基于光子计数探测器的X射线双能骨密度仪投影分解算法

Projective Decomposition Algorithms for X-Ray Dual-Energy Bone Densitometer Based on Photon Counting Detectors

对可应用于X射线骨密度仪的各种双能投影分解算法(包括曲面拟合法、查找表法、等值线拟合法、神经网络法)进行了对比研究。光子计数探测器具有高能量分辨率、低噪声的优点,由碲锌镉光子计数探测器系统进行投影数据采集可以提高分解精度。选用铝(Al)和聚甲基丙烯酸甲酯(PMMA)作为基材料,分别代表骨骼和软组织,对基材料不同厚度组合进行校准实验,建立高、低能投影数据查找表;分别使用上述投影分解算法建立反向查找表;在反向查找表的范围内选取了9个测试点,用上述分解方法分别进行分解,计算各种分解方法的分解偏差和运行时间并进行对比。结果表明,在分解精度上,曲面拟合法、查找表法、等值线拟合法、神经网络法对于Al的分解偏差分别为0.11%~3.68%、0~2.86%、0.07%~3.23%、0.41%~4.18%,PMMA的分解偏差则分别为0.11%~3.42%、0.44%~5.33%、0.02%~2.83%、0.09%~4.89%。在分解速度方面,曲面拟合法和等值线拟合法高出其他两种方法约一个数量级。等值线拟合法不管是分解精度还是速度都具有较大的优势。

Four dual-energy projective decomposition algorithms that may be applied to the X-ray bone densitometry, including surface fitting method, lookup table method, contour fitting method, and neural network method, are studied and compared. The photon counting detector has high energy resolution and low noise. The projection data acquired by the multi energy bin photon counting detector on a bench-top imaging setup helps to improve the decomposition precision. Aluminum （Al） and polymethyl methacrylate （PMMA） are selected as the base materials to represent bone and soft tissue respectively. Combinations with different base material thicknesses are used for calibration experiments to build lookup tables for high energy and low energy projections. The four projective decomposition algorithms mentioned above are used to establish the inverse lookup table, nine test points are selected in table and are decomposed by the four decomposition algorithms, and the decomposition deviation and running time of various algorithms are calculated and compared. The results show that Al thicknesses with a bias of 0.11%-3.68%, 0-2.86%, 0.07%-3.23% and 0.41%-4.18%, PMMA thicknesses with a bias of 0.11%-3.42%, 0.44%-5.330%, 0.02%-2.83% and 0.09%-4.89% are estimated by the surface fitting method, the lookup table method, the contour fitting method and the neural networks method, respectively. Compared to the lookup table method and the neural network method, the surface fitting method and the contour fitting method are faster by about an order of magnitude. The results suggest that the contour fitting method is superior in terms of decomposition accuracy and rate.