Université de Lyon, CREATIS, CNRS UMR5220, Inserm U1044, INSA-Lyon, Université Lyon 1, Centre Léon Bérard 69373, France.
University of Lyon, Université Claude Bernard Lyon 1, CNRS/IN2P3, IP2I Lyon, F-69622, Villeurbanne, France.
Phys Med Biol. 2021 Feb 20;66(5):055014. doi: 10.1088/1361-6560/abde9a.
A method is proposed to model by a generative adversarial network the distribution of particles exiting a patient during Monte Carlo simulation of emission tomography imaging devices. The resulting compact neural network is then able to generate particles exiting the patient, going towards the detectors, avoiding costly particle tracking within the patient. As a proof of concept, the method is evaluated for single photon emission computed tomography (SPECT) imaging and combined with another neural network modeling the detector response function (ARF-nn). A complete rotating SPECT acquisition can be simulated with reduced computation time compared to conventional Monte Carlo simulation. It also allows the user to perform simulations with several imaging systems or parameters, which is useful for imaging system design.
提出了一种使用生成对抗网络（GAN）对正电子发射断层扫描成像设备的蒙特卡罗模拟中从患者体内逸出的粒子分布进行建模的方法。然后，生成的紧凑神经网络能够生成从患者体内逸出、朝向探测器的粒子，避免在患者体内进行昂贵的粒子追踪。作为概念验证，该方法针对单光子发射计算机断层扫描（SPECT）成像进行了评估，并与另一个建模探测器响应函数（ARF-nn）的神经网络相结合。与传统的蒙特卡罗模拟相比，该方法可以大大减少计算时间来模拟完整的旋转 SPECT 采集。它还允许用户使用几种成像系统或参数进行模拟，这对于成像系统设计很有用。
