SPAM-Net: Spectral Prior-Guided Attentive Mamba Network for Sparse-View Multienergy CT Reconstruction

Abstract

Multienergy computed tomography (CT) has demonstrated considerable clinical utility owing to its superior material differentiation capabilities. Emerging multisource CT (MSCT) systems further provide inherent advantages for efficient multienergy imaging. However, measuring attenuation of the same object across multiple energy spectra yields redundant information. In this context, we propose a complementary sparse-view multienergy data measurement protocol and use the combined projection data acquired within the same scan to reconstruct a structurally well-preserved prior image (PI). The degraded multienergy CT (MECT) images, along with the PI, are then fed into SPAM-Net to achieve artifact suppression and image quality enhancement. SPAM-Net comprises two core modules: the spectral prior-guided content-aware fusion module (SPCAFM), specifically designed for MECT, which performs differentiated and content-aware fusion of degraded multienergy images and the PI via multidimensional feature extraction and a noise- and content-driven dual-gating mechanism; and the Mamba-based attentive state-space module (ASSM), designed to effectively suppress globally distributed sparse-view artifacts by modeling nonlocal structural correlations and long-range spatial dependencies through semantic guided neighboring (SGN) rearrangement and an attentive state-space equation (ASE). This combined strategy effectively mitigates structural artifacts and improves image uniformity. Both qualitative and quantitative results show that the proposed method outperforms competing methods in detail preservation, artifact reduction, and subsequent material decomposition accuracy, suggesting the potential of complementary sparse-view sampling for dose-efficient MECT imaging in future MSCT systems.