Fast Monte Carlo-Based Inverse Planning for Prostate Brachytherapy by Using Deep Learning

Abstract

Inverse planning is an essential tool for optimizing the delivered radiation dose on low-dose-rate (LDR) prostate brachytherapy. Clinical inverse planning systems use the TG-43 dose computation formalism in order to perform a fast optimization. However, this method is an approximation that often leads to a dose overestimation, resulting on sub-optimal plans. Alternatively, Monte Carlo simulation (MCS) can be used to obtain an accurate dose distribution, but considerably increasing the estimation time. We propose a fast inverse planning method for LDR prostate brachytherapy that uses a deep convolutional neural network (DCNN) trained on a GPU-based MCS generated database to estimate the dose distribution on the prostate and organs at risk. Segmentations of the organs and seeds' positions are given as the DCNN input. The mean percent error on the test set was -1.19±0.94% within the prostate. The DCNN was used to estimate the dosimetric parameters in each organ for every configuration of the optimization loop. The dosimetric parameters of the final DCNN-based brachytherapy plans were in good agreement compared to the same plans recalculated with a full MCS. The proposed inverse planning based on DCNN was capable to reach an equivalent level of accuracy with Monte Carlo with a runtime in less than 1 min using conventional GPU card.

Embedding expertise knowledge into inverse treatment planning for low‐dose‐rate brachytherapy of hepatic malignancies

“…The similar initialization settings with the number of needles were also applied in inverse treatment planning of low-dose-rate prostate brachytherapy. 10 In other planning methods, 21,33 the number of seeds was calculated using empirical formulas, while the number of needles was optimized as a parameter. Although the DVH metrics of the dose distribution demonstrate the advantage of knowledge-based initialization, in the future, we plan to investigate the impact of reducing the number of needles on the quality of the generated plans.…”

“…As mentioned in the introduction, no TPS has been specifically developed for LDR brachytherapy of liver malignancies, and hence, we can only compare the execution time of our TPS to those utilized in other clinical applications. Villa et al 21 reported an execution time of approximately 1 min for their prostate brachytherapy plan generation, excluding GTV and OAR contouring. Guthier et al 10 reported an execution time of approximately 35 s for their inverse treatment planning, which also does not include GTV and OAR contouring.…”

“…However, model‐based dose calculation is computationally expensive and is only suitable for applications with high tolerance for computational time. To address this issue, inference of the dose field using deep learning methods can achieve high computational efficiency, 21 but has yet to be validated in clinical applications. Considering the computational efficiency advantage of the TG‐43 formula, we used it as a tool for calculating the dose field in this work.…”

Automatic planning for head and neck seed implant brachytherapy based on deep convolutional neural network dose engine

“…Manual design of plan is often necessary, and the quality of the plan depends heavily on the clinician's experience. Efforts have been made to develop automatic SIBT planning methods to improve planning efficiency and quality consistency 14–17 . However, current methods are primarily designed for seed position optimization within possible parallel needle paths determined by co‐planar templates 17 and are not suitable for H&N SIBT due to complex anatomy and ample path blockages by bones, vessels, and other OARs.…”

“…As the most accurate dose calculation method, Monte Carlo simulation (MCS) has been widely applied in the plan dose evaluation for brachytherapy. However, its high computation complexity hinders the usage in routine clinical planning 15,22 . To facilitate accurate dose evaluation during plan optimization, several endeavors have been made to develop fast analytical dose computation methods by modifying the TG‐43 formula to account for heterogeneous tissue 23–27 .…”