NeuralDAO: Incorporating neural network generated dose into direct aperture optimization for end‐to‐end IMRT planning

Abstract

Thecurrent practice in intensity-modulated radiation therapy (IMRT) planning almost always includes different dose calculation strategies for plan optimization and final dose verification. The accurate Monte Carlo (MC) dose algorithm is considered to be time-consuming for the optimization. Thus a fast, simplified dose algorithm is used in the optimization. The significant differences between the optimized dose and the delivered dose lead to tediously planning loops and potentially suboptimal solutions. This work aims to develop an IMRT optimization algorithm to minimize the dose discrepancy so that the delivered dose can be optimized in a holistic, end-to-end manner. Methods: The proposed algorithm, namely NeuralDAO, integrates a neural dose network into the column generation (CG) direct aperture optimization (DAO) formulation for step-and-shoot IMRT planning. The neural dose network is designed and trained to produce doses of MC-level accuracy within few milliseconds. Its differentiability is fully exploited to compute gradients for identifying potential aperture shapes. A prototype of NeuralDAO was developed in PyTorch and available to the public. Five lung patient cases have been studied. Dosimetric accuracy was compared with the MC dose. Plan quality and time were compared with a state-of -the-art (SoA) dose-correct algorithm. Statistical analysis was performed by Wilcoxon signed-rank test. Results:The average gamma passing rate at 2 mm/2% is 99.7% between the optimized and delivered doses. The convergence process produced by Neural-DAO is virtually identical to that produced by an MC-based DAO. The average dose calculation time is 12.1 ms for an aperture on GPU. One session of optimization took 10-36 min. Compared with the SoA, better conformity index and homogeneity index were observed for the target. The esophagus was significantly spared. Significant reductions were observed for the replanning number and the planning time. Conclusions: A new DAO algorithm based on the neural dose network has been developed. The results suggest that this algorithm minimizes the discrepancy between the optimized and delivered doses, which offers a promising approach to reduce the time and effort required in IMRT planning. This work demonstrates the possibility of applying the neural network in IMRT optimization. It is of great potential to extend this algorithm to other treatment modalities.

An ultra-fast deep-learning-based dose engine for prostate VMAT via knowledge distillation framework with limited patient data

“…For example, the calculation time per arc was approximately 3.3 s (input generation: 2.9 s and model inference: 0.4 s) and 180 s on CPU-only devices using the proposed dose engine and the Monaco MC module, respectively. In addition, comparing to other DL-based dose engines (Kontaxis et al 2020, Liu et al 2021, Tsekas et al 2021, our engine provides an additional speed boost due to its compact size, which allows for a much faster dose calculation speed than for a large size network. The student model decreases the inference time of the teacher model by 31.3%, 38.9%, and 60.0% on an A100 GPU, a RTX 3080 GPU, and a CPU, respectively.…”

“…Unlike some of the DL-based engines (Kontaxis et al 2020, Liu et al 2021, Tsekas et al 2021, which can only calculate an individual segment per inference, the proposed dose engine is able to provide a composite VMAT arc dose at one time using the composite doses in water as the input. Given that hundreds of sampled CPs are typically included in a VMAT arc, our dose engine demonstrates a great advantage over the segment/CP-based DL dose engines.…”

“…The DL-based dose engines with their ultra-fast calculation speed allow a considerable reduction in the planning time. One research group (Liu et al 2021) has successfully integrated their DL-based dose engine (large-scale convolutional neural network (CNN)) into direct aperture optimization for IMRT inverse planning. They found that using the DL-based engine to calculate dose for each iteration can substantially reduce the planning time (up to 53% on average) and minimize the discrepancies between the optimized doses and the final plan doses (computed by high-accuracy MC algorithm).…”

“…This framework was also found to be effective in IMRT dose calculation within a 1.5 T magnetic field (Tsekas et al 2021). The other strategy (Peng et al 2019, Fu et al 2020, Xing et al 2020a, 2020b, Liu et al 2021 turns the dose calculation task into a dose conversion problem (from low-accuracy doses to high-accuracy ones) using a DL model as a de-noising tool. Xing et al (2020a) trained a DL-based dose engine for 7-field (fixed) IMRT plans using doses acquired from an inaccurate (but fast and low-cost) algorithm (Lu and Chen 2010) and a highly-accurate algorithm (S/C algorithm (Ahnesjö 1989)) as the input and reference outputs, respectively.…”

“…To overcome these challenges, a dose engine should ideally possess not only a high level of efficiency but also a high level of accuracy for inverse planning. In recent years, many deep learning (DL) methods (Peng et al 2019, Fu et al 2020, Kontaxis et al 2020, Xing et al 2020a, 2020b, Liu et al 2021, Tsekas et al 2021 have been proposed for such dose engines. Two major strategies have been implemented for DL-based dose engines.…”