在SPECT图像重建中，有限角（LA）条件会导致频率成分的丢失，从而在对应未采集投影角度范围的方向上扭曲重建的断层图像。尽管传统迭代图像重建方法已在有限角条件下用于改善重建图像质量，但其图像质量仍不适合临床应用。我们提出了一种基于端到端深度图像先验（DIP）框架的有限角SPECT图像重建方法，以提高重建图像质量。所提出的有限角SPECT图像重建方法是一种端到端的DIP框架，将前向投影模型纳入损失函数以优化神经网络。通过结合一个二值掩模（指示测量投影数据中的每个数据点是否已被采集），该方法恢复了未采集的投影数据并重建了更少失真的图像。该方法使用20个数值仿真目标和临床患者数据进行了评估。在数值模拟中，与现有的基于反投影的方法相比，所提出的方法在PSNR和SSIM方面表现更优。我们通过特定对象的调制传递函数分析了在模拟和临床患者数据中的重建断层图像的频域响应，以评估重建方法对不同频率物体的响应。结果表明，所提出的方法显著改善了几乎所有空间频率的响应，即使在未采集的投影角度范围内也是如此。结果证明，所提出的方法可以重建更少失真的断层图像。所提出的基于端到端DIP的重建方法通过将二值掩模纳入损失函数，恢复了丢失的频率成分并减轻了有限角条件下的图像畸变。

In SPECT image reconstruction, limited-angle (LA) conditions lead to a loss of frequency components, which distort the reconstructed tomographic image along directions corresponding to the non-collected projection angle range. Although conventional iterative image reconstruction methods have been used to improve the reconstructed images in LA conditions, the image quality is still unsuitable for clinical use. We propose a LA SPECT image reconstruction method that uses an end-to-end deep image prior (DIP) framework to improve reconstructed image quality. The proposed LA SPECT image reconstruction is an end-to-end DIP framework which incorporates a forward projection model into the loss function to optimise the neural network. By also incorporating a binary mask that indicates whether each data point in the measured projection data has been collected, the proposed method restores the non-collected projection data and reconstructs a less distorted image. The proposed method was evaluated using 20 numerical phantoms and clinical patient data. In numerical simulations, the proposed method outperformed existing back-projection-based methods in terms of PSNR and SSIM. We analysed the reconstructed tomographic images in the frequency domain using an object-specific modulation transfer function, in simulations and on clinical patient data, to evaluate the response of the reconstruction method to different frequencies of the object. The proposed method significantly improved the response to almost all spatial frequencies, even in the non-collected projection angle range. The results demonstrate that the proposed method reconstructs a less distorted tomographic image. The proposed end-to-end DIP-based reconstruction method restores lost frequency components and mitigates image distortion under LA conditions by incorporating a binary mask into the loss function.