Monte Carlo (MC)-based simulation is the most precise method in scatter correction for Cone-beam CT (CBCT). Nonetheless, the existing MC methods cannot be fully applied in clinical due to its low efficiency. The traditional MC simulations perform calculations via a particle-by-particle scheme, which leads to high computation costs because abundant photons do not reach the X-ray detector in transport. The conventional approaches cannot control where the particle ends. Hence, it unavoidably waste lots of time in transporting numerous photons that have no contribution to the signal at the detector, yielding a low computational efficiency. To solve the problem, an innovative GPU-based Metropolis MC (gMMC) method was proposed. Compared with the traditional ones, the Metropolis based algorithm utilizes a path-by-path sampling method. The method can automatically control each particle path and eventually accelerate the convergence. In this paper, we firstly take planning CT image as prior information because of its precise CT value, and utilize gMMC to estimate scatter signal. Then the scatter signals are removed from the raw CBCT projections. Afterwards, FDK reconstruction is performed to obtain the corrected image,some accelerating strategies including reducing photon history number, pixels sampling, projection angles sampling and reconstructed image down-sampling achieve adaptive fast CBCT image reconstruction. For having high computational efficiency, we implemented the whole workflow on a 4-GPU workstation. In order to verify the feasibility of the the method, the experiment of several cases are conducted including simulation, phantom, and real patient cases. Results indicate that the image contrast becomes better, the scatter artifacts are eliminated. The maximum error (e_max), the minimum error (e_min), the 95th percentile error (e_95%), average error (¯e) are reduced from 264, 56, 14 and 21 HU to 28, 10, 3 and 7 HU in full-fan case, and from 387, 5, 19 and 95 HU to 39, 2, 2 and 6 HU in the half-fan case. In terms of computation time, the MC simulation time of all cases is within 2.5 seconds, and the total time is within 15 seconds.

基于蒙特卡洛（MC）的模拟是锥束CT（CBCT）散射校正中最精确的方法。然而，由于其低效率，现有的MC方法不能完全应用于临床。传统的MC模拟通过逐个粒子方案执行计算，由于大量光子在传输过程中未到达X射线检测器，因此这会导致较高的计算成本。传统方法无法控制粒子在何处结束。因此，在传输对检测器的信号无贡献的大量光子时，不可避免地要浪费大量时间，从而导致计算效率低下。为了解决该问题，提出了一种创新的基于GPU的Metropolis MC（gMMC）方法。与传统算法相比，基于Metropolis的算法采用逐路径采样方法。该方法可以自动控制每个粒子路径并最终加速收敛。在本文中，由于其精确的CT值，我们首先将计划CT图像作为先验信息，然后利用gMMC估计散射信号。然后从原始CBCT投影中删除散射信号。然后，通过FDK重建获得校正后的图像，减少光子历史数，像素采样，投影角度采样和重建图像下采样等一些加速策略实现了自适应快速CBCT图像重建。为了提高计算效率，我们在4-GPU工作站上实现了整个工作流程。为了验证该方法的可行性，进行了多个案例的实验，包括模拟，幻影和实际患者案例。结果表明图像对比度变得更好，消除了散射伪影。最大误差（ë_最大值），最小误差（ë_分钟），第95百分位误差（Ë _95％ ），平均误差（¯ ê）选自264，56，14和21 HU减少到28，图10，图3和7 HU在全风扇机箱，半风扇机箱从387、5、19和95 HU增至39、2、2和6 HU。在计算时间方面，所有情况的MC仿真时间均在2.5秒以内，总时间在15秒以内。