Purpose
Clinical dosimetric calculations used in modern brachytherapy are based upon the TG-43 formalism, but do not account for actual tissue heterogeneities. Model-based dose calculation algorithms, such as a commercially-available planning system implementing a collapsed cone algorithm, enable this and approximate Monte Carlo dosimetry. However, dosimetric discrepancies remain and real-time dose visualisation is not possible due to long calculation times.
We propose a machine learning algorithm trained on Monte Carlo data that can achieve rapid density-aware dosimetry, and present a proof-of-concept.
Methods and materials
We generated Monte Carlo dosimetry for an Ir-192 source in a water phantom using the DOSZYXnrc program from EGSnrc (v2020) [1]. The dose per initial particle was scored for a 12x12x12 cm3 box with 0.2 mm cubic voxels, with a statistical uncertainty of <2%.
We designed a convolutional neural network based on the previously described 3D U-Net [2] and implemented it in TensorFlow (version 1.15). An overview of the network architecture is shown in Figure 1. Data were divided into a training and validation set...
Results
Monte Carlo dosimetry was generated for 10 random source positions in a water phantom, 8 of which were used in model training. We found network depth, choice of loss function and initial neuron weights to be the most critical model hyperparameters requiring optimisation. As an example, the impact of network depth on model training speed and final accuracy are illustrated in Figure 2. We see increasing accuracy with deeper networks, but there are diminishing returns beyond 3 layers. Nevertheless, a 6-layer network, being the most...
Conclusion
We have successfully trained a convolutional neural network to approximate Monte Carlo dosimetry for an Ir-192 source in a water phantom. Our model generates its predictions in a fraction of a second, potentially enabling real-time dosimetry. It is encouraging that our model could predict dose for three sources despite only being trained on a single source, indicating that the underlying physical principles can be modelled by the neural network.
This work demonstrated feasibility and potential utility of machine learning techniques in brachytherapy dosimetry. Further development...
References
1. Kawrakow I, Rogers DWO, Mainegra-Hing E, Tessier F, Townson RW, Walters BRB. EGSnrc toolkit for Monte Carlo simulation of ionizing radiation transport, doi:10.4224/40001303 [release v2020] (2000).
2. Cicek O, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O (2016). 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. CoRR,abs/1606.06650.