An interpretable cascaded residual iterative network for sparse-view spectral CT imaging

Sparse-view spectral tomographic image reconstruction represents a typical ill-posed inverse problem, resulting in distortion in image structures and noise surging in basis materials. Nowadays, deep learning (DL) has emerged as a state-of-the-art method in spectral image reconstruction and quantitative material analysis. However, interpretability, generalizability, and data consistency are still challenges for the existing DL-based methods. Additionally, there is no general network framework capable of simultaneously handling a series of dependent tasks in spectral imaging. This study aimed to establish a general framework for integrating multi-scene spectral imaging issues. The spectral imaging tasks, which interact in a cyclic manner during the iterative process and are optimized together.

The interpretable cascaded residual iterative network (ICRIN) for spectral tomographic reconstruction and material decomposition was established. First, as a general iterative framework based on hybrid-domain networks, ICRIN integrates physical model-driven, compressed sensing (CS), and data-driven priors to promote model stability and data consistency. Second, a residual iterative mechanism is employed to extract residual image features, which are further emphasized by a transformer attention module. Third, an interpretable objective function is established using the alternating minimization method to jointly optimize spectral images and decomposed materials. Fourth, a feedback mechanism is employed to improve the stability and performance of ICRIN in both tasks. Numerical simulations were conducted on eight patients and real preclinical experiments on 126 mouse slices to evaluate the performance of the proposed model.

Qualitative and quantitative comparisons between ICRIN and other state-of-the-art methods were conducted. The interpretability and generalizability of the ICRIN model were verified using the change curves of the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) indicators as the number of iterations increased. After iterations, the highest PSNR improvements for low- and high-energy spectral images and bone and tissue materials were approximately 6.9, 6.6, 4.0, and 8.4 dB, respectively. After the introduction of the feedback mechanism, the reconstructed images increased by approximately 3 dB, while the material images improved by approximately 1-3 dB.

This study established a general iterative framework, referred to as ICRIN, and discussed its advantages in terms of interpretability, generalizability, and data consistency in a mathematical modeling context. ICRIN could be applied across a wider range of spectral computed tomography (CT) imaging tasks, enabling clinical multi-task imaging and material quantification.