Small beams, fast predictions: a comparison of machine learning dose prediction models for proton minibeam therapy.

BACKGROUND: Dose calculations for novel radiotherapy cancer treatments such as proton minibeam radiation therapy is often done using full Monte Carlo (MC) simulations. As MC simulations can be very time consuming for this kind of application, deep learning models have been considered to accelerate dose estimation in cancer patients. PURPOSE: This work systematically evaluates the dose prediction accuracy, speed and generalization performance of three selected state-of-the-art deep learning models for dose prediction applied to the proton minibeam therapy. The strengths and weaknesses of those models are thoroughly investigated, helping other researchers to decide on a viable algorithm for their own application. METHODS: The following recently published models are compared: first, a 3D U-Net model trained as a regression network, second, a 3D U-Net trained as a generator of a generative adversarial network (GAN) and third, a dose transformer model which interprets the dose prediction as a sequence translation task. These models are trained to emulate the result of MC simulations. The dose depositions of a proton minibeam with a diameter of 800µm and an energy of 20-100 MeV inside a simple head phantom calculated by full Geant4 MC simulations are used as a case study for this comparison. The spatial resolution is 0.5 mm. Special attention is put on the evaluation of the generalization performance of the investigated models. RESULTS: Dose predictions with all models are produced in the order of a second on a GPU, the 3D U-Net models being fastest with an average of 130 ms. An investigated 3D U-Net regression model is found to show the strongest performance with overall 61.0 % ± $\%\pm$ 0.5% of all voxels exhibiting a deviation in energy deposition prediction of less than 3% compared to full MC simulations with no spatial deviation allowed. The 3D U-Net models are observed to show better generalization performance for target geometry variations, while the transformer-based model shows better generalization with regard to the proton energy. CONCLUSIONS: This paper reveals that (1) all studied deep learning models are significantly faster than non-machine learning approaches predicting the dose in the order of seconds compared to hours for MC, (2) all models provide reasonable accuracy, and (3) the regression-trained 3D U-Net provides the most accurate predictions.

A self-organizing map for adaptive processing of structured data.

Recent developments in the area of neural networks produced models capable of dealing with structured data. Here, we propose the first fully unsupervised model, namely an extension of traditional self-organizing maps (SOMs), for the processing of labeled directed acyclic graphs (DAGs). The extension is obtained by using the unfolding procedure adopted in recurrent and recursive neural networks, with the replicated neurons in the unfolded network comprising of a full SOM. This approach enables the discovery of similarities among objects including vectors consisting of numerical data. The capabilities of the model are analyzed in detail by utilizing a relatively large data set taken from an artificial benchmark problem involving visual patterns encoded as labeled DAGs. The experimental results demonstrate clearly that the proposed model is capable of exploiting both information conveyed in the labels attached to each node of the input DAGs and information encoded in the DAG topology.