X射线的能谱信息在计算机断层扫描(CT)中有重要应用, 准确的能谱信息可以帮助实现能谱CT, 并对传统CT重建中因射束硬化造成的伪影进行有效校正等。 在涉及CT成像的领域, 如医学成像和工业检测等, 大多数应用中都使用高功率X射线管。 这种X射线管可以在短时间内产生大量的X射线, 从而提高成像效率, 并获取高分辨率和高清晰度的图像。 然而, 由于高功率X射线管的出射光子通量过高, 直接测量其能谱变得较为困难。 目前, 常用的能谱获取方法是能谱估计, 该方法通过获取不同厚度模体的投影数据, 建立能谱、 衰减系数与投影数据之间的方程组, 最终求解该方程组以获得能谱。 能谱估计的方法依赖于算法和数据的准确性, 由于方程的严重病态, 在缺乏初始值的情况下, 求解出的能谱会缺失重要的特征信息。 为此, 提出了一种结合直接测量和神经网络模型预测的新方法。 通过使用一定厚度的铁片降低X射线的光子通量, 使其达到探测器可检测的水平, 然后逐步添加薄厚度的铁片, 获取一系列能谱数据。 这些数据用于训练神经网络模型, 以精准预测出射能谱。 仿真数据的测试结果表明, 所提神经网络模型在预测能谱方面表现优异, 归一化后的预测能谱与真实值之间的均方根误差仅为0.000 031。 然而, 由于实验条件的限制, 实测数据量相对较少, 为了提升模型的广泛适用性和预测精度, 研究进一步结合了迁移学习的方法。 首先, 利用大量的GEANT4模拟数据训练神经网络, 然后通过实验测得的少量能谱数据对模型进行微调。 实验数据的预测结果显示, 归一化后的能谱与真实值的均方根误差为0.000 194。 与传统的能谱估计方法相比, 所提出的方法不仅在精度上有显著提升, 而且还能准确捕捉到能谱的重要特征信息。 这一方法的优势在于通过神经网络强大的学习能力, 克服了传统方法对初始值和病态方程求解的依赖, 从而大大提高了预测的稳定性与准确性。 仿真数据和实验数据均验证了该方法的可行性。

The energy spectrum of X-rays plays a critical role in computed tomography (CT) applications. Accurate spectral information is crucial for implementing spectral CT and effectively correcting artifacts resulting from beam hardening in conventional CT reconstruction. In the fields of medical imaging and industrial inspection, high-power X-ray tubes are commonly used due to their ability to produce a large flux of X-rays in a short time, thereby improving imaging efficiency and enabling the acquisition of high-resolution and high-contrast images. However, directly measuring the energy spectrum of high-power X-ray tubes becomes challenging due to their excessively high photon flux.Currently, energy spectrum estimation is the predominant method for obtaining spectral information. This approach involves acquiring projection data from phantoms of varying thicknesses and establishing a system of equations that relate the spectrum, attenuation coefficients, and projection data. By solving this system, the energy spectrum can be obtained. However, the accuracy of this method relies heavily on the precision of algorithms and data. Due to the severely ill-posed nature of the equations, the reconstructed spectrum often lacks critical spectral features when suitable initial values are not provided. To address this limitation, this study proposes a novel method that combines direct measurement with neural network-based prediction. The photon flux of the X-rays is reduced to a detectable level by introducing a certain thickness of iron filters. A series of spectral data is then acquired by sequentially adding thin iron filters of varying thicknesses. These data are used to train a neural network model, enabling precise prediction of the emitted energy spectrum. Simulation results demonstrate the outstanding performance of the proposed neural network model in predicting energy spectra, achieving a root mean square error (RMSE) of only 0.000 031 between the normalized predicted spectrum and the ground truth. However, due to experimental limitations, the amount of measured data is relatively small. To enhance the model's generalizability and prediction accuracy, this study integrates transfer learning. Specifically, a neural network is first trained on a large dataset generated using GEANT4 simulations, and then fine-tuned using a small amount of experimentally measured spectral data. Prediction results based on experimental data show an RMSE of 0.000 194 between the normalized predicted spectrum and the ground truth. Compared to traditional spectral estimation methods, the proposed approach not only achieves significantly higher accuracy but also effectively captures critical spectral features. The advantages of this method lie in leveraging the powerful learning capabilities of neural networks, thereby overcoming the dependency on initial values and ill-posed equation solving in traditional methods, and greatly improving prediction stability and accuracy. Both simulation and experimental results validate the feasibility of the proposed approach.