Multi-stage dual-domain progressive network with synergistic training for sparse-view CT reconstruction.

Abstract

Sparse-view computed tomography (SVCT) significantly reduces patient radiation exposure and accelerates data acquisition by using fewer projections for tomographic reconstruction. However, reconstructed images often suffer from severe artifacts, which limit their diagnostic utility. Additionally, existing CT reconstruction methods require separate training for different sparse-view scenarios, which is cumbersome and lacks the flexibility to adapt to new situations. In this paper, we propose a synergistic reconstruction method for diverse sparse-view CT scenarios. Specifically, we introduce a Multi-view Synergistic Training Strategy (MSTS) that enables a single model to handle various sparse-view CT scenarios, particularly ultra-sparse cases. Moreover, we also present the Multi-stage Dual-domain Progressive Reconstruction Network (MDPRNet), which employs a multi-stage architecture and Cross-stage Feature Adapter (CFA) to progressively reconstruct from sinograms to image domains, generating images with fine spatial details. Extensive experiments on both popular clinical CT datasets and our constructed lung abnormality dataset demonstrate that MDPRNet outperforms state-of-the-art methods across a wide range of sparse-view scenarios. The source code is available at https://github.com/hbchen98/MDPRNet.