Sparse-view spectral CT reconstruction via a coupled subspace representation and score-based generative model.

Background: Spectral computed tomography (CT) demonstrates significant potential for clinical application by providing rich structural and compositional information about scanned objects. However, sparse-view scanning introduces streak artifacts during image reconstruction, severely degrading image quality. Conventional regularization-based methods exhibit inherent limitations in preserving fine details and edge structures. To address this challenge, this study aimed to enhance reconstruction quality by developing a novel framework that synergistically integrates subspace decomposition with deep generative priors, effectively leveraging both low-rank properties and data-driven representations inherent to spectral CT images. Methods: To address these challenges, we proposed an unsupervised reconstruction framework for sparse-view imaging that synergistically integrates subspace representation with a score-based generative model (SGM), which exploits intrinsic information in the measurement signals. This framework leverages the low-rank prior of the subspace representation to guide the SGM in generating images that highly coincide with the ground truth. Specifically, high-dimensional spectral CT images are first decomposed into orthogonal subspace basis components and corresponding eigen-images, effectively reducing dimensionality while preserving spectral correlations. Subsequently, we employed a data-driven SGM to learn the statistical distribution of the image. This deep prior knowledge effectively supplements the limitations of low-rank regularization in capturing complex probability distribution of image. Afterward, we integrated an efficient alternating optimization algorithm that alternately updates subspace coefficients, enforcing consistency between physical measurements and learned priors. This integration results in a synergetic effect between model-driven low-rank priors and the data-driven distribution learning, significantly enhancing the accuracy of image and the model's generalization across diverse datasets. Results: In the simulation experiment, compared with the optimal comparison algorithm (Wavelet-SGM), the proposed algorithm has increased the peak signal-to-noise ratio (PSNR) by at least 3dB, and the structural similarity index measure (SSIM) by 2.54%. In the real data experiment, the results of this paper were the closest to the ground truth, with minimum error. Both qualitative and quantitative analysis demonstrated the promising and competitive performance of the proposed method in preserving details and reducing streaking artifacts. Conclusions: Our framework established a new paradigm for spectral CT reconstruction through the synthesis of the model-driven low-rank prior with a data-driven deep prior, which yielded mutual enhancement and complementarity, collectively improving the overall quality of the reconstructed images. This dual mechanism enables comprehensive utilization of measurement signals while preventing hallucinated structures-a critical advancement for clinical applications where artifact-induced misdiagnosis carries significant risks. Our experimental results clearly demonstrate that the proposed method significantly outperforms baseline methods. On the whole, our work introduces a robust and practical sparse-view spectral CT reconstruction technique that exhibits exceptional detail preservation capabilities.