该文针对大动物单光子发射断层成像（single photon emission computed tomography，SPECT）系统传输矩阵测量耗时长的问题, 提出并验证了一种基于二维Gauss函数拟合的系统传输矩阵计算方法。该方法利用投影概率密度函数（projection probability density function，PPDF）在图像域上的连续性，通过采集稀疏点源数据即可构建系统传输矩阵。实验基于37 210个体素的实测数据（总采集时间124 h），利用1/4和1/9稀疏采样对该方法进行了评估。通过相对均方根误差（relative root mean square error，RRMSE）和结构相似性指示（structure similarity index measure，SSIM）定量评估系统传输矩阵精度，并利用热圆柱重建图像评价系统传输矩阵重建性能，与直接测量、重心Lagrange插值和三次样条插值方法进行了对比。结果表明，该方法在1/4和1/9稀疏条件下获得的系统传输矩阵的RRMSE分别为0.023±0.068和0.035±0.095，SSIM分别为0.997±0.003和0.996±0.006；重建结果可清晰分辨3.5 mm热圆柱，图像质量优于对比方法；系统传输矩阵计算时间分别为1.9 h和1.8 h。所提方法在保证成像精度的同时显著降低了测量和计算成本，为大动物SPECT系统传输矩阵的高效获取提供了一种可行途径。

Objective: Large-animal single-photon emission computed tomography (SPECT) systems are crucial for preclinical cardiovascular research. An accurate system matrix (or system response matrix) is essential for high-quality iterative image reconstruction. However, directly measuring the system matrix on large-animal SPECT systems is often prohibitively time-consuming due to the extensive field of view and high spatial resolution needs. To address the lengthy measurement process for the system matrix in large-animal SPECT, this work proposes and validates a calculation method based on two-dimensional (2D) Gaussian fitting. This method leverages the inherent continuity of the projection probability density function (PPDF) in the image domain. Instead of measuring the system response for each voxel, the proposed approach demonstrates that an accurate system matrix can be built by acquiring only sparse point-source measurement data. Using 2D Gaussian fitting, the method effectively models the system's spatial response and connects sparse data points to synthesize the full matrix. Methods: The complete dataset of point-source projections for 37 210 voxels was collected on the large-animal SPECT system, with a total acquisition time of 124 h. Down-sampled subsets at ratios of 1/4 and 1/9 were created from the full dataset to simulate accelerated protocols and to thoroughly evaluate the feasibility and robustness of the proposed method. The accuracy of the fitted PPDFs was quantitatively assessed against the fully measured ground truth using two metrics: the relative root mean square error (RRMSE) and the structural similarity index measure (SSIM). Additionally, to evaluate the reconstruction performance of the system matrix derived from the fitted PPDFs both qualitatively and quantitatively, 3.5 and 4 mm hot-rod phantom images were reconstructed. The performance of the proposed method was compared with several traditional approaches, including fully sampled direct measurement, barycentric Lagrange interpolation, and cubic spline interpolation. Results: Quantitative analyses showed exceptional fidelity in the estimated system matrices. The matrix computed under the 1/4 sparse sampling condition achieved an RRMSE of 0.023 ± 0.068 and an SSIM of 0.997 ± 0.003. Even with the more aggressive 1/9 sparse sampling, the method produced an RRMSE of 0.035 ± 0.095 and an SSIM of 0.996 ± 0.006. The system matrix generated with this method successfully resolved 3.5-mm hot rods. Under the 1/4 sparse sampling, the reconstructed images of 3.5 and 4 mm rods had RRMSE values of 0.795 and 0.654 against the ground truth, and SSIM values of 0.981 and 0.988. For the same sampling, the reconstructed images showed RRMSE values of 1.042 and 0.797, with SSIM values of 0.971 and 0.981. These imaging results were visually and quantitatively superior to those obtained with other methods. Importantly, the computational time remained efficient, with calculations taking only 1.9 hours and 1.8 hours for the 1/4 and 1/9 sparse datasets, respectively. Conclusions: The proposed 2D Gaussian fitting approach effectively overcomes the traditional limitations in generating system matrices for large-animal SPECT systems. It significantly reduces measurement time and computational costs without sacrificing tomographic image quality. This method presents a practical and efficient solution for acquiring precise system matrices in large-animal SPECT imaging.