Development of an accurate separation of two radionuclides with an artificial neural network in dual radioisotope SPECT data acquisition

本文提出了一种新的方法，估计初级光子与人工神经网络的双同位素SPECT研究。靶同位素是I-123和Tc-99 m，用于心肌成像（Tc-99 m MIBI和I-123 BMIPP）。这两种放射性核素的峰值能量相近。为了估计初级光子，我们使用了一个神经网络，它有三层：一个输入层有十个单元，一个隐藏层有二十个单元，一个输出层有两个单元。作为输入单元的输入值，我们使用计数比，该计数比是在120至180 keV的能量范围内用窄能量窗口（6 keV）采集的计数与用宽窗口采集的总计数的比率。输出为I-123的初级计数比，以及Tc-99 m和I-123的初级计数比。有了这些主要的计数比率和总计数，我们直接计算像素的主要计数。用蒙特卡罗模拟计算的真能谱对神经网络进行训练。模拟表明，可以实现对初级光子的准确估计。

This paper presents a new method for estimating primary photons with an artificial neural network in a dual isotope SPECT study. The target isotopes are I-123 and Tc-99 m which are used for myocardial imaging (Tc-99 m MIBI and I-123 BMIPP). These two radionuclides have close photopeak energies. To estimate the primary photons we used a neural network which had three layers : one input layer with ten units, one hidden layer with twenty units and one output layer with two units. As input values to the input units, we used count ratios which were the ratios of the counts acquired with narrow energy windows (6 keV) to the total count acquired with a broad window in an energy range from 120 to 180 keV. The outputs were a primary count ratio of I-123, and a primary count ratio of Tc-99 m and I-123. With these primary count ratios and the total count we calculated the primary count of the pixel directly. The neural network was trained true energy spectra calculated by a Monte Carlo simulation. The simulation showed that an accurate estimation of primary photons was accomplished.