ANN configurations (3 different sets of input features × 4 different sets of number of hidden nodes) was simulated 10 times with randomly selected 80% of data for training and the remaining data for validation.

RESULTS

Artificial neural networks with one hidden layer, consisting of 10 nodes, and 10 input features provided optimum results. Once the feature sets were extracted, the time required for the network training was on the order of a few minutes, and the time required to perform an output calculation per field was only fraction of a second. More than 95% of clinical intensity-modulated radiation therapy (IMRT) segments were calculated within ± 3.0% modeling error for Varian Truebeam (90% and ±3.3% for Elekta Infinity). A total of 3320 volumetric-modulated arc therapy (VMAT) segments from Truebeam were calculated using the ANN trained with IMRT fields. More than 95% of the cumulative VMAT beam segments were within 3.6% modeling error, similar to the performance for IMRT segments. In general the modeling error was found to be inversely proportional to the size and intensity of the beam segment.

CONCLUSIONS

A prototype ANN has been developed for predicting the signals of the IQM system, with substantially less efforts compared to the analytic model. The performance of the ANN was found to be at least equivalent to that of the analytic method, in terms of average and maximum error, for 6 MV beams on both Varian TrueBeam and Elekta Infinity platforms.