Toward AI-Driven Proton Dose Calculation: Development and Evaluation of 3D and 2D Sequential Neural Network Design

Abstract

AI-driven proton dose calculation may provide both the accuracy and computational speed needed during memory-intensive operations (e.g., robust treatment planning) and could ultimately pave the way to realizing a seamless adaptive proton therapy workflow. Recently, multiple approaches have been devised to effectively exploit the capabilities of the neural networks in simulating the proton transport in matter. Primarily, the three-dimensional (3D) proton beam dose distribution estimation was carried out by either training a 3D Convolutional Neural Network (3D-CNN) or by transforming the problem into a sequence of two-dimensional (2D) slices in the beam's eye view. In this study, we compare the performance of the two approaches for estimating the dose for single pencil beams (PB) and we discuss the strength and weaknesses of each approach, primarily in terms of the underlying inherent model characteristics.The two models were trained on two distinct datasets created via gold standard Monte Carlo (MC) simulation. A dataset of 10000 Topas MC simulations on artificial water phantoms exhibiting sizable cuboid inhomogeneities of varying dimensions, densities, and alignment along the beam trajectory, and a dataset of 12000 MC simulations on real-world lung patient case with different impinging proton energies, gantry angles, and isocenter shifts. While the latter dataset is meant to demonstrate the performance of each model in challenging real-life scenarios, the former is designed to monitor the behavior of each model in extreme heterogeneous scenarios. The 3D-CNN is constituted of a U-net model with a state-of-the-art ResNet encoder. The sequenced approach is incorporating a Long Short-Term Memory (LSTM) model as the frontend to transmit relevant information along the beam, and a fully connected neural network as a backend to generate the 2D dose distribution in each step along the beam trajectory.The response of the two models to simple and extreme clinical scenarios was investigated extensively in terms of range, lateral spread, and the profile deformation of the Bragg peaked region. Both models demonstrate comparable accuracy in comparison to the gold standard MC simulations resulting in 96.5 % (3D-CNN) and 97.7 % (LSTM) mean gamma index pass rate ([1%, 3mm]) when executed on the set-aside test set of the lung dataset. The 3D-CNN proves to be capable of extracting spatio-temporal features between the heterogeneities and the consequent dose deformations, however, the LSTM network exhibits a better propagation of the heterogeneities for higher ranged PBs.Both approaches to AI-driven dose calculation in proton therapy were shown to be viable candidates to estimate dose distributions at the PB level; however, in extreme scenarios of highly heterogeneous tissues and high-energy protons, advanced procedures are warranted for the design of the 3D model neural network.