Département de Physique, Université de Montréal, Pavillon Roger-Gaudry, 2900 Boulevard Édouard-Montpetit, Montréal, Québec, H3T 1J4, Canada.
National Physical Laboratory, Acoustics and Ionising Radiation Team, Hampton Road, Teddington, TW11 0LW, United Kingdom.
Med Phys. 2017 Oct;44(10):5293-5302. doi: 10.1002/mp.12489. Epub 2017 Sep 4.
To propose a new formalism allowing the characterization of human tissues from multienergy computed tomography (MECT) data affected by noise and to evaluate its performance in estimating proton stopping powers (SPR).
A recently published formalism based on principal component analysis called eigentissue decomposition (ETD) is adapted to the context of noise using a Bayesian estimator. The method, named Bayesian ETD, uses the maximum a posteriori fractions of eigentissues in each voxel to determine physical parameters relevant for proton beam dose calculation. Simulated dual-energy computed tomography (DECT) data are used to evaluate the performance of the proposed method to estimate SPR and to compare it to the initially proposed maximum-likelihood ETD and to a state-of-the-art ρ - Z formalism. To test the robustness of each method towards clinical reality, three different levels of noise are implemented, as well as variations in elemental composition and density of reference tissues. The impact of using more than two energy bins to determine SPR is also investigated by simulating MECT data using two to five energy bins. Finally, the impact of using MECT over DECT for range prediction is evaluated using a probabilistic model.
For simulated DECT data of reference tissues, the Bayesian ETD approach systematically gives lower root-mean-square (RMS) errors with negligible bias. For a medium level of noise, the RMS errors on SPR are found to be 2.78%, 2.76% and 1.53% for ρ - Z, maximum-likelihood ETD, and Bayesian ETD, respectively. When variations are introduced to the elemental composition and density, all implemented methods give similar performances at low noise. However, for a medium noise level, the proposed Bayesian method outperforms the two others with a RMS error of 1.94%, compared to 2.79% and 2.78% for ρ - Z and maximum-likelihood ETD, respectively. When more than two energy spectra are used, the Bayesian ETD is able to reduce RMS error on SPR using up to five energy bins. In terms of range prediction, Bayesian ETD with four energy bins in realistic conditions reduces proton beam range uncertainties by a factor of up to 1.5 compared to ρ - Z.
The Bayesian ETD is shown to be more robust against noise than similar methods and a promising approach to extract SPR from noisy DECT data. In the advent of commercially available multi-energy CT or photon-counting CT scanners, the Bayesian ETD is expected to allow extracting more information and improve the precision of proton therapy beyond DECT.
提出一种新的形式主义方法，用于描述受噪声影响的多能量计算机断层扫描（MECT）数据中的人体组织，并评估其在估计质子阻止本领（SPR）方面的性能。
对最近发表的一种基于主成分分析的形式主义方法，即特征组织分解（ETD），使用贝叶斯估计器进行了适用于噪声的调整。该方法称为贝叶斯 ETD，它使用每个体素中特征组织的最大后验分数来确定与质子束剂量计算相关的物理参数。使用模拟的双能计算机断层扫描（DECT）数据来评估所提出方法估计 SPR 的性能，并将其与最初提出的最大似然 ETD 和最先进的ρ-Z 形式主义进行比较。为了测试每种方法对临床实际情况的稳健性，实施了三个不同水平的噪声以及参考组织的元素组成和密度变化。还通过使用两个到五个能量bins 来模拟 MECT 数据，研究了确定 SPR 时使用多个能量bins 的影响。最后，使用概率模型评估了 MECT 在射程预测方面相对于 DECT 的影响。
对于参考组织的模拟 DECT 数据，贝叶斯 ETD 方法系统地给出了较低的均方根（RMS）误差，且偏差可忽略不计。对于中等噪声水平，发现ρ-Z、最大似然 ETD 和贝叶斯 ETD 分别在 SPR 上的 RMS 误差为 2.78%、2.76%和 1.53%。当元素组成和密度发生变化时，所有实施的方法在低噪声水平下都表现出相似的性能。然而，对于中等噪声水平，与ρ-Z 和最大似然 ETD 相比，所提出的贝叶斯方法的 RMS 误差为 1.94%，表现更优，分别为 2.79%和 2.78%。当使用多个能量谱时，贝叶斯 ETD 可以使用多达五个能量bins 来降低 SPR 的 RMS 误差。在射程预测方面，在现实条件下使用四个能量bins 的贝叶斯 ETD 可将质子束射程不确定性降低多达 1.5 倍，与ρ-Z 相比。
与类似方法相比，贝叶斯 ETD 显示出对噪声的更强鲁棒性，并且是从噪声 DECT 数据中提取 SPR 的一种很有前途的方法。随着商用多能量 CT 或光子计数 CT 扫描仪的出现，预计贝叶斯 ETD 将允许从 DECT 中提取更多信息并提高质子治疗的精度。
