Determination of the Interaction Position of Gamma Photons in Monolithic Scintillators Using Neural Network Fitting

Determination of the Interaction Position of Gamma Photons in Monolithic Scintillators Using Neural Network Fitting
复制标题

DOI:
10.1109/tns.2016.2515163
复制
发表时间:
2016-02-01
影响因子:
1.8
通讯作者:
Benlloch, J. M.
Benlloch, J. M.
中科院分区:
工程技术3区
文献类型:
--
作者:
Conde, P.;Iborra, A.;Benlloch, J. M.

文献摘要

被引文献

相似文献

在基于单片放大器的正电子发射断层成像(PET)探测器中,需要根据光电探测器像素上的光分布(LD)来估计光子相互作用位置。由于闪烁体体积的有限尺寸,LD的对称性在除了晶体中心之外的任何地方都被截断。这种效应产生了对朝向边缘的相互作用位置的差的估计,当使用诸如重心(CoG)的线性算法时,这是特别关键的情况。当所有的晶面都涂成黑色时,除了与光电探测器接触的晶面之外,可以假设LD的行为符合平方反比定律,从而提供了一个简单的理论模型。使用该LD模型,可以通过将每个事件拟合到理论分布来确定相互作用坐标。在这个意义上,神经网络(NN)的使用已被证明是一种有效的替代更传统的拟合技术,如非线性最小二乘法(LS)。多层感知器是一种神经网络,它可以很好地对非线性函数建模,并且可以在新数据出现时进行训练以准确地概括。在这项工作中,我们已经证明了神经网络的能力,以近似的LD和提供的相互作用坐标的7-光子与两个不同的光电探测器设置。一个实验设置是基于模拟硅光电倍增管(SiPMs)和电荷分配二极管网络,而第二个设置是基于数字SiPMs(dSiPMs)。在这两个实验中,NN最小化了边界效应。模拟和dSiPMs方法分别获得整个晶体表面的平均空间分辨率为1.9 +/- 0.2 mm和1.7 +/- 0.2 mm。
In Positron Emission Tomography (PET) detectors based on monolithic scintillators, the photon interaction position needs to be estimated from the light distribution (LD) on the photodetector pixels. Due to the finite size of the scintillator volume, the symmetry of the LD is truncated everywhere except for the crystal center. This effect produces a poor estimation of the interaction positions towards the edges, an especially critical situation when linear algorithms, such as Center of Gravity (CoG), are used. When all the crystal faces are painted black, except the one in contact with the photodetector, the LD can be assumed to behave as the inverse square law, providing a simple theoretical model. Using this LD model, the interaction coordinates can be determined by means of fitting each event to a theoretical distribution. In that sense, the use of neural networks (NNs) has been shown to be an effective alternative to more traditional fitting techniques as nonlinear least squares (LS). The multilayer perceptron is one type of NN which can model non-linear functions well and can be trained to accurately generalize when presented with new data. In this work we have shown the capability of NNs to approximate the LD and provide the interaction coordinates of 7-photons with two different photodetector setups. One experimental setup was based on analog Silicon Photomultipliers (SiPMs) and a charge division diode network, whereas the second setup was based on digital SiPMs (dSiPMs). In both experiments NNs minimized border effects. Average spatial resolutions of 1.9 +/- 0.2 mm and 1.7 +/- 0.2 mm for the entire crystal surface were obtained for the analog and dSiPMs approaches, respectively.